Feature point detection on textureless surfaces remains a fundamental challenge in computer vision due to the absence of discernible color and brightness gradients. From the imaging mechanism perspective, micro-geometry structures of textureless surfaces provide physically stable cues for feature point extraction despite the absence of visual distinctiveness. Therefore, we propose a novel feature point detection method, which reconstructs surface micro-geometry structures from a single RGB image and leverages these micro-geometry structures for feature extraction, without relying on specialized equipment or complex deep learning models. Specifically, our method establishes a novel framework that models the light-surface interaction to analyze phase modulation in reflected light. Then it reconstructs underlying micro-geometry structures through Gabor Kernel-based spectral analysis, enabling accurate quantification of surface height variations from phase information. This information forms the foundation of our proposed Concave-Convex Index (CCI), a robust geometric descriptor that achieves stable feature characterization through geometry-aware measurements. Extensive evaluations on TUM, T-LESS, Shape2.5D datasets and self-collected images, demonstrate our method's superior capability in extracting stably distributed and highly repeatable feature points, even when visible texture or brightness gradients vanish. Our method offers a novel perspective for reliable feature point detection on challenging textureless surfaces across diverse materials and illumination conditions.

Artificial intelligence (AI) agents are systems capable of perceiving their environment, autonomously planning and executing tasks. Recent advancements in Large Language Models (LLMs) have introduced a transformative paradigm for AI agents, enabling them to interact with external resources and tools through prompt techniques. This advancement has significantly extended the capabilities of LLMs, positioning LLM-based AI Agents as an important research area. In such agents, the workflow integrates developer-written code, which manages framework construction and logic control, with LLM-generated natural language that enhances dynamic decision-making and interaction. However, inconsistencies between LLM outputs and developer logic can lead to defects, such as tool invocation failures. These issues introduce specific risks, leading to various defects in LLM-based AI Agents, including service interruptions and incorrect output. Despite the importance of these issues, there is a lack of systematic work that focuses on analyzing LLM-based AI Agents to uncover defects in their code. To address this gap, we present the first study focused on identifying and detecting defects in LLM Agents. We collected and analyzed 14,754 relevant developer reports from StackOverflow and GitHub. We further filtered 2,604 valid posts to define and classify eight types of agent code defects. Then, we designed a static analysis tool, named Agentable, to detect these defects. Agentable leverages Code Property Graphs (CPGs) and LLMs to analyze Agent workflows by efficiently identifying specific code patterns and analyzing natural language descriptions. To evaluