Search and recommendation (S&R) are fundamental components of modern commercial platforms, enabling users to access and explore information efficiently. User behaviors in these scenarios reflect different aspects of user intent, providing an opportunity for joint modeling of S&R. However, effectively leveraging search logs to enhance recommendation remains a challenging task. Existing methods often encode S&R histories either jointly or separately; however, they tend to regard all search signals as equally informative, thereby neglecting that many search behaviors can be irrelevant or even detrimental to recommendation performance. In practice, however, search histories frequently contain noisy or outdated behaviors that may introduce spurious correlations and degrade recommendation performance. Motivated by the human decision-making process, where one first identifies recommendation intent and then selectively reasons about relevant search signals, we propose LCR-SER, a latent cross reasoning method for search-enhanced recommendation. LCR-SER first encodes the user's S&R history into a unified latent representation that captures users' global interests. It then performs iterative reasoning in the latent space to dynamically identify informative search signals that are most relevant to the recommendation. To further guide this reasoning process, we introduce contrastive learning to align the reasoning states with the target items. In addition, we employ reinforcement learning to directly optimize ranking-oriented metrics, enabling LCR-SER to refine its reasoning strategy toward improved recommendation performance. Experiments on public datasets demonstrate that LCR-SER consistently outperforms strong baselines, validating the effectiveness of latent reasoning in enhancing search-aware recommendation.

Generative models are widely used to produce synthetic images with annotations, alleviating the burden of image collection and annotation for training deep visual models. However, challenges such as limited image diversity, noisy pseudo labels, and domain gaps between synthetic and real images often undermine their effectiveness in downstream visual tasks. This paper introduces the Iterative Self-Training with Class-Aware Text-to-Image Synthesis (IST-CATS) framework, which addresses these challenges by integrating a class-aware text-to-image synthesis (CATS) component with an iterative self-training (IST) strategy. CATS innovatively introduces a class-aware chain approach to generate detailed descriptions. These descriptions act as prompts for a diffusion model, enabling the creation of a diverse of images accompanied by distinguishable objects against the background. The generated images can be easily pseudo-labeled by an unsupervised instance segmentation method, and then noisy pseudo labels can be effectively purified by a novel feature similarity-based filtering mechanism. The generated images underpin our IST, which progressively enhances vision models and refines pseudo labels through self-training and our proposed label filtering strategy (LabFilt). LabFilt meticulously improves the quality of pseudo labels by employing class-adaptive techniques at both the pixel and object levels, ensuring refined pseudo-label accuracy. IST-CATS demonstrates superior performance in object detection and semantic segmentation compared to traditional synthetic and semi/weakly-supervised methods, effectively addressing data collection and annotation challenges.

Weicong Qin, Yi Xu, Weijie Yu, Chenglei Shen, Ming He, Jianping Fan, Xiao Zhang, Jun Xu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

In modern online platforms, search and recommendation (S&R) often coexist, offering opportunities for performance improvement through search-enhanced approaches. Existing studies show that incorporating search signals boosts recommendation performance. However, the effectiveness of these methods relies heavily on rich search interactions. They primarily benefit a small subset of users with abundant search behavior, while offering limited improvements for the majority of users who exhibit only sparse search activity. To address the problem of sparse search data in search-enhanced recommendation, we face two key challenges : (1) how to learn useful search features for users with sparse search interactions, and (2) how to design effective training objectives under sparse conditions. Our idea is to leverage the features of users with rich search interactions to enhance those of users with sparse search interactions. Based on this idea, we propose GSERec, a method that utilizes message passing on the User-Code Graphs to alleviate data sparsity in Search-Enhanced Recommendation. Specifically, we utilize Large Language Models (LLMs) with vector quantization to generate discrete codes, which connect similar users and thereby construct the graph. Through message passing on this graph, embeddings of users with rich search data are propagated to enhance the embeddings of users with sparse interactions. To further ensure that the message passing captures meaningful information from truly similar users, we introduce a contrastive loss to better model user similarities. The enhanced user representations are then integrated into downstream search-enhanced recommendation models. Experiments on three real-world datasets show that GSERec consistently outperforms baselines, especially for users with sparse search behaviors.

Personalized search systems in e-commerce platforms increasingly involve user interactions with AI assistants, where users consult about products, usage scenarios, and more.Leveraging consultation to personalize search services is trending.Existing methods typically rely on semantic similarity to align historical consultations with current queries due to the absence of 'value' labels, but we observe that semantic similarity alone often fails to capture the true value of consultation for personalization.To address this, we propose a consultation value assessment framework that evaluates historical consultations from three novel perspectives: (1) Scenario Scope Value, (2) Posterior Action Value, and (3) Time Decay Value.Based on this, we introduce VAPS, a value-aware personalized search model that selectively incorporates high-value consultations through a consultation-user action interaction module and an explicit objective that aligns consultations with user actions.Experiments on both public and commercial datasets show that VAPS consistently outperforms baselines in both retrieval and ranking tasks.Codes are available at https://github.com/E-qin/VAPS.

Accurate 6D object pose estimation from RGB images is crucial for various computer vision applications, such as augmented reality, robotic manipulation and autonomous driving. Existing methods often rely on extensive labeled data, either manually annotated or synthetically generated, which can be laborious and impractical for real-world deployment. To address these challenges, we propose OK-POSE, a keypoint-based 6D object pose estimation method that leverages relative transformations between viewpoints for training. By utilizing pairs of images with object annotations and relative transformation information, OK-POSE automatically learns to detect 3D keypoints of objects, enabling geometrically and visually consistent pose estimation. The simplicity and accessibility of obtaining relative transformation information, which can be acquired from inexpensive binocular cameras or common smartphone devices, significantly reduce labeling costs and mitigate domain gap issues associated with synthetic data. Experimental results demonstrate that OK-POSE achieves competitive performance compared to methods relying on explicit 3D annotations or object 3D models. Moreover, we provide insights into the data collection process and introduce OK-POSE++, an enhanced version with optimized network architecture and loss functions, yielding further improvements in performance. Our approach offers a practical solution for 6D object pose estimation, suitable for real-world applications in scenarios where extensive 3D annotations or object models are unavailable. The code is released athttps://github.com/acmff22/OKPOSE.

Deep learning-based object detection models rely heavily on large-scale and precise annotations for training. However, manually annotating bounding-box annotations for such data is both time-consuming and costly, especially when dealing with high-resolution satellite imagery containing densely packed small-sized objects. To alleviate the burden of manual annotation, we propose a simple yet effective approach, called progressive self-training object detection (PSTDet), to enable accurate object detection in remote sensing imagery without relying on manual annotations. Our PSTDet framework consists of two main components: initial pseudo label generation (IPLG) and progressive self-training with relabeling (PST-R). In IPLG, we leverage unsupervised image clustering, unsupervised instance detection, and geometric constraints to automatically generate high-quality bounding-box annotations for the initial training dataset. This innovative approach significantly reduces the time and expense associated with data annotation, laying a solid foundation for the subsequent progressive self-training stage. The annotations produced by IPLG serve as the training data for PST-R, which enhances the detector and pseudo labels through progressive self-training and our proposed noisy pseudo label filtering strategy (NPLFilter). Our NPLFilter purifies the quality of pseudo labels by integrating geometric constraints, prior knowledge, and category-adaptive thresholds. Experimental results demonstrate that our method achieves significant performance improvement on challenging NWPU VHR-10.v2 and DIOR datasets. Notably, our method far outperforms state-of-the-art weakly supervised methods and compares favorably with fully supervised methods.

In this article, we investigate a new problem of generating a variety of multi-view fashion designs conditioned on a human pose and texture examples of arbitrary sizes, which can replace the repetitive and low-level design work for fashion designers. To solve this challenging multi-modal image translation problem, we propose a novel Photo-reAlistic fashIon desigN synThesis (PAINT) framework, which decomposes the framework into three manageable stages. In the first stage, we employ a Layout Generative Network (LGN) to transform an input human pose into a series of person semantic layouts. In the second stage, we propose a Texture Synthesis Network (TSN) to synthesize textures on all transformed semantic layouts. Specifically, we design a novel attentive texture transfer mechanism for precisely expanding texture patches to the irregular clothing regions of the target fashion designs. In the third stage, we leverage an Appearance Flow Network (AFN) to generate the fashion design images of other viewpoints from a single-view observation by learning 2D multi-scale appearance flow fields. Experimental results demonstrate that our method is capable of generating diverse photo-realistic multi-view fashion design images with fine-grained appearance details conditioned on the provided multiple inputs. The source code and trained models are available at https://github.com/gxl-groups/PAINT .

Image-based virtual try-on aims at transferring a target in-shop garment onto a reference person, and has garnered significant attention from the research communities recently. However, previous methods have faced severe challenges in handling occlusion problems. To address this limitation, we classify occlusion problems into three types based on the reference person’s arm postures: single-arm occlusion , two-arm non-crossed occlusion , and two-arm crossed occlusion . Specifically, we propose a novel Occlusion-Free Virtual Try-On Network (OF-VTON) that effectively overcomes these occlusion challenges. The OF-VTON framework consists of two core components: (i) a new Recurrent Appearance Flow based Deformation (RAFD) model that robustly aligns the in-shop garment to the reference person by adopting a multi-task learning strategy . This model jointly produces the dense appearance flow to warp the garment and predicts a human segmentation map to provide semantic guidance for the subsequent image synthesis model. (ii) a powerful Multi-mask Image SynthesiS (MISS) model that generates photo-realistic try-on results by introducing a new mask generation and selection mechanism . Experimental results demonstrate that our proposed OF-VTON significantly outperforms existing state-of-the-art methods by mitigating the impact of occlusion problems. Our code is available at https://github.com/gxl-groups/OF-VTON .

Building benchmarks to systemically analyze different capabilities of video question answering (VideoQA) models is challenging yet crucial. Existing benchmarks often use non-compositional simple questions and suffer from language biases, making it difficult to diagnose model weaknesses incisively. A recent benchmark AGQA [8] poses a promising paradigm to generate QA pairs automatically from pre-annotated scene graphs, enabling it to measure diverse reasoning abilities with granular control. However, its questions have limitations in reasoning about the fine-grained semantics in videos as such information is absent in its scene graphs. To this end, we present ANetQA, a large-scale benchmark that supports fine-grained compositional reasoning over the challenging untrimmed videos from ActivityNet [4]. Similar to AGQA, the QA pairs in ANetQA are automatically generated from annotated video scene graphs. The fine-grained properties of ANetQA are reflected in the following: (i) untrimmed videos with fine-grained semantics; (ii) spatio-temporal scene graphs with fine-grained taxonomies; and (iii) diverse questions generated from fine-grained templates. ANetQA attains 1.4 billion unbalanced and 13.4 million balanced QA pairs, which is an order of magnitude larger than AGQA with a similar number of videos. Comprehensive experiments are performed for state-of-the-art methods. The best model achieves 44.5% accuracy while human performance tops out at 84.5%, leaving sufficient room for improvement.

Abstract Point-level temporal action localization (PTAL) aims to locate action instances in untrimmed videos with only one timestamp annotation for each action instance. Existing methods adopt the localization-by-classification paradigm to locate action boundaries in the temporal class activation map (TCAM) by thresholding, also known as TCAM-based method. However, TCAM-based methods are limited by the gap between classification and localization tasks, since TCAM is generated by a classification network. To address this issue, we propose a re-training framework for the PTAL task, also known as LPR. This framework consists of two stages: pseudo-label generation and re-training. In the pseudo-label generation stage, we propose a feature embedding module based on a transformer encoder to capture global context features and optimize pseudo-labels’ quality by leveraging point-level annotations. In the re-training stage, LPR uses the above pseudo-labels as supervision to locate action instances with a temporal action localization network rather than generating TCAMs. Furthermore, to alleviate the effects of label noise in the pseudo-labels, we propose a joint learning classification module (JLCM) in the re-training stage. This module contains two classification sub-modules that simultaneously predict action categories and are guided by a jointly determined clean set for network training. The proposed framework achieves state-of-the-art localization performance on both the THUMOS’14 and BEOID datasets.

The current social media is flooded with hyper realistic face-synthetic videos due to the explosion of DeepFake technology that has brought a serious impact on human society security, which calls for further exploring on deepfake video detection methods. Existing methods attempt to isolated capture spatial artifacts or extract the homogeneous temporal inconsistency to detect deepfake video, but little attention has been paid to the exploitation of dynamic spatial-temporal inconsistency. To mitigate this issue, in this paper, we propose a novel Multi-Rate Excitation Network (MRE-Net) to effectively excite dynamic spatial-temporal inconsistency from the perspective of multiple rates for deepfake video detection. The proposed MRE-Net is composed of two components: Bipartite Group Sampling (BGS) and multiple rate branches. The BGS draws the entire video into multiple bipartite groups with different rates to cover various face motion dynamic evolution. We further design multiple rate branches to capture both short-term and long-term spatial-temporal inconsistency from corresponding bipartite groups of BGS. Concretely, for the early stages of the multi-rate branches, Momentary Inconsistency Excitation (MIE) module is developed to encode the spatial artifacts and intra-group short-term temporal inconsistency. Meanwhile, for the last stages of the multi-rate branches, Longstanding Inconsistency Excitation (LIE) module is constructed to perceive inter-group long-term temporal dynamics. Extensive experiments and visualizations conducted on four popular datasets demonstrate the effectiveness of the proposed method against state-of-the-art deepfake detection methods.

Most existing image privacy protection works focus mainly on the privacy of photo owners and their friends, but lack the consideration of other people who are in the background of the photos and the related location privacy issues. In fact, when a person is in the background of someone else’s photos, he/she may be unintentionally exposed to the public when the photo owner shares the photo online. Not only a single visited place could be exposed, attackers may also be able to piece together a person’s travel route from images. In this article, we propose a novel image privacy protection system, called LAMP, which aims to light up the location awareness for people during online image sharing. The LAMP system is based on a newly designed location-aware multi-party image access control model. Unlike previous works on small scales, the LAMP system is highly efficient and scalable as it can enforce privacy protection for billions of users on social networks in real time. The LAMP system automatically detects the user’s occurrences on photos regardless the user is the photo owner or not. Once a user is identified and the location of the photo is deemed sensitive according to the user’s privacy policy, the user’s face will be replaced with a synthetic face. A prototype of the system was implemented and evaluated to demonstrate its applicability in the real world.

The performance of deep networks for object detection in remote sensing images (RSIs) largely depends on the availability of large-scale training images whose labels are given at the bounding-box level through a labor-intensive manual labeling process. To alleviate the huge burden of providing bounding-box annotations manually for movable objects, we propose a new approach, called dynamic automatic learning (DAL), to progressively learn object detectors. Specifically, a novel initial annotation generation (IAG) strategy is first designed to produce bounding-box annotations for movable objects in multi-temporal remote sensing images. During this process, image-level labels need to be manually labeled for the generated candidates. Next, a detection network learns the detection knowledge from multi-temporal remote sensing images with bounding-box annotations and then transfers the knowledge to generate pseudo boxes for the unlabeled data. Finally, with these pseudo boxes, the object detector can be optimized for generating accurate pseudo boxes iteratively. Furthermore, we introduce a pseudo box filtering (PBF) strategy to purify the quality of pseudo boxes to obtain accurate supervision. Our experiments on the challenging NWPU VHR-10.v2 and DIOR datasets have demonstrated that our DAL approach can achieve competitive results compared to state-of-the-art methods.

Cloud detection is one of the key technologies in the field of remote sensing. Although extensive deep learning-based cloud detection methods achieve good performance, their detection results in confusing areas such as cloud boundaries and thin clouds are often not satisfactory due to the potential inter-class similarity and intra-class inconsistency of objects. To this end, we propose a Boundary-Aware Bilateral Fusion network (BABFNet), which effectively enhances cloud detection in confusing areas by introducing a boundary prediction branch as an auxiliary. To avoid the loss of details, the boundary prediction branch is designed to run at full resolution with a shallow architecture, while some Semantic Enhancement Modules (SEMs) are used to supplement high-level semantic information by introducing multi-level encoder features of the cloud detection branch. This feature sharing in turn drives the cloud detection branch to focus more on cloud boundaries during training. At the end of the network, a Bilateral Fusion Module (BFM) is added for information complementarity between features from these two branches. The features from the cloud detection branch provide multi-scale features to the boundary prediction branch for more accurate boundary prediction, while the features from the boundary prediction branch further serve as prior knowledge to help the cloud detection branch aggregate contextual information. To verify the effectiveness of the proposed method, we select four different networks as cloud detection branches and conduct comparative experiments on two public datasets, GF-1 WFV and MODIS. The experimental results show that the proposed method significantly enhances cloud detection in confusing areas.

Deep learning has achieved promising performance on semantic segmentation, but few works focus on semantic segmentation at the fine-grained level. Fine-grained semantic segmentation requires recognizing and distinguishing hundreds of sub-categories. Due to the high similarity of different sub-categories and large variations in poses, scales, rotations, and color of the same sub-category in the fine-grained image set, the performance of traditional semantic segmentation methods will decline sharply. To alleviate these dilemmas, a new approach, named Class Guided Channel Weighting Network (CGCWNet), is developed in this paper to enable fine-grained semantic segmentation. For the large intra-class variations, we propose a Class Guided Weighting (CGW) module, which learns the image-level fine-grained category probabilities by exploiting second-order feature statistics, and use them as global information to guide semantic segmentation. For the high similarity between different sub-categories, we specially build a Channel Relationship Attention (CRA) module to amplify the distinction of features. Furthermore, a Detail Enhanced Guided Filter (DEGF) module is proposed to refine the boundaries of object masks by using an edge contour cue extracted from the enhanced original image. Experimental results on PASCAL VOC 2012 and six fine-grained image sets show that our proposed CGCWNet has achieved state-of-the-art results.

The rapidly growing application of pedestrian images has aroused wide concern on visual privacy protection because personal information is under the risk of privacy disclosure. Anonymization is regarded as an effective solution by identity obfuscation. Most recent methods focus on face, but it is not enough when the presence of human body carries lots of identifiable information. This paper presents a new delegate-based utility preserving synthesis (DUPS) approach for pedestrian image anonymization. This is challenging because one may expect that the anonymized image can still be useful in various computer vision tasks. We model DUPS as an adaptive translation process from source to target. To provide a comprehensive identity protection, we first perform anonymous delegate sampling based on image-level differential privacy. To synthesize anonymous images, we then introduce an adaptive translation network and optimize it with a multi-task loss function. Our approach is theoretically sound and can generate diverse results by preserving data utility. The experiments on multiple datasets show that DUPS can not only achieve superior anonymization performance against deep pedestrian recognizers, but also can obtain a better tradeoff between privacy protection and utility preservation compared with state-of-the-art methods.

Image classification has been widely adopted by current social media applications. Compared with fully supervised classification, semisupervised classification attracts more attention because it is commonly observed that category labels are only available for a small portion of images while most images on social media platforms do not have labels. To this end, we propose a two-stage semisupervised learning framework. In the first stage, we train two Self-supervised Models (SSMs). One model is initialized by predicting the rotation angles of pretransformed training images and then further trained by the labeled images. The other model is initialized by making consistent predictions for the transformed images in color, shape, and quality from the same sample image, and then further trained by the labeled images. In the second stage, we fuse the two SSMs through deep mutual learning, which enhances each of the two SSMs with the complementary information provided by the other such that the correct prediction could be shared. Experimental results on CIFAR and Caltech-256 data sets demonstrate the effect of the proposed framework.

In recent years, a large number of convolutional neural networks-based cloud detection algorithms have been proposed for remote sensing image preprocessing and most of them have an encoder-decoder structure. However, downsampling and upsampling operations, as the basic components of these methods, inevitably lead to the loss of detailed information in high-level features, which affects cloud detection performance. At the same time, the physical characteristics of the cloud, such as the variable size and irregular structure, also put forward requirements for the multi-scale feature representation ability of the network. To this end, we propose a novel cloud detection network named DMNet, which contains a Dense Feature Enhancement Module (DFEM) and a Multi-scale Context Fusion Spatial Attention Module (MCFSAM). DFEM aims to achieve information complementarity by exploiting the different properties of the features at different levels of the encoder, so as to strengthen the detailed information of high-level features and make the low-level features have more semantics. MCFSAM introduces a Multi-scale Context Fusion Block (MCFB) in spatial attention, which enables the network to densely capture contextual information at different scales and further emphasize useful features in the spatial dimension. Extensive experiments on GF-1 wide field-of-view satellite imagery (GF-1 WFV) dataset and Moderate-Resolution Imaging Spectroradiometer (MODIS) dataset demonstrate that our method outperforms other state-of-the-art cloud detection algorithms.

Object detection from remote sensing images (RSIs) is a basic topic in the area of aviation and satellite image processing, which has great effects on geological disaster detection, agricultural planning, and land utilization. However, it is always faced with several severe difficulties. For instance, the scale of the target spans over a very wide range, and the difference between the target size and the image size is huge, as some targets only account for a dozen pixels compared with the RSI of the megapixel level. In this work, an innovative object detection network (GLNet) is proposed for remote sensing imagery. Our approach integrates global context clues extracted by the multiscale perception (MSP) module and local spatial contextual correlations encoded by Clip long short-term memory (Clip-LSTM). These rich semantic features are further exploited to design a self-adapted anchor (SA) module to alleviate the scale variations in RSIs. Extensive experiments are executed on several public easily accessed benchmarks, including DOTA, NWPU VHR-10, and DIOR. Experimental results have demonstrated that our GLNet outperforms numerous latest methods.

Video object segmentation is a challenging task in computer vision because the appearances of target objects might change drastically along the time in the video. To solve this problem, space-time memory (STM) networks are exploited to make use of the information from all the intermediate frames between the first frame and the current frame in the video. However, fully using the information from all the memory frames may make STM not practical for long videos. To overcome this issue, a novel method is developed in this paper to select the reference frames adaptively. First, an adaptive selection criterion is introduced to choose the reference frames with similar appearance and precise mask estimation, which can efficiently capture the rich information of the target object and overcome the challenges of appearance changes, occlusion, and model drift. Secondly, bi-matching (bi-scale and bi-direction) is conducted to obtain more robust correlations for objects of various scales and prevents multiple similar objects in the current frame from being mismatched with the same target object in the reference frame. Thirdly, a novel edge refinement technique is designed by using an edge detection network to obtain smooth edges from the outputs of edge confidence maps, where the edge confidence is quantized into ten sub-intervals to generate smooth edges step by step. Experimental results on the challenging benchmark datasets DAVIS-2016, DAVIS-2017, YouTube-VOS, and a Long-Video dataset have demonstrated the effectiveness of our proposed approach to video object segmentation.

The existing publicly available datasets with pixel-level labels contain limited categories, and it is difficult to generalize to the real world containing thousands of categories. In this paper, we propose an approach to generate object masks with detailed pixel-level structures/boundaries automatically to enable semantic image segmentation of thousands of targets in the real world without manually labelling. A Guided Filter Network (GFN) is first developed to learn the segmentation knowledge from an existed dataset, and such GFN then transfers the learned segmentation knowledge to generate initial coarse object masks for the target images. These coarse object masks are treated as pseudo labels to self-optimize the GFN iteratively in the target images. Our experiments on six image sets have demonstrated that our proposed approach can generate object masks with detailed pixel-level structures/boundaries, whose quality is comparable to the manually-labelled ones. Our proposed approach also achieves better performance on semantic image segmentation than most existing weakly-supervised, semi-supervised, and domain adaptation approaches under the same experimental conditions.

As one of the fundamental visual tasks in the unmanned driving area, lane detection attracts increasing attention. In practical applications, lane points are very difficult to localize because they usually appear to be sparse and incomplete due to the influence of illumination and environment. Conventional lane detection methods rely on coarse features and carefully designed postprocessing to detect the lane lines. However, these methods are usually slow, and the stability and generalization ability are unsatisfactory. In this paper, we propose a novel lane detection network – Fast-HBNet(Fast-Hybrid Branch Network), which exploits both global semantic information and spatial contexts. To enlarge receptive fields and encode more detailed information, the compound transformation is employed and the proposed hybrid branch network extracts four diverse feature maps with different receptive fields and spatial contexts. Besides, we design a Hierarchical Feature Learning (HFL) module to learn lane features from the scale, channel, and spatial levels to enhance the generalization ability of our detector. These features are further selectively coalesced to generate unified lane feature maps with large receptive fields and rich detailed information. In other words, our network can encode the global semantic information from the high-resolution feature maps and the fine-grained details in the low-resolution feature maps. Experimental results conducted on the TuSimple (2017) and CULane [Panet al.(2018)] datasets demonstrate that the proposed Fast-HBNet outperforms numerous state-of-the-art lane detectors in both speed and accuracy. Particularly, Fast-HBNet achieves an accuracy of 96.88% on the TuSimple dataset at a speed of 76 FPS.

Although deep learning has demonstrated its outstanding performance on image classification, most well-known deep networks make efforts to optimize both their structures and their node weights for recognizing fewer (e.g., no more than 1000) object classes. Therefore, it is attractive to extend or mixture such well-known deep networks to support large-scale image classification. According to our best knowledge, how to adaptively and effectively fuse multiple CNNs for large-scale image classification is still under-explored. On this basis, a deep mixture algorithm is developed to support large-scale image classification in this paper. First, a soft spectral clustering method is developed to construct a two-layer ontology (group layer and category layer) by assigning large numbers of image categories into a set of groups according to their inter-category semantic correlations, where the semantically-related image categories under the neighbouring group nodes may share similar learning complexities. Then, such two-layer ontology is further used to generate the task groups, in which each task group contains partial image categories with similar learning complexities and one particular base deep network is learned. Finally, a gate network is learned to combine all base deep networks with fewer diverse outputs to generate a mixture network with larger outputs. Our experimental results on ImageNet10K have demonstrated that our proposed deep mixture algorithm can achieve very competitive results (top 1 accuracy: 32.13%) on large-scale image classification tasks.

Few-shot classification aims to recognize new classes by learning reliable models from very few available samples. It could be very challenging when there is no intersection between the alreadyknown classes (base set) and the novel set (new classes). To alleviate this problem, we propose to evolve the network (for the base set) via label propagation and self-supervision to shrink the distribution difference between the base set and the novel set. Our network evolution approach transfers the latent distribution from the already-known classes to the unknown (novel) classes by: (a) label propagation of the novel/new classes (novel set); and (b) design of dual-task to exploit a discriminative representation to effectively diminish the overfitting on the base set and enhance the generalization ability on the novel set. We conduct comprehensive experiments to examine our network evolution approach against numerous state-of-the-art ones, especially in a higher way setup and cross-dataset scenarios. Notably, our approach outperforms the second best state-of-the-art method by a large margin of 3.25% for one-shot evaluation over miniImageNet.

The growing application of face images and modern AI technology has raised another important concern in privacy protection. In many real scenarios like scientific research, social sharing and commercial application, lots of images are released without privacy processing to protect people's identity. In this paper, we develop a novel effective de-identification generative adversarial network (DeIdGAN) for face anonymization by seamlessly replacing a given face image with a different synthesized yet realistic one. Our approach consists of two steps. First, we anonymize the input face to obfuscate its original identity. Then, we use our designed de-identification generator to synthesize an anonymized face. During the training process, we leverage a pair of identity-adversarial discriminators to explicitly constrain identity protection by pushing the synthesized face away from the predefined sensitive faces to resist re-identification and identity invasion. Finally, we validate the effectiveness of our approach on public datasets. Compared with existing methods, our approach can not only achieve better identity protection rates but also preserve superior image quality and data reusability, which suggests the state-of-the-art performance.

Abstract Image classification is one of the mainstream tasks of computer vision. However, the most existing methods use labels of the same granularity level for training. This leads to ignoring the hierarchy that may help to differentiate different visual objects better. Embedding hierarchical information into the convolutional neural networks (CNNs) can effectively regulate the semantic space and thus reduce the ambiguity of prediction. To this end, a multi‐task learning framework, named as Hierarchical Bilinear Convolutional Neural Network (HB‐CNN), is developed by seamlessly integrating CNNs with multi‐task learning over the hierarchical visual concept structures. Specifically, the labels with a tree structure are used as the supervision to hierarchically train multiple branch networks. In this way, the model can not only learn additional information (e.g. context information) as the coarse‐level category features, but also focus the learned fine‐level category features on the object properties. To smoothly pass hierarchical conceptual information and encourage feature reuse, a connectivity pattern is proposed to connect features at different levels. Furthermore, a bilinear module is embedded to generalise various orderless texture feature descriptors so that our model can capture more discriminative features. The proposed method is extensively evaluated on the CIFAR‐10, CIFAR‐100, and ‘Orchid’ Plant image sets. The experimental results show the effectiveness and superiority of our method.

Multi-keyword query is widely supported in text search engines. However, an analogue in image retrieval systems, multi-object query, is rarely studied. Meanwhile, traditional object-based image retrieval methods often involve multiple steps separately. In this work, we propose a weakly-supervised Deep Multiple Instance Hashing (DMIH) approach for multi-object image retrieval. Our DMIH approach, which leverages a popular CNN model to build the end-to-end relation between a raw image and the binary hash codes of its multiple objects, can support multi-object queries effectively and integrate object detection with hashing learning seamlessly. We treat object detection as a binary multiple instance learning (MIL) problem and such instances are automatically extracted from multi-scale convolutional feature maps. We also design a conditional random field (CRF) module to capture both the semantic and spatial relations among different class labels. For hashing training, we sample image pairs to learn their semantic relationships in terms of hash codes of the most probable proposals for owned labels as guided by object predictors. The two objectives benefit each other in a multi-task learning scheme. Finally, a two-level inverted index method is proposed to further speed up the retrieval of multi-object queries. Our DMIH approach outperforms state-of-the-arts on public benchmarks for object-based image retrieval and achieves promising results for multi-object queries.

Multi-view subspace clustering (MVSC) finds a shared structure in latent low-dimensional subspaces of multi-view data to enhance clustering performance. Nonetheless, we observe that most existing MVSC methods neglect the diversity in multi-view data by considering only the common knowledge to find a shared structure either directly or by merging different similarity matrices learned for each view. In the presence of noise, this predefined shared structure becomes a biased representation of the different views. Thus, in this article, we propose a MVSC method based on coupled low-rank representation to address the above limitation. Our method first obtains a low-rank representation for each view, constrained to be a linear combination of the view-specific representation and the shared representation by simultaneously encouraging the sparsity of view-specific one. Then, it uses the k -block diagonal regularizer to learn a manifold recovery matrix for each view through respective low-rank matrices to recover more manifold structures from them. In this way, the proposed method can find an ideal similarity matrix by approximating clustering projection matrices obtained from the recovery structures. Hence, this similarity matrix denotes our clustering structure with exactly k connected components by applying a rank constraint on the similarity matrix’s relaxed Laplacian matrix to avoid spectral post-processing of the low-dimensional embedding matrix. The core of our idea is such that we introduce dynamic approximation into the low-rank representation to allow the clustering structure and the shared representation to guide each other to learn cleaner low-rank matrices that would lead to a better clustering structure. Therefore, our approach is notably different from existing methods in which the local manifold structure of data is captured in advance. Extensive experiments on six benchmark datasets show that our method outperforms 10 similar state-of-the-art compared methods in six evaluation metrics.

In this article, a novel ensemble model, called Multiple Kernel Ensemble Learning (MKEL), is developed by introducing a unified ensemble loss. Different from the previous multiple kernel learning (MKL) methods, which attempt to seek a linear combination of basis kernels as a unified kernel, our MKEL model aims to find multiple solutions in corresponding Reproducing Kernel Hilbert Spaces (RKHSs) simultaneously. To achieve this goal, multiple individual kernel losses are integrated into a unified ensemble loss. Therefore, each model can co-optimize to learn its optimal parameters by minimizing a unified ensemble loss in multiple RKHSs. Furthermore, we apply our proposed ensemble loss into the deep network paradigm and take the sub-network as a kernel mapping from the original input space into a feature space, named Deep-MKEL (D-MKEL). Our D-MKEL model can utilize the diversified deep individual sub-networks into a whole unified network to improve the classification performance. With this unified loss design, our D-MKEL model can make our network much wider than other traditional deep kernel networks and more parameters are learned and optimized. Experimental results on several mediate UCI classification and computer vision datasets demonstrate that our MKEL model can achieve the best classification performance among comparative MKL methods, such as Simple MKL, GMKL, Spicy MKL, and Matrix-Regularized MKL. On the contrary, experimental results on large-scale CIFAR-10 and SVHN datasets concretely show the advantages and potentialities of the proposed D-MKEL approach compared to state-of-the-art deep kernel methods.

Unsupervised domain adaptation (UDA) for person re-identification (re-ID) is a challenging task due to large variations in human classes, illuminations, camera views, and so on. Currently, existing UDA methods focus on two-domain adaptation and are generally trained on one labeled source set and adapted on the other unlabeled target set. In this article, we put forward a new issue on person re-ID, namely, unsupervised multi-target domain adaptation (UMDA). It involves one labeled source set and multiple unlabeled target sets, which is more reasonable for practical real-world applications. Enabling UMDA has to learn the consistency for multiple domains, which is significantly different from the UDA problem. To ensure distribution consistency and learn the discriminative embedding, we further propose the Camera Identity-guided Distribution Consistency method that performs an alignment operation for multiple domains. The camera identities are encoded into the image semantic information to facilitate the adaptation of features. According to our knowledge, this is the first attempt on the unsupervised multi-target domain adaptation learning. Extensive experiments are executed on Market-1501, DukeMTMC-reID, MSMT17, PersonX, and CUHK03, and our method has achieved very competitive re-ID accuracy in multi-target domains against numerous state-of-the-art methods.

To generate an image caption, firstly, the content of the image should be fully understood; and then the semantic information contained in the image should be described using a phrase or statement that conforms to certain grammatical rules. Thus, it requires techniques from both computer vision and natural language processing to connect the two different media forms together, which is highly challenging. To adaptively adjust the effect of visual information and language information on the captioning process, in this paper, the part of speech information is proposed to novelly integrate with image captioning models based on the encoder-decoder framework. First, a part of speech prediction network is proposed to analyze and model the part of speech sequences for the words in natural language sentences; then, different mechanisms are proposed to integrate the part of speech guidance information with merge-based and inject-based image captioning models, respectively; finally, according to the integrated frameworks, a multi-task learning paradigm is proposed to facilitate model training. Experiments are conducted on two widely used image captioning datasets, Flickr30 k and COCO, and the results have validated that the image captions generated by the proposed method contain more accurate visual information and comply with language habits and grammar rules better.

To leverage deep learning for image aesthetics assessment, one critical but unsolved issue is how to seamlessly incorporate the information of image aspect ratios to learn more robust models. In this paper, an adaptive fractional dilated convolution (AFDC), which is aspect-ratio-embedded, composition-preserving and parameter-free, is developed to tackle this issue natively in convolutional kernel level. Specifically, the fractional dilated kernel is adaptively constructed according to the image aspect ratios, where the interpolation of nearest two integer dilated kernels are used to cope with the misalignment of fractional sampling. Moreover, we provide a concise formulation for mini-batch training and utilize a grouping strategy to reduce computational overhead. As a result, it can be easily implemented by common deep learning libraries and plugged into popular CNN architectures in a computation-efficient manner. Our experimental results demonstrate that our proposed method achieves state-of-the-art performance on image aesthetics assessment over the AVA dataset.

The state-of-art 6D object pose detection methods use convolutional neural networks to estimate objects' 6D poses from RGB images. However, they require huge numbers of images with explicit 3D annotations such as 6D poses, 3D bounding boxes and 3D keypoints, either obtained by manual labeling or inferred from synthetic images generated by 3D CAD models. Manual labeling for a large number of images is a laborious task, and we usually do not have the corresponding 3D CAD models of objects in real environment. In this paper, we develop a keypoint-based 6D object pose detection method (and its deep network) called Object Keypoint based POSe Estimation (OK-POSE). OK-POSE employs relative transformation between viewpoints for training. Specifically, we use pairs of images with object annotation and relative transformation information between their viewpoints to automatically discover objects' 3D keypoints which are geometrically and visually consistent. Then, the 6D object pose can be estimated using a keypoint-based geometric reasoning method with a reference viewpoint. The relative transformation information can be easily obtained from any cheap binocular cameras or most smartphone devices, thus greatly lowering the labeling cost. Experiments have demonstrated that OK-POSE achieves acceptable performance compared to methods relying on the object's 3D CAD model or a great deal of 3D labeling. These results show that our method can be used as a suitable alternative when there are no 3D CAD models or a large number of 3D annotations.

Object detection on remote sensing images plays an important role in image analysis and understanding. The generation of anchors is a significant factor due to the diverse variations in object size and clustered small objects in remote sensing images. In this paper, we focus on this issue, and propose an anchor generator with continuously variable scales leading by semantic information and build a Scalable Object Detection Network (SODNet) to detect remote sensing targets. The proposed network exploits semantic information to guide the generation of anchors and improves the target scale variation to adapt to different remote sensing images. We execute extensive experiments on widely employed NWPU VHR-10 dataset and DOTA dataset against numerous state-of-the-art methods, and the experimental results demonstrate the performance of the proposed method, especially on small targets.

People constantly share their photographs with others through various social media sites. With the aid of the privacy settings provided by social media sites, image owners can designate scope of sharing, e.g., close friends and acquaintances. However, even if the owner of a photograph carefully sets the privacy setting to exclude a given individual who is not supposed to see the photograph, the photograph may still eventually reach a wider audience, including those clearly undesired through unanticipated channels of disclosure, causing a privacy breach. Moreover, it is often the case that a given image involves multiple stakeholders who are also depicted in the photograph. Due to various personalities, it is even more challenging to reach agreement on the privacy settings for these multi-owner photographs. In this paper, we propose a privacy risk reminder system, called REMIND, which estimates the probability that a shared photograph may be seen by unwanted people-through the social graph-who are not included in the original sharing list. We tackle this problem from a novel angle by digging into the big data regarding image sharing history. Specifically, the social media providers possess a huge amount of image sharing information (e.g., what photographs are shared with whom) of their users. By analyzing and modeling such rich information, we build a sophisticated probability model that efficiently aggregates the image disclosure probabilities along different possible image propagation chains and loops. If the computed disclosure probability indicates high risks of privacy breach, a reminder is issued to the image owner to help revise the privacy settings (or, at least, inform the user about this accidental disclosure risk). The proposed REMIND system also has a nice feature of policy harmonization that helps resolve privacy differences in multi-owner photographs. We have carried out a user study to validate the rationale of our proposed solutions and also conducted experimental studies to evaluate the efficiency of the proposed REMIND system.

There are two key components that can be leveraged for visual tracking: (a) object appearances; and (b) object motions. Many existing techniques have recently employed deep learning to enhance visual tracking due to its superior representation power and strong learning ability, where most of them employed object appearances but few of them exploited object motions. In this work, a deep spatial and temporal network (DSTN) is developed for visual tracking by explicitly exploiting both the object representations from each frame and their dynamics along multiple frames in a video, such that it can seamlessly integrate the object appearances with their motions to produce compact object appearances and capture their temporal variations effectively. Our DSTN method, which is deployed into a tracking pipeline in a coarse-to-fine form, can perceive the subtle differences on spatial and temporal variations of the target (object being tracked), and thus it benefits from both off-line training and online fine-tuning. We have also conducted our experiments over four largest tracking benchmarks, including OTB-2013, OTB-2015, VOT2015, and VOT2017, and our experimental results have demonstrated that our DSTN method can achieve competitive performance as compared with the state-of-the-art techniques. The source code, trained models, and all the experimental results of this work will be made public available to facilitate further studies on this problem.

In multi-task learning, multiple interrelated tasks are jointly learned to achieve better performance. In many cases, if we can identify which tasks are related, we can also clearly identify which tasks are unrelated. In the past, most researchers emphasized exploiting correlations among interrelated tasks while completely ignoring the unrelated tasks that may provide valuable prior knowledge for multi-task learning. In this paper, a new approach is developed to hierarchically learn a tree of multi-task metrics by leveraging prior knowledge about both the related tasks and unrelated tasks. First, a visual tree is constructed to hierarchically organize large numbers of image categories in a coarse-to-fine fashion. Over the visual tree, a multi-task metric classifier is learned for each node by exploiting both the related and unrelated tasks, where the learning tasks for training the classifiers for the sibling child nodes under the same parent node are treated as the interrelated tasks, and the others are treated as the unrelated tasks. In addition, the node-specific metric for the parent node is propagated to its sibling child nodes to control inter-level error propagation. Our experimental results demonstrate that our hierarchical metric learning algorithm achieves better results than other state-of-the-art algorithms.

In this article, a discriminative fast hierarchical learning algorithm is developed for supporting multiclass image classification, where a visual tree is seamlessly integrated with multitask learning to achieve fast training of the tree classifier hierarchically (i.e., a set of structural node classifiers over the visual tree). By partitioning a large number of categories hierarchically in a coarse-to-fine fashion, a visual tree is first constructed and further used to handle data imbalance and identify the interrelated learning tasks automatically (e.g., the tasks for learning the node classifiers for the sibling child nodes under the same parent node are strongly interrelated), and a multitask SVM classifier is trained for each nonleaf node to achieve more effective separation of its sibling child nodes at the next level of the visual tree. Both the internode visual similarities and the interlevel visual correlations are utilized to train more discriminative multitask SVM classifiers and control the interlevel error propagation effectively, and a stochastic gradient descent (SGD) algorithm is developed for learning such multitask SVM classifiers with higher efficiency. Our experimental results have demonstrated that our fast hierarchical learning algorithm can achieve very competitive results on both the classification accuracy rates and the computational efficiency.

Most of the existing methods for anomaly detection use only positive data to learn the data distribution, thus they usually need a pre-defined threshold at the detection stage to determine whether a test instance is an outlier. Unfortunately, a good threshold is vital for the performance and it is really hard to find an optimal one. In this paper, we take the discriminative information implied in unlabeled data into consideration and propose a new method for anomaly detection that can learn the labels of unlabelled data directly. Our proposed method has an end-to-end architecture with one encoder and two decoders that are trained to model inliers and outliers’ data distributions in a competitive way. This architecture works in a discriminative manner without suffering from overfitting, and the training algorithm of our model is adopted from SGD, thus it is efficient and scalable even for large-scale datasets. Empirical studies on 7 datasets including KDD99, MNIST, Caltech-256, and ImageNet etc. show that our model outperforms the state-of-the-art methods.

In this paper, a deep embedding algorithm is developed to achieve higher accuracy rates on large-scale image classification. By adapting the importance of the object classes to their error rates, our deep embedding algorithm can train multiple complementary deep networks sequentially, where each of them focuses on achieving higher accuracy rates for different subsets of object classes in an easy-to-hard way. By integrating such complementary deep networks to generate an ensemble network, our deep embedding algorithm can improve the accuracy rates for the hard object classes (which initially have higher error rates) at certain degrees while effectively preserving high accuracy rates for the easy object classes. Our deep embedding algorithm has achieved higher overall accuracy rates on large scale image classification.

The image fusion method based on sparse representation in the single‐scale image domain has produced better fusion results than the classic methods based on multi‐scale analysis nowadays. However, due to the limited number of dictionary atoms, it is difficult to provide an accurate description for image details in the sparse‐representation‐based image fusion methods, and it requires a lot of time. A novel dictionary is constructed with non‐subsampled contourlet transform and sparse representation by using the proposed simultaneous strategy. Then the novel dictionary could combine the sparsity attribute of the learning dictionary with a multi‐scale feature of non‐subsampled contourlet transform. Moreover, the simultaneous strategy is combined with this novel dictionary so that sparse coefficients can be represented with the same dictionary atoms and thus they can be compared in a reasonable and accurate way. Finally, the image fusion method along with this novel dictionary is proposed and named non‐subsampled contourlet transform (NSCT)–simultaneous sparse representation (SSR). Experimental results show that the proposed fusion method NSCT–SSR, with its more excellent fusion effect and better anti‐noise capability, outperforms the existing fusion methods, which are based on both multi‐scale domain and sparse representation in the single‐scale image domain.

For remote sensing image fusion, infrared and visible images have very different brightness due to their disparate imaging mechanisms, the result of which is that nontarget regions in the infrared image often affect the fusion of details in the visible image. This letter proposes a novel infrared and visible image fusion method basing hybrid representation learning by combining dictionary-learning-based joint sparse representation (JSR) and nonnegative sparse representation (NNSR). In the proposed method, different fusion strategies are adopted, respectively, for the mean image, which represents the primary energy information, and for the deaveraged image, which contains important detail features. Since the deaveraged image contains a large amount of high-frequency details information of the source image, JSR is utilized to sparsely and accurately extract the common and innovation features of the deaveraged image, thus, accurately merging high-frequency details in the deaveraged image. Then, the mean image represents low-frequency and overview features of the source image, according to NNSR, mean image is classified well-directed to different feature regions and then fused, respectively. Such proposed method, on the one hand, can eliminate the impact on fusion result suffering from very different brightness causing by different imaging mechanism between infrared and visible image; on the other hand, it can improve the readability and accuracy of the result fusion image. Experimental result shows that, compared with the classical and state-of-the-art fusion methods, the proposed method not only can accurately integrate the infrared target but also has rich background details of the visible image, and the fusion effect is superior.

In this paper, a deep mixture of diverse experts algorithm is developed to achieve more efficient learning of a huge (mixture) network for large-scale visual recognition application. First, a two-layer ontology is constructed to assign large numbers of atomic object classes into a set of task groups according to the similarities of their learning complexities, where certain degrees of inter-group task overlapping are allowed to enable sufficient inter-group message passing. Second, one particular base deep CNNs with M+1 outputs is learned for each task group to recognize its M atomic object classes and identify one special class of "not-in-group", where the network structure (numbers of layers and units in each layer) of the well-designed deep CNNs (such as AlexNet, VGG, GoogleNet, ResNet) is directly used to configure such base deep CNNs. For enhancing the separability of the atomic object classes in the same task group, two approaches are developed to learn more discriminative base deep CNNs: (a) our deep multi-task learning algorithm that can effectively exploit the inter-class visual similarities; (b) our two-layer network cascade approach that can improve the accuracy rates for the hard object classes at certain degrees while effectively maintaining the high accuracy rates for the easy ones. Finally, all these complementary base deep CNNs with diverse but overlapped outputs are seamlessly combined to generate a mixture network with larger outputs for recognizing tens of thousands of atomic object classes. Our experimental results have demonstrated that our deep mixture of diverse experts algorithm can achieve very competitive results on large-scale visual recognition.

The applications of isometric 3-D objects have recently received sufficient attention and, thus, it is very attractive to retrieve such isometric 3-D objects from large-scale collections. Although existing approaches have presented some interesting ideas, their performance is limited to their ability on feature representation. To improve the performance of 3-D object (shape) recognition, some recent algorithms prefer using complicated deep neural networks to learn discriminative features, but they consume huge amounts of computing resources. Instead, this paper presents a more effective solution by seamlessly integrating the traditional local descriptor with a deep pointwise convolutional network to extract 1-D features for shape recognition and retrieval. To reduce the costs of designing a complicated deep network, the first step of our algorithm is to describe the shape deformation by sampling a set of intrinsic point descriptors. Then, we introduce a simple yet effective pointwise convolutional network to integrate these descriptors as a global feature and the learning process can be significantly accelerated with the help of downsampling. Furthermore, a knowledge transfer strategy is used to upgrade our feature by compensating for information loss. Finally, we carry out experimental evaluations over popular shape benchmarks, and the results suggest that our approach exhibits superior accuracy rates and robustness on shape recognition and retrieval.

As the amount of available video data has grown substantially, automatic video classification has become an urgent yet challenging task. Most video classification methods focus on acquiring discriminative spacial visual features and motion patterns for video representation, especially deep learning methods, which have achieved very good results on action recognition problems. However, the performance of most of these methods drastically degenerates for more generic video classification tasks where the video contents are much more complex. Thus, in this paper, the mid-level semantics of videos are exploited to bridge the semantic gap between low-level features and high-level video semantics. Inspired by the term ``frequency-inverse document frequency'', a word weighting method for the problem of text classification is introduced to the video domain. The visual objects in videos are regarded as the words in texts, and two new weighting methods are proposed to encode videos by weighting visual objects according to the characteristics of videos. In addition, the semantic similarities between video categories and visual objects are introduced from the text domain as privileged information to facilitate classifier training on the obtained semantic representations of videos. The proposed semantic encoding method (semantic stream) is then fused with the popular two-stream CNN model for the final classification results. Experiments are conducted on two large-scale complex video datasets, CCV and ActivityNet. The experimental results validate the effectiveness of the proposed methods.

With the increasing popularity of 3D models, retrieving deformable 3D objects is becoming a crucial task. The state-of-the-art methods use complex deep neural networks to address this problem, which require lots of computational resources. In this paper, we develop a more effective solution by using point convolution. Our algorithm takes local point descriptors as the input and produces a global vector for shape retrieval. To save the efforts of designing complex deep convolutional neural network (CNN), we first use intrinsic point descriptors to describe the shape deformations. Then, a simple but effective point CNN network is developed to integrate the local shape information by performing subspace compression and fusion, which depends on an end-to-end learning process to link the local and global information for discriminative shape representation. The experimental results on popular benchmarks have verified that our algorithm is able to outperform the state-of-the-art methods.

Automatically classifying large scale of video data is an urgent yet challenging task. To bridge the semantic gap between low-level features and high-level video semantics, we propose a method to represent videos with their mid-level semantics. Inspired by the problem of text classification, we regard the visual objects in videos as the words in documents, and adapt the TF-IDF word weighting method to encode videos by visual objects. Some extensions upon the proposed method are also made according to the characteristics of videos. We integrate the proposed semantic encoding method with the popular two-stream CNN model for video classification. Experiments are conducted on two large-scale video datasets, CCV and ActivityNet. The experimanetal results validates the effectiveness of our method.

To configure successful privacy settings for social image sharing, two issues are inseparable: 1) content sensitiveness of the images being shared; and 2) trustworthiness of the users being granted to see the images. This paper aims to consider these two inseparable issues simultaneously to recommend fine-grained privacy settings for social image sharing. For achieving more compact representation of image content sensitiveness (privacy), two approaches are developed: 1) a deep network is adapted to extract 1024-D discriminative deep features; and 2) a deep multiple instance learning algorithm is adopted to identify 280 privacy-sensitive object classes and events. Second, users on the social network are clustered into a set of representative social groups to generate a discriminative dictionary for user trustworthiness characterization. Finally, both the image content sensitiveness and the user trustworthiness are integrated to train a tree classifier to recommend fine-grained privacy settings for social image sharing. Our experimental studies have demonstrated both the efficiency and the effectiveness of our proposed algorithms.

In this paper, a layer-wise mixture model (LMM) is developed to support hierarchical visual recognition, where a Bayesian approach is used to automatically adapt the visual hierarchy to the progressive improvements of the deep network along the time. Our LMM algorithm can provide an end-to-end approach for jointly learning: (a) the deep network for achieving more discriminative deep representations for object classes and their inter-class visual similarities; (b) the tree classifier for recognizing large numbers of object classes hierarchically; and (c) the visual hierarchy adaptation for achieving more accurate assignment and organization of large numbers of object classes. By learning the tree classifier, the deep network and the visual hierarchy adaptation jointly in an end-to-end manner, our LMM algorithm can achieve higher accuracy rates on hierarchical visual recognition. Our experiments are carried on ImageNet1K and ImageNet10K image sets, which have demonstrated that our LMM algorithm can achieve very competitive results on the accuracy rates as compared with the baseline methods.

Visual question answering (VQA) is challenging, because it requires a simultaneous understanding of both visual content of images and textual content of questions. To support the VQA task, we need to find good solutions for the following three issues: 1) fine-grained feature representations for both the image and the question; 2) multimodal feature fusion that is able to capture the complex interactions between multimodal features; and 3) automatic answer prediction that is able to consider the complex correlations between multiple diverse answers for the same question. For fine-grained image and question representations, a "coattention" mechanism is developed using a deep neural network (DNN) architecture to jointly learn the attentions for both the image and the question, which can allow us to reduce the irrelevant features effectively and obtain more discriminative features for image and question representations. For multimodal feature fusion, a generalized multimodal factorized high-order pooling approach (MFH) is developed to achieve more effective fusion of multimodal features by exploiting their correlations sufficiently, which can further result in superior VQA performance as compared with the state-of-the-art approaches. For answer prediction, the Kullback-Leibler divergence is used as the loss function to achieve precise characterization of the complex correlations between multiple diverse answers with the same or similar meaning, which can allow us to achieve faster convergence rate and obtain slightly better accuracy on answer prediction. A DNN architecture is designed to integrate all these aforementioned modules into a unified model for achieving superior VQA performance. With an ensemble of our MFH models, we achieve the state-of-the-art performance on the large-scale VQA data sets and win the runner-up in VQA Challenge 2017.

Visual question answering (VQA) is challenging because it requires a simultaneous understanding of both the visual content of images and the textual content of questions. The approaches used to represent the images and questions in a fine-grained manner and questions and to fuse these multimodal features play key roles in performance. Bilinear pooling based models have been shown to outperform traditional linear models for VQA, but their high-dimensional representations and high computational complexity may seriously limit their applicability in practice. For multimodal feature fusion, here we develop a Multi-modal Factorized Bilinear (MFB) pooling approach to efficiently and effectively combine multi-modal features, which results in superior performance for VQA compared with other bilinear pooling approaches. For fine-grained image and question representation, we develop a `co-attention' mechanism using an end-to-end deep network architecture to jointly learn both the image and question attentions. Combining the proposed MFB approach with co-attention learning in a new network architecture provides a unified model for VQA. Our experimental results demonstrate that the single MFB with co-attention model achieves new state-of-theart performance on the real-world VQA dataset. Code available at https://github.com/yuzcccc/mfb.

Bilinear convolutional neural networks (BCNN) model, the state-of-the-art in fine-grained image recognition, fails in distinguishing the categories with subtle visual differences. We design a novel BCNN model guided by user click data (C-BCNN) to improve the performance via capturing both the visual and semantical content in images. Specially, to deal with the heavy noise in large-scale click data, we propose a weakly supervised learning approach to learn the C-BCNN, namely W-C-BCNN. It can automatically weight the training images based on their reliability. Extensive experiments are conducted on the public Clickture-Dog dataset. It shows that: (1) integrating CNN with click feature largely improves the performance; (2) both the click data and visual consistency can help to model image reliability. Moreover, the method can be easily customized to medical image recognition. Our model performs much better than conventional BCNN models on both the Clickture-Dog and medical image dataset.

In this paper, a deep mixture algorithm is developed to support large-scale visual recognition (e.g., recognizing tens of thousands of object classes) by seamlessly combining a set of base deep CNNs (AlexNet) with diverse task spaces, e.g., such base deep CNNs (i.e., diverse experts) are trained to recognize different subsets of tens of thousands of object classes rather than the same set of object classes. Our experimental results have demonstrated that our deep mixture algorithm can achieve very competitive results on large-scale visual recognition.

This paper aims to simultaneously consider two inseparable issues for privacy setting recommendation: (1) sensitiveness of visual content of the images being shared; and (2) trustworthiness of users being granted. First, an object-based approach is developed for image content sensitiveness (privacy) representation. Secondly, the users on a social network are clustered into a set of representative social groups to generate a discriminative dictionary for user trustworthiness characterization. Finally, a tree classifier is trained hierarchically to recommend appropriate privacy settings for image sharing.

Multi-keyword query is widely supported in text search engines. However, an analogue in image retrieval systems, multi-object query, is rarely studied. Meanwhile, traditional object-based image retrieval methods often involve multiple steps separately and need expensive location labeling for detecting objects. In this work, we propose a weakly-supervised Deep Multiple Instance Hashing (DMIH) framework for object-based image retrieval. DMIH integrates object detection and hashing learning on the basis of a popular CNN model to build the end-to-end relation between a raw image and the binary hashing codes of multiple objects in it. Specifically, we cast the object detection of each object class as a binary multiple instance learning problem where instances are object proposals extracted from multi-scale convolutional feature maps. For hashing training, we sample image pairs to learn their semantic relationships in terms of hash codes of the most probable proposals for owned labels as guided by object predictors. The two objectives benefit each other in learning. DMIH outperforms state-of-the-arts on public benchmarks for object-based image retrieval and achieves promising results for multi-object queries.

To achieve automatic recommendation of privacy settings for image sharing, a new tool called iPrivacy (image privacy) is developed for releasing the burden from users on setting the privacy preferences when they share their images for special moments. Specifically, this paper consists of the following contributions: 1) massive social images and their privacy settings are leveraged to learn the object-privacy relatedness effectively and identify a set of privacy-sensitive object classes automatically; 2) a deep multi-task learning algorithm is developed to jointly learn more representative deep convolutional neural networks and more discriminative tree classifier, so that we can achieve fast and accurate detection of large numbers of privacy-sensitive object classes; 3) automatic recommendation of privacy settings for image sharing can be achieved by detecting the underlying privacy-sensitive objects from the images being shared, recognizing their classes, and identifying their privacy settings according to the object-privacy relatedness; and 4) one simple solution for image privacy protection is provided by blurring the privacy-sensitive objects automatically. We have conducted extensive experimental studies on real-world images and the results have demonstrated both the efficiency and effectiveness of our proposed approach.

In this paper, a hierarchical deep multi-task learning (HD-MTL) algorithm is developed to support large-scale visual recognition (e.g., recognizing thousands or even tens of thousands of atomic object classes automatically). First, multiple sets of multi-level deep features are extracted from different layers of deep convolutional neural networks (deep CNNs), and they are used to achieve more effective accomplishment of the coarseto- fine tasks for hierarchical visual recognition. A visual tree is then learned by assigning the visually-similar atomic object classes with similar learning complexities into the same group, which can provide a good environment for determining the interrelated learning tasks automatically. By leveraging the inter-task relatedness (inter-class similarities) to learn more discriminative group-specific deep representations, our deep multi-task learning algorithm can train more discriminative node classifiers for distinguishing the visually-similar atomic object classes effectively. Our hierarchical deep multi-task learning (HD-MTL) algorithm can integrate two discriminative regularization terms to control the inter-level error propagation effectively, and it can provide an end-to-end approach for jointly learning more representative deep CNNs (for image representation) and more discriminative tree classifier (for large-scale visual recognition) and updating them simultaneously. Our incremental deep learning algorithms can effectively adapt both the deep CNNs and the tree classifier to the new training images and the new object classes. Our experimental results have demonstrated that our HD-MTL algorithm can achieve very competitive results on improving the accuracy rates for large-scale visual recognition.

The manifold of the dataset turns out to be quite useful in refining the retrieval results, and the diffusion process provides an efficient solution by careful selection of the similarity neighborhood which is usually modeled as the K-nearest neighborhood (KNN) graph. However, existing works are sensitive to the topology noises induced by the first K neighbors. In this paper, we tackle the problem by studying metric transformation which aims at finding new functional relationship to dig the latent similarity. The advantage of the approach lies in its robustness towards the varying K values; that is to say, it could preserve high similarity performances even if K is very large. Except for discussing only the global KNN (i.e. the same K for all neighborhoods) graph, we also investigate to specify a different K for each neighborhood by incorporating the new penalized consensus information (PCI). We show that PCI works superior compared with the original consensus information for denoising. Experiments on multiple affinity matrices have corroborated the superiority of our method with surprising good results.

In this paper, an automatic image-text alignment algorithm is developed for achieving more accurate indexing and retrieval of large-scale web images. First, large-scale web pages are crawled, where the informative images and their most relevant auxiliary text blocks are extracted. Second, parallel image clustering is performed to partition large-scale informative web images into a large number of clusters. By grouping the visually-similar (near-duplicate) web images into the same cluster, our parallel image clustering algorithm can significantly reduce the huge uncertainty on the relatedness between the web images and their auxiliary text terms, which can provide a good starting point for supporting automatic image-text alignment. Finally, a relevance re-ranking algorithm is developed to identify the most relevant visual text terms for the visually-similar web images in the same cluster. Our experiments on large-scale web images have obtained very positive results.

In this paper, we aim at improving the discriminative jointly dictionaries for large-scale image classification. Sparse representation is a popular tool for image classification. Visual dictionary is very critical to the classification performance. A visual tree is constructed according to the visual similarity, in which the higher layer represents the coarser membership and the lower layer represents the finer membership. Jointly dictionary is learned according to the visual tree. Bregman iterative algorithm is implemented to solve the optimal problem of joint dictionary learning, which makes the solution accurate and the running speed fast. Furthermore, we try to implement the pre-trained features learned from the convolution neural network (CNN) to represent an image, and the residual error of the sparse representation is utilized for image classification. The experimental results demonstrate that the CNN feature is more distinct than SIFT, and the hierarchical classification framework with the Bregman iteration algorithm can greatly improve the performance of classification.

In this paper, a hierarchical multi-task structural learning algorithm is developed to support large-scale plant species identification, where a visual tree is constructed for organizing large numbers of plant species in a coarse-to-fine fashion and determining the inter-related learning tasks automatically. For a given parent node on the visual tree, it contains a set of sibling coarse-grained categories of plant species or sibling fine-grained plant species, and a multi-task structural learning algorithm is developed to train their inter-related classifiers jointly for enhancing their discrimination power. The inter-level relationship constraint, e.g., a plant image must first be assigned to a parent node (high-level non-leaf node) correctly if it can further be assigned to the most relevant child node (low-level non-leaf node or leaf node) on the visual tree, is formally defined and leveraged to learn more discriminative tree classifiers over the visual tree. Our experimental results have demonstrated the effectiveness of our hierarchical multi-task structural learning algorithm on training more discriminative tree classifiers for large-scale plant species identification.

In this paper, a hierarchical learning algorithm is developed for classifying large-scale patient records, e.g., categorizing large-scale patient records into large numbers of known patient categories (i.e., thousands of known patient categories) for automatic treatment stratification. Our hierarchical learning algorithm can leverage tree structure to train more discriminative max-margin classifiers for high-level nodes and control interlevel error propagation effectively. By ruling out unlikely groups of patient categories (i.e., irrelevant high-level nodes) at an early stage, our hierarchical approach can achieve log-linear computational complexity, which is very attractive for big data applications. Our experiments on one specific medical domain have demonstrated that our hierarchical approach can achieve very competitive results on both classification accuracy and computational efficiency as compared with other state-of-the-art techniques.

Real-world datasets often involve multiple views of data items, e.g., a Web page can be described by both its content and anchor texts of hyperlinks leading to it; photos in Flickr could be characterized by visual features, as well as user contributed tags. Different views provide information complementary to each other. Synthesizing multi-view features can lead to a comprehensive description of the data items, which could benefit many data analytic applications. Unfortunately, the simple idea of concatenating different feature vectors ignores statistical properties of each view and usually incurs the “curse of dimensionality” problem. We propose Multi-view Concept Learning (MCL), a novel nonnegative latent representation learning algorithm for capturing conceptual factors from multi-view data. MCL exploits both multi-view information and label information. The key idea is to learn a common latent space across different views which (1) captures the semantic relationships between data items through graph embedding regularization on labeled items, and (2) allows each latent factor to be associated with a subset of views via sparseness constraints. In this way, MCL could capture flexible conceptual patterns hidden in multi-view features. Experiments on a toy problem and three real-world datasets show that MCL performs well and outperforms baseline methods.

This paper addresses a general feature indexing and retrieval scenario in which a set of features detected in the image can retrieve a relevant class of objects, or classes of objects. The main idea behind those features for general object retrieval is that they are capable of identifying and localizing some small regions or parts of the potential object. We propose a set of criteria which take advantage of the learned features to find regions in the image which likely belong to an object. We further use the features' localization capability to localize the full object of interest and its extents. The proposed approach improves the recognition performance and is very efficient. Moreover, it has the potential to be used in automatic image understanding or annotation since it can uncover regions where the objects can be found in an image.

Learning discriminative dictionaries for image content representation plays a critical role in visual recognition. In this paper, we present a joint dictionary learning (JDL) algorithm which exploits the inter-category visual correlations to learn more discriminative dictionaries. Given a group of visually correlated categories, JDL simultaneously learns one common dictionary and multiple category-specific dictionaries to explicitly separate the shared visual atoms from the category-specific ones. The problem of JDL is formulated as a joint optimization with a discrimination promotion term according to the Fisher discrimination criterion. A visual tree method is developed to cluster a large number of categories into a set of disjoint groups, so that each of them contains a reasonable number of visually correlated categories. The process of image category clustering helps JDL to learn better dictionaries for classification by ensuring that the categories in the same group are of strong visual correlations. Also, it makes JDL to be computationally affordable in large-scale applications. Three classification schemes are adopted to make full use of the dictionaries learned by JDL for visual content representation in the task of image categorization. The effectiveness of the proposed algorithms has been evaluated using two image databases containing 17 and 1,000 categories, respectively.

In this paper, a social health support system is developed to assist both ACL (anterior cruciate ligament) patients and clinicians on making better decisions and choices for ACL reconstruction and rehabilitation. By providing a good platform to enable more effective sharing of personal expertise and ACL treatments, our social health support system can allow: (1) ACL patients to identify the best-matching social groups and locate the most suitable expertise for personal health management; and (2) clinicians to easily locate the best-matching ACL patients and learn from well-done treatments, so that they can make better decisions for new ACL patients (who have similar ACL injuries and close social principles with those best-matching ACL patients) and prescribe safer and more effective knee rehabilitation treatments.

In this paper, a novel framework is developed to achieve effective summarization of large-scale image collection by treating the problem of automatic image summarization as the problem of dictionary learning for sparse representation, e.g., the summarization task can be treated as a dictionary learning task (i.e., the given image set can be reconstructed sparsely with this dictionary). For image set of a specific category or a mixture of multiple categories, we have built a sparsity model to reconstruct all its images by using a subset of most representative images (i.e., image summary); and we adopted the simulated annealing algorithm to learn such sparse dictionary by minimizing an explicit optimization function. By investigating their reconstruction ability under sparsity constrain and diversity constrain, we have quantitatively measure the performance of various summarization algorithms. Our experimental results have shown that our dictionary learning for sparse representation algorithm can obtain more accurate summary as compared with other baseline algorithms.

Object recognition is challenging especially when the objects from different categories are visually similar to each other. In this paper, we present a novel joint dictionary learning (JDL) algorithm to exploit the visual correlation within a group of visually similar object categories for dictionary learning where a commonly shared dictionary and multiple category-specific dictionaries are accordingly modeled. To enhance the discrimination of the dictionaries, the dictionary learning problem is formulated as a joint optimization by adding a discriminative term on the principle of the Fisher discrimination criterion. As well as presenting the JDL model, a classification scheme is developed to better take advantage of the multiple dictionaries that have been trained. The effectiveness of the proposed algorithm has been evaluated on popular visual benchmarks.

Research of graffiti character recognition and retrieval, as a branch of traditional optical character recognition (OCR), has started to gain attention in recent years. We have investigated the special challenge of the graffiti image retrieval problem and propose a series of novel techniques to overcome the challenges. The proposed bounding box framework locates the character components in the graffiti images to construct meaningful character strings and conduct image-wise and semantic-wise retrieval on the strings rather than the entire image. Using real world data provided by the law enforcement community to the Pacific Northwest National Laboratory, we show that the proposed framework outperforms the traditional image retrieval framework with better retrieval results and improved computational efficiency.

In this paper, a novel data-driven algorithm is developed for achieving quantitative characterization of the semantic gaps directly in the visual feature space, where the visual feature space is the common space for concept classifier training and automatic concept detection. By supporting quantitative characterization of the semantic gaps, more effective inference models can automatically be selected for concept classifier training by: (1) identifying the image concepts with small semantic gaps (i.e., the isolated image concepts with high inner-concept visual consistency) and training their one-against-all SVM concept classifiers independently; (2) determining the image concepts with large semantic gaps (i.e., the visually-related image concepts with low inner-concept visual consistency) and training their inter-related SVM concept classifiers jointly; and (3) using more image instances to achieve more reliable training of the concept classifiers for the image concepts with large semantic gaps. Our experimental results on NUS-WIDE and ImageNet image sets have obtained very promising results.

In this paper, each image is viewed as a bag of local regions, as well as it is investigated globally. A novel method is developed for achieving multi-label multi-instance image annotation, where image-level (bag-level) labels and region-level (instance-level) labels are both obtained. The associations between semantic concepts and visual features are mined both at the image level and at the region level. Inter-label correlations are captured by a co-occurence matrix of concept pairs. The cross-level label coherence encodes the consistency between the labels at the image level and the labels at the region level. The associations between visual features and semantic concepts, the correlations among the multiple labels, and the cross-level label coherence are sufficiently leveraged to improve annotation performance. Structural max-margin technique is used to formulate the proposed model and multiple interrelated classifiers are learned jointly. To leverage the available image-level labeled samples for the model training, the region-level label identification on the training set is firstly accomplished by building the correspondences between the multiple bag-level labels and the image regions. JEC distance based kernels are employed to measure the similarities both between images and between regions. Experimental results on real image datasets MSRC and Corel demonstrate the effectiveness of our method.

Recently, video sharing services such as YouTube and Daily-motion have become popular and many videos taken with mobile video cameras are uploaded to such a video sharing service. However, such videos can infringe on the privacy right of people in the videos because they may contain privacy sensitive information (PSI) of the people, i.e., their appearances. This strongly motivates us to develop a technique to generate privacy-protected videos. In this paper, we propose a novel system for automatic generation of privacy-protected videos based on background estimation. In most conventional techniques, objects that contain PSI are detected and obscured by, e.g., blurring. Conversely, in our system, background pixels are estimated and then substituted with intended human objects that are essential for the camera person's capture intention. We quantitatively evaluate our system to demonstrate its potential applicability.

To support more precise query translation for English-Chinese Bi-Directional Cross-Language Information Retrieval (CLIR), we have developed a novel framework by integrating a semantic network to characterize the correlations between multiple inter-related text terms of interest and learn their inter-related statistical query translation models. First, a semantic network is automatically generated from large-scale English-Chinese bilingual parallel corpora to characterize the correlations between a large number of text terms of interest. Second, the semantic network is exploited to learn the statistical query translation models for such text terms of interest. Finally, these inter-related query translation models are used to translate the queries more precisely and achieve more effective CLIR. Our experiments on a large number of official public data have obtained very positive results.

In this paper, we present a novel framework to achieve effective summarization of large-scale web images by treating the problem of automatic image summarization as the problem of dictionary learning for sparse coding, e.g., the summary of a given image set can be treated as a sparse representation of the given image set (i.e., sparse dictionary for the given image set). For a given semantic category (i.e., certain object class or image concept), we build a sparsity model to reconstruct all its relevant images by using a subset of most representative images (i.e., image summary); and a stepwise basis selection algorithm is developed to learn such sparse dictionary (i.e., image summary) by minimizing an explicit optimization function. By investigating their reconstruction ability, the reconstruction Mean Square Error (MSE) is adapted to objectively measure the performance of various algorithms for automatic image summarization. Our experimental results demonstrate that our dictionary learning for sparse representation algorithm can obtain more accurate summary as compared with other baseline algorithms for automatic image summarization.

It is very attractive to exploit weakly-labeled image dataset for multi-label annotation applications. In our paper the meaning of the terminology weakly labeled is threefold: i) only a small subset of the available images are labeled; ii) even for the labeled image, the given labels may be uncorrect or incomplete; iii) the given labels do not provide the exact object locations in the images. A novel method is developed to predict the multiple labels for images and to provide region-level labels for the objects. We cluster the image regions to learn several region-exemplars and predict the label vector for each image region as a locally weighted average of the label vectors on exemplars. By investigating the label confidence matrix for the region-exemplars from different perspectives (column picture and row picture), we sufficiently leverage the visual contexts, the semantic contexts, and the consistency between similarities in the visual feature space and semantic label space. Experimental results on real web images demonstrate the effectiveness of the proposed method.

In this paper, a structured max-margin learning algorithm is developed to achieve more effective training of a large number of inter-related classifiers for multilabel image annotation application. To leverage multilabel images for classifier training, each multilabel image is partitioned into a set of image instances (image regions or image patches) and an automatic instance label identification algorithm is developed to assign multiple labels (which are given at the image level) to the most relevant image instances. A K-way min-max cut algorithm is developed for automatic instance clustering and kernel weight determination, where multiple base kernels are seamlessly combined to address the issue of huge intra-concept visual diversity more effectively. Second, a visual concept network is constructed for characterizing the inter-concept visual similarity contexts more precisely in the high-dimensional multimodal feature space. The visual concept network is used to determine the inter-related learning tasks directly in the feature space rather than in the label space because feature space is the common space for classifier training and image classification. Third, a parallel computing platform is developed to achieve more effective learning of a large number of inter-related classifiers over the visual concept network. A structured max-margin learning algorithm is developed by incorporating the visual concept network, max-margin Markov networks and multitask learning to address the issue of huge inter-concept visual similarity more effectively. By leveraging the inter-concept visual similarity contexts for inter-related classifier training, our structured max-margin learning algorithm can significantly enhance the discrimination power of the inter-related classifiers. Our experiments have also obtained very positive results for a large number of object classes and image concepts.

To leverage large-scale weakly-tagged images for computer vision tasks (such as object detection and scene recognition), a novel cross-modal tag cleansing and junk image filtering algorithm is developed for cleansing the weakly-tagged images and their social tags (i.e., removing irrelevant images and finding the most relevant tags for each image) by integrating both the visual similarity contexts between the images and the semantic similarity contexts between their tags. Our algorithm can address the issues of spams, polysemes and synonyms more effectively and determine the relevance between the images and their social tags more precisely, thus it can allow us to create large amounts of training images with more reliable labels by harvesting from large-scale weakly-tagged images, which can further be used to achieve more effective classifier training for many computer vision tasks.

The growing popularity of video sharing services such as YouTube enables us to upload and share consumer generated videos (CGVs) easily, resulting in disclosure of the privacy sensitive information (PSI) of persons, i.e., their appearances. Therefore, we need a technique for automatically protecting the privacy in CGVs; however, the main problem is how to determine PSI regions automatically. In this paper, we propose a novel system for automatically protecting the privacy in CGVs. The proposed system tackles the problem of determining PSI regions by using an intended human object detector that detects human objects which the camera person wanted to capture to achieve his/her capture intention. In addition, the proposed system adopts several PSI obscuring methods such as blocking out, blurring and seam carving. We present the results of subjective evaluations of a privacy protected video in terms of the visual quality and acceptability of PSI disclosure, as well as the performance of the intended human object detector.

Large-scale loosely-tagged images (i.e., multiple object tags are given loosely at the image level) are available on Internet, and it is very attractive to leverage such loosely-tagged images for automatic image annotation applications. In this paper, a multi-task structured SVM algorithm is developed to leverage both the inter-object correlations and the loosely-tagged images for achieving more effective training of a large number of inter-related object classifiers. To leverage the loosely-tagged images for object classifier training, each loosely-tagged image is partitioned into a set of image instances (image regions) and a multiple instance learning algorithm is developed for instance label identification by automatically identifying the correspondences between multiple tags (given at the image level) and the image instances. An object correlation network is constructed for characterizing the inter-object correlations explicitly and identifying the inter-related learning tasks automatically. To enhance the discrimination power of a large number of inter-related object classifiers, a multi-task structured SVM algorithm is developed to model the inter-task relatedness more precisely and leverage the inter-object correlations for classifier training. Our experiments on a large number of inter-related object classes have provided very positive results.

For automatic object detection tasks, large amounts of training images are usually labeled to achieve more reliable training of the object classifiers; this is cost-expensive since it requires hiring professionals to label large-scale training images. When a large number of object classes come into view, the issue of obtaining a large enough amount of the labeled training images becomes more critical. There are three potential solutions to reduce the burden for image labeling: (1) allowing people to provide the object labels loosely at the image level rather than at the object level (e.g., loosely-tagged images without identifying the exact object locations in the images); (2) harnessing large-scale collaboratively-tagged images that are available on the Internet; and, (3) developing new machine learning algorithms that can directly leverage large-scale collaboratively- or loosely-tagged images for achieving more effective training of a large number of object classifiers. Based on these observations, a multi-task multi-label multiple instance learning (MTML-MIL) algorithm is developed in this paper by leveraging both interobject correlations and large-scale loosely-labeled images for object classifier training. By seamlessly integrating multi-task learning, multi-label learning, and multiple instance learning, our MTML-MIL algorithm can achieve more accurate training of a large number of inter-related object classifiers (where an object network is constructed for determining the inter-related learning tasks directly in the feature space rather than in the label space). Our experimental results have shown that our MTML-MIL algorithm can achieve higher detection accuracy rates for automatic object detection.

Digital video now plays an important role in supporting more profitable online patient training and counseling, and integration of patient training videos from multiple competitive organizations in the health care network will result in better offerings for patients. However, privacy concerns often prevent multiple competitive organizations from sharing and integrating their patient training videos. In addition, patients with infectious or chronic diseases may not want the online patient training organizations to identify who they are or even which video clips they are interested in. Thus, there is an urgent need to develop more effective techniques to protect both video content privacy and access privacy . In this paper, we have developed a new approach to construct a distributed Hippocratic video database system for supporting more profitable online patient training and counseling. First, a new database modeling approach is developed to support concept-oriented video database organization and assign a degree of privacy of the video content for each database level automatically. Second, a new algorithm is developed to protect the video content privacy at the level of individual video clip by filtering out the privacy-sensitive human objects automatically. In order to integrate the patient training videos from multiple competitive organizations for constructing a centralized video database indexing structure, a privacy-preserving video sharing scheme is developed to support privacy-preserving distributed classifier training and prevent the statistical inferences from the videos that are shared for cross-validation of video classifiers. Our experiments on large-scale video databases have also provided very convincing results.

As digital cameras and online photo sharing services become more popular, digital images are growing exponentially on the Internet. Thus supporting more effective retrieval from large-scale image collections has become a challenging issue for CBIR community. In this paper, we have developed a human-centered computing framework for assisting users to access large-scale online image collections: (a) filtering out junk images from Google images; and (b) exploring large-scale Flickr images.

In this paper we propose an imbalanced classification algorithm to extract informative images from web news pages. Our algorithm resolve the difficult problem based on two approaches. First, we limit our dataset to a specific application area so that the patterns of the informative images can be captured by existing classification algorithms. Second, we propose an automatic negative samples filtering algorithm to eliminate most negative samples, so that the classification training data is rebalanced. Because most classification algorithms have reduced performance on imbalanced training data, our algorithm improves the overall performance significantly. In addition, our approach is inherently robust to new web sites and style/layout change of web sites.

Photos taken by human beings significantly differ from the pictures that are taken by a surveillance camera or a vision sensor on a robot, e.g., human beings may intentionally capture photos to express his/her feeling or record a memorial scene. Such a creative photo capture process is accomplished by adjusting two factors: (1) the parameters setting of a camera; and (2) the position between the camera and the interesting objects or scenes. To enable automatic understanding and interpretation of the semantics of photos, it is very important to take all these factors into account. Unfortunately, most existing algorithms for image understanding focus on only the content of the images while completely ignoring these two important factors. In this paper, we have developed a new algorithm to calculate what the interestingness of the photographer is and what the core content of a photo is. The gained information (i.e., attended regions and attention of the photographer) is further used to support more effective photo classification and retrieval. Our experiments on 70,000+ photos taken by 200+ different models of cameras have obtained very positive results.

In this paper, we have developed a novel framework calledJustClickto enable personalized image recommendation via exploratory search from large-scale collections of Flickr images. First, a topic network is automatically generated to summarize large-scale collections of Flickr images at a semantic level. Hyperbolic visualization is further used to enable interactive navigation and exploration of the topic network, so that users can gain insights of large-scale image collections at the first glance, build up their mental query models interactively and specify their queries (i.e., image needs) more precisely by selecting the image topics on the topic network directly. Thus, our personalized query recommendation framework can effectively address both the problem of query formulation and the problem of vocabulary discrepancy and null returns. Second, a small set of most representative images are recommended for the given image topic according to their representativeness scores. Kernel principal component analysis and hyperbolic visualization are seamlessly integrated to organize and layout the recommended images (i.e., most representative images) according to their nonlinear visual similarity contexts, so that users can assess the relevance between the recommended images and their real query intentions interactively. An interactive interface is implemented to allow users to express their time-varying query intentions precisely and to direct ourJustClicksystem to more relevant images according to their personal preferences. Our experiments on large-scale collections of Flickr images show very positive results.

The keyword-based Google images search engine is now becoming very popular for online image search. Unfortunately, only the text terms that are explicitly or implicitly linked with the images are used for image indexing but the associated text terms may not have exact correspondence with the underlying image semantics, thus the keyword-based Google images search engine may return large amounts of junk images which are irrelevant to the given keyword-based queries. Based on this observation, we have developed an interactive approach to filter out the junk images from keyword-based Google images search results and our approach consists of the following major components. a) A kernel-based image clustering technique is developed to partition the returned images into multiple clusters and outliers. b) Hyperbolic visualization is incorporated to display large amounts of returned images according to their nonlinear visual similarity contexts, so that users can assess the relevance between the returned images and their real query intentions interactively and select one or multiple images to express their query intentions and personal preferences precisely. c) An incremental kernel learning algorithm is developed to translate the users' query intentions and personal preferences for updating the mixture-of-kernels and generating better hypotheses to achieve more accurate clustering of the returned images and filter out the junk images more effectively. Experiments on diverse keyword-based queries from Google images search engine have obtained very positive results. Our junk image filtering system is released for public evaluation at: http://www.cs.uncc.edu/~jfan/google-demo/.

In this paper, we have developed an interactive system to enable personalized news video recommendation. First, multi-modal information channels (audio, video and closed captions) are seamlessly integrated and synchronized to achieve more reliable news topic detection, and the contextual relationships between the news topics are extracted automatically. Second, topic network and hyperbolic visualization are seamlessly integrated to achieve interactive navigation and exploration of large-scale collections of news videos at the topic level, so that users can have a good global overview of large-scale collections of news videos at the first glance. In such interactive topic network navigation and exploration process, the users' personal background knowledge can be taken into consideration for obtaining the news topics of interest interactively, building up their mental models of news needs precisely and formulating their searches easily by selecting the visible news topics on the screen directly. Our system can further recommend the relevant web news, the new search directions, and the most relevant news videos according to their importance and representativeness scores.

In this paper, we have developed a new scheme for achieving multilevel annotations of large-scale images automatically. To achieve more sufficient representation of various visual properties of the images, both the global visual features and the local visual features are extracted for image content representation. To tackle the problem of huge intraconcept visual diversity, multiple types of kernels are integrated to characterize the diverse visual similarity relationships between the images more precisely, and a multiple kernel learning algorithm is developed for SVM image classifier training. To address the problem of huge interconcept visual similarity, a novel multitask learning algorithm is developed to learn the correlated classifiers for the sibling image concepts under the same parent concept and enhance their discrimination and adaptation power significantly. To tackle the problem of huge intraconcept visual diversity for the image concepts at the higher levels of the concept ontology, a novel hierarchical boosting algorithm is developed to learn their ensemble classifiers hierarchically. In order to assist users on selecting more effective hypotheses for image classifier training, we have developed a novel hyperbolic framework for large-scale image visualization and interactive hypotheses assessment. Our experiments on large-scale image collections have also obtained very positive results.

In this paper, we have proposed a novel framework for mining multilevel image semantics via hierarchical classification. To bridge the semantic gap more successfully, salient objects are used to characterize the intermediate image semantics effectively. The salient objects are defined as the connected image regions that capture the dominant visual properties linked to the corresponding physical objects in an image. To achieve a more reliable and tractable concept learning in high-dimensional feature space, a novel algorithm calledproduct of mixture-experts(PoM) is proposed to reduce the size of training images and speed up concept learning. A novel hierarchical concept learning algorithm is proposed by incorporating concept ontology and multitask learning to enhance the discrimination power of the concept models and reduce the computational complexity for learning the concept models for large amount of image concepts, which may have huge intra-concept variations and inter-concept similarities on their visual properties. A hyperbolic image visualization algorithm has been developed for allowing users to specify their queries easily and assess the query results interactively. Our experiments on large-scale image collections have also obtained very positive results.

For online medical education purposes, we have developed a novel scheme to incorporate the results of semantic video classification to select the most representative video shots for generating concept-oriented summarization and skimming ofsurgery education videos. First, salient objects are used as the video patterns for feature extraction to achieve a good representation of the intermediate video semantics. The salient objects are defined as the salient video compounds that can be used to characterize the most significant perceptual properties of the corresponding real world physical objects in a video, and thus the appearances of such salient objects can be used to predict the appearances of the relevant semantic video concepts in a specific video domain. Second, a novelmulti-modal boostingalgorithm is developed to achieve more reliable video classifier training by incorporating feature hierarchy and boosting to dramatically reduce both the training cost and the size of training samples, thus it can significantly speed up SVM (support vector machine) classifier training. In addition, the unlabeled samples are integrated to reduce the human efforts on labeling large amount of training samples. Finally, the results of semantic video classification are incorporated to enable concept-oriented video summarization and skimming. Experimental results in a specific domain ofsurgery education videosare provided.

Recently more and more researches focus on the concept extraction from unstructured video data. To bridge the semantic gap between the low-level features and the high-level video concepts, a mid-level understanding of the video contents, i.e., salient object is detected based on the techniques of image segmentation and machine learning. Specifically, 21 salient object detectors are developed and tested on TRECVID 2005 development video corpus. In addition, a boosting method is proposed to select the most representative features to achieve a higher performance than only using single modality, and lower complexity than taking all features into account.

For large-scale classification problems, the training samples can be clustered beforehand as a downsampling pre-process, and then only the obtained clusters are used for training. Motivated by such assumption, we proposed a classification algorithm, Support Cluster Machine (SCM), within the learning framework introduced by Vapnik. For the SCM, a compatible kernel is adopted such that a similarity measure can be handled not only between clusters in the training phase but also between a cluster and a vector in the testing phase. We also proved that the SCM is a general extension of the SVM with the RBF kernel. The experimental results confirm that the SCM is very effective for largescale classification problems due to significantly reduced computational costs for both training and testing and comparable classification accuracies. As a by-product, it provides a promising approach to dealing with privacy-preserving data mining problems.

In this paper, a hierarchical classification framework has been proposed for bridging the semantic gap effectively and achieving multi-level image annotation automatically. First, the semantic gap between the low-level computable visual features and users' real information needs is partitioned into four smaller gaps, and multiple approachesallare proposed to bridge these smaller gaps more effectively. To learn more reliable contextual relationships between the atomic image concepts and the co-appearances of salient objects, a multi-modal boosting algorithm is proposed. To enable hierarchical image classification and avoid inter-level error transmission, a hierarchical boosting algorithm is proposed by incorporating concept ontology and multi-task learning to achieve hierarchical image classifier training with automatic error recovery. To bridge the gap between the computable image concepts and the users' real information needs, a novel hyperbolic visualization framework is seamlessly incorporated to enable intuitive query specification and evaluation by acquainting the users with a good global view of large-scale image collections. Our experiments on large-scale image databases have also obtained very positive results.

Most existing content-based video retrieval (CBVR) systems are now amenable to support automatic low-level feature extraction, but they still have limited effectiveness from a user's perspective because of the semantic gap. Automatic video concept detection via semantic classification is one promising solution to bridge the semantic gap. To speed up SVM video classifier training in high-dimensional heterogeneous feature space, a novel multimodal boosting algorithm is proposed by incorporating feature hierarchy and boosting to reduce both the training cost and the size of training samples significantly. To avoid the inter-level error transmission problem, a novel hierarchical boosting scheme is proposed by incorporating concept ontology and multitask learning to boost hierarchical video classifier training through exploiting the strong correlations between the video concepts. To bridge the semantic gap between the available video concepts and the users' real needs, a novel hyperbolic visualization framework is seamlessly incorporated to enable intuitive query specification and evaluation by acquainting the users with a good global view of large-scale video collections. Our experiments in one specific domain of surgery education videos have also provided very convincing results.

Robust detection of a large dictionary of salient objects in natural image database is of fundamental importance to image retrieval systems. We review three popular frameworks for salient object detection, i.e., segmentation-based method, grid-based method and part-based method and discuss their advantages and limitations. We argue that using these frameworks individually is generally not enough to handle a large number of salient object classes accurately because of the intrinsic diversity of salient object features. Motivated by this observation, we have proposed a new system which combines the merits of these frameworks into one single hybrid system. The system automatically selects the appropriate modeling method for each individual object class using J measure and shape variance. We conduct comparison experiments on two popular image dataset – Corel and LabelMe. Empirical results have shown that the proposed hybrid method is more general and can handle much more salient object classes in a robust manner.

The performance of image classifiers largely depends on two inter-related issues:(1)suitable frameworks for image content representation and automatic feature extraction;(2) effective algorithms for image classifier training and feature subset selection. To address the first issue, a multiresolution grid-based framework is proposed for image content representation and feature extraction to bypass the time-consuming and erroneous process for image segmentation. To address the second issue, a hierarchical boosting algorithm is proposed by incorporating feature hierarchy and boosting to scale up SVM image classifier training in high-dimensional feature space. The high-dimensional multi-modal heterogeneous visual features are partitioned into multiple low-dimensional single-modal homogeneous feature subsets and each of them characterizes certain visual property of images. For each homogeneous feature subset, principal component analysis (PCA)is performed to exploit the feature correlations and a weak classifier is learned simultaneously. After the weak classifiers for different feature subsets and grid sizes are available, they are combined to boost an optimal classifier for the given object class or image concept, and the most representative feature subsets and grid sizes are selected. Our experiments on a specific domain of natural images have obtained very positive results.

Automatic image annotation is a promising solution to enable keyword-based semantic image retrieval. In this demo, we present our image search engine I-Search implemented using a multi-level semantic image annotation algorithm. By incorporating a domain-specific ontology into the autogenerated annotations, we are able to organize large scale natural image databases into hierarchical structures for browsing and keyword-based searching without referring to external text information.

Most existing systems for content-based video retrieval (CBVR) are now amenable to support automatic low-level video content analysis and feature extraction, but they have limited effectiveness from a user's perspective. To support semantic video retrieval via keywords, we have proposed a novel framework by incorporating the concept ontology to enable more effective modeling and representation of semantic video concepts. Specifically, this novel framework includes: (a) Using the salient objects to achieve a middle-level understanding of the semantics of video contents; (b) Building a domain dependent concept ontology to enable multi-level modeling and representation of semantic video concepts; (c) Developing a multi-task boosting technique to achieve hierarchical video classifier training for automatic multi-level video annotation. The experimental results in a certain domain of medical education videos are also provided.

Motivated by Google's great success on text document retrieval and recent progresses of semantic video understanding, researchers begin to build new generation of video retrieval systems that are able to support semantic sensitive video retrieval via keywords. Unfortunately, these systems are not able to provide satisfactory results for the masses because of several inter-related challenging problems. We have proposed novel algorithms to resolve some of these problems. Firstly, the salient object based semantic classification algorithm is proposed to extract semantic concepts of video clips. Secondly, the video visualization based interactive retrieval framework is proposed to help users input semantic and visual queries efficiently and effectively. Finally, the concept-oriented skimming algorithm is proposed to help users efficiently check search results.

As the content of everyday news reports is unpredictable, keyword based news search engine can't provide effective services to audiences because the audiences may not be able to figure out proper keywords to search. In this paper, a novel framework is proposed to help audiences browse and retrieve news video clips without the need of keywords. Interesting keyframes and keywords are automatically extracted from news video clips and visually represented according to their interestingness and informativeness measurement. A computational approach is also developed to quantify the interestingness measurement of video clips. The keyframes and keywords are carefully organized so that the audiences can find news stories of interest at first glance.

In this paper we have addressed confidentiality and privacy for video surveillance databases. First we discussed our overall approach for suspicious event detection. Next we discussed an access control model and accedes control algorithms for confidentiality. Finally we discuss privacy preserving video surveillance. Our goal is build a comprehensive system that can detect suspicious events, ensure confidentiality as well as privacy.

International audience

In this paper, we have proposed a novel framework to achieve more effective classifier training by using unlabeled samples. By integrating concept hierarchy for semantic image concept organization, a hierarchical mixture model is proposed to enable multi-level image concept modeling and hierarchical classifier training. To effectively learn the base-level classifiers for the atomic image concepts at the first level of the concept hierarchy, we have proposed a novel adaptive EM algorithm to achieve more effective classifier training with higher prediction accuracy. To effectively learn the classifiers for the higher-level semantic image concepts, we have also proposed a novel technique for classifier combining by using hierarchical mixture model. The experimental results on two large-scale image databases are also provided.

International audience

Automatic image annotation is a promising solution to enable semantic image retrieval via keywords. In this paper, we propose a multi-level approach to annotate the semantics of natural scenes by using both the dominant image components (salient objects) and the relevant semantic concepts. To achieve automatic image annotation at the content level, we use salient objects as the dominant image components for image content representation and feature extraction. To support automatic image annotation at the concept level, a novel image classification technique is developed to map the images into the most relevant semantic image concepts. In addition, Support Vector Machine (SVM) classifiers are used to learn the detection functions for the pre-defined salient objects and finite mixture models are used for semantic concept interpretation and modeling. An adaptive EM algorithm has been proposed to determine the optimal model structure and model parameters simultaneously. We have also demonstrated that our algorithms are very effective to enable multi-level annotation of natural scenes in a large-scale image dataset.

Effective video skimming requires a good understanding of the semantics of video contents. However, more existing systems for content-based video retrieval (CBVR) can only support low-level video analysis, but they have limited effectiveness on achieving semantic-sensitive video understanding. In this paper, we have developed a novel framework to achieve concept-oriented video skimming and it consists of three parts: (a) using salient objects for semantic-sensitive video content representation; (b) using finite mixture models for semantic video concept modeling and classification; (c) enabling concept-oriented video skimming via semantic video classification.

Multi-level annotation of images is a promising solution to enable more effective semantic image retrieval by using various keywords at different semantic levels. In this paper, we propose a multi-level approach to annotate the semantics of natural scenes by using both the dominant image components and the relevant semantic concepts. In contrast to the well-known image-based and region-based approaches, we use the salient objects as the dominant image components to achieve automatic image annotation at the content level. By using the salient objects for image content representation, a novel image classification technique is developed to achieve automatic image annotation at the concept level. To detect the salient objects automatically, a set of detection functions are learned from the labeled image regions by using Support Vector Machine (SVM) classifiers with an automatic scheme for searching the optimal model parameters. To generate the semantic concepts, finite mixture models are used to approximate the class distributions of the relevant salient objects. An adaptive EM algorithm has been proposed to determine the optimal model structure and model parameters simultaneously. We have also demonstrated that our algorithms are very effective to enable multi-level annotation of natural scenes in a large-scale dataset.

Semantic video classification has become an active research topic to enable more effective video retrieval and knowledge discovery from large-scale video databases. However, most existing techniques for classifier training require a large number of hand-labeled samples to learn correctly. To address this problem, we have proposed a semi-supervised framework to achieve incremental classifier training by integrating a limited number of labeled samples with a large number of unlabeled samples. Specifically, this emi-supervised framework includes: (a) Modeling the semantic video concepts by using the finite mixture models to approximate the class distributions of the relevant salient objects; (b) Developing an adaptive EM algorithm to integrate the unlabeled samples to achieve parameter estimation and model selection simultaneously; The experimental results in a certain domain of medical videos are also provided.

Digital video now plays an important role in medical education, health care, telemedicine and other medical applications. Several content-based video retrieval (CBVR) systems have been proposed in the past, but they still suffer from the following challenging problems: semantic gap, semantic video concept modeling, semantic video classification, and concept-oriented video database indexing and access. In this paper, we propose a novel framework to make some advances toward the final goal to solve these problems. Specifically, the framework includes: 1) a semantic-sensitive video content representation framework by using principal video shots to enhance the quality of features; 2) semantic video concept interpretation by using flexible mixture model to bridge the semantic gap; 3) a novel semantic video-classifier training framework by integrating feature selection, parameter estimation, and model selection seamlessly in a single algorithm; and 4) a concept-oriented video database organization technique through a certain domain-dependent concept hierarchy to enable semantic-sensitive video retrieval and browsing.

Recent advances in digital video compression and networks have made video more accessible than ever. However, the existing content-based video retrieval systems still suffer from the following problems. 1) Semantics-sensitive video classification problem because of the semantic gap between low-level visual features and high-level semantic visual concepts; 2) Integrated video access problem because of the lack of efficient video database indexing, automatic video annotation, and concept-oriented summary organization techniques. In this paper, we have proposed a novel framework, called ClassView, to make some advances toward more efficient video database indexing and access. 1) A hierarchical semantics-sensitive video classifier is proposed to shorten the semantic gap. The hierarchical tree structure of the semantics-sensitive video classifier is derived from the domain-dependent concept hierarchy of video contents in a database. Relevance analysis is used for selecting the discriminating visual features with suitable importances. The Expectation-Maximization (EM) algorithm is also used to determine the classification rule for each visual concept node in the classifier. 2) A hierarchical video database indexing and summary presentation technique is proposed to support more effective video access over a large-scale database. The hierarchical tree structure of our video database indexing scheme is determined by the domain-dependent concept hierarchy which is also used for video classification. The presentation of visual summary is also integrated with the inherent hierarchical video database indexing tree structure. Integrating video access with efficient database indexing tree structure has provided great opportunity for supporting more powerful video search engines.

To achieve more efficient video indexing and access, we introduce a video database management framework and strategies for video content structure and events mining. The video shot segmentation and representative frame selection strategy are first utilized to parse the continuous video stream into physical units. Video shot grouping, group merging, and scene clustering schemes are then proposed to organize the video shots into a hierarchical structure using clustered scenes, scenes, groups, and shots, in increasing granularity from top to bottom. Then, audio and video processing techniques are integrated to mine event information, such as dialog, presentation and clinical operation, from the detected scenes. Finally, the acquired video content structure and events are integrated to construct a scalable video skimming tool which can be used to visualize the video content hierarchy and event information for efficient access. Experimental results are also presented to evaluate the performance of the proposed framework and algorithms.

As digital cameras become more affordable, digital video now plays an important role in medical education and healthcare. In this paper, we propose a novel framework to facilitate semantic classification of surgery education videos. Specifically, the framework includes: (a) semantic-sensitive video content characterization via principal video shots, (b) semantic video classification via a mixture Gaussian model to bridge the semantic gap between low-level visual features and semantic visual concepts in a specific surgery education video domain.

Content-based video database access control is becoming very important, but it depends on the progresses of the following related research issues: (a) efficient video analysis for supporting semantic visual concept representation; (b) effective video database indexing structure; (c) the development of suitable video database models; and (d) the development of access control models tailored to the characteristics of video data. In this paper, we propose a novel approach to support multilevel access control in video databases. Our access control technique combines a video database indexing mechanism with a hierarchical organization of visual concepts (i.e., video database indexing units), so that different classes of users can access different video elements or even the same video element with different quality levels according to their permissions. These video elements, which, in our access control mechanism, are used for specifying the authorization objects, can be a semantic cluster, a subcluster, a video scene, a video shot, a video frame, or even a salient object (i.e., region of interest). In the paper, we first introduce our techniques for obtaining these multilevel video access units. We also propose a hierarchical video database indexing technique to support our multilevel video access control mechanism. Then, we present an innovative access control model which is able to support flexible multilevel access control to video elements. Moreover, the application of our multilevel video database modeling, representation, and indexing for MPEG-7 is discussed.

In our project, we are adopting a new approach for handling video data. We view the video as a well-defined data type with its own description, parameters and applicable methods. The system is based on PREDATOR, an open-source object-relational DBMS. PREDATOR uses Shore as the underlying storage manager. Supporting video operations (storing, searching-by-content and streaming) and new query types (query-by-example and multi-feature similarity searching) requires major changes in many of the traditional system components. More specifically, the storage and buffer manager has to deal with huge volumes of data with real-time constraints. Query processing has to consider the video methods and operators in generating, optimizing and executing the query plans.

Face detection is important in video/image content analysis and organization since the most important object in those media is often a human being. We propose a facial feature based omni-face detection algorithm. While utilizing the skin color model for face cue detection, a pairwise skin region refinement strategy is applied to eliminate the errors incurred by the skin model. Then, a region based adaptive threshold selection scheme is employed for facial feature segmentation. After the facial feature filtering, an orientation, pose and scale invariant face verification strategy is utilized to verify the detected face candidate regions. Experimental results demonstrate successful detection over a wide variety of facial variation in background, scale, view and orientation from different types of video collections.

This paper investigates the problem of augmenting labeled data with unlabeled data to improve classification accuracy. This is significant for many applications such as image classification where obtaining classification labels is expensive, while large unlabeled examples are easily available. We investigate an Expectation Maximization (EM) algorithm for learning from labeled and unlabeled data. The reason why unlabeled data boosts learning accuracy is because it provides the information about the joint probability distribution. A theoretical argument shows that the more unlabeled examples are combined in learning, the more accurate the result. We then introduce B-EM algorithm, based on the combination of EM with bootstrap method, to exploit the large unlabeled data while avoiding prohibitive I/O cost. Experimental results over both synthetic and real data sets that the proposed approach has a satisfactory performance.

1. SYSTEM TECHNICAL DESCRIPTION The ClassMiner system demonstrates a fully implemented tool for scalable video skimming and summarization. The key technology in the system is the integrated medical video content structure and events mining process, which was presented in a paper at the SIGMOD workshop on Data Mining and Knowledge Discovery [1]. As the system architecture in Fig. 1 indicates, we first apply a general video shot segmentation and key-frame selection scheme to parse the video stream into physical units. Then, the video group detection, scene detection and clustering strategies are executed to mine the video content structure. Various visual and audio feature processing techniques are utilized to detect some semantic cues, such as slides, face and speaker changes, etc. within the video, and these detection results are joined together to mine three types of events (presentation, dialog, clinical operation) from the detected video scenes. Finally, a scalable video skimming and summarization tool is constructed based on the mined video content structure and event information to help the user visualize and access video content.

Automatic image segmentation and object extraction play an important role in supporting content-based image coding, indexing, and retrieval. However, the low-level visual homogeneity critical (like color, texture, intensity, and so on) for segmentation do not lead to semantic objects directly because a semantic object can contain totally different gray levels, color, or texture. We propose an automatic model-based semantic object extraction algorithm by integrating object seeds with their region constraint graphs (perceptual models). Images are first partitioned into a set of homogeneous regions with accurate boundaries by integrating the results obtained by similarity-based region growing and edge detection procedures. We propose a 1-D fast entropic thresholding technique for determining the thresholds used in region growing and edge detection automatically. The object seeds, which are the intuitive and representative parts of semantic objects, are then distinguished from these homogeneous image regions. A seeded region aggregation procedure is used for merging the adjacent regions of a detected object seed to give a semantic object according to the perceptual model of the object. We focus on semantic human object generation by taking faces as object seeds and using a ratio-based perceptual model.

We propose a new automatic image segmentation method. Color edges in an image are first obtained automatically by combining an improved isotropic edge detector and a fast entropic thresholding technique. After the obtained color edges have provided the major geometric structures in an image, the centroids between these adjacent edge regions are taken as the initial seeds for seeded region growing (SRG). These seeds are then replaced by the centroids of the generated homogeneous image regions by incorporating the required additional pixels step by step. Moreover, the results of color-edge extraction and SRG are integrated to provide homogeneous image regions with accurate and closed boundaries. We also discuss the application of our image segmentation method to automatic face detection. Furthermore, semantic human objects are generated by a seeded region aggregation procedure which takes the detected faces as object seeds.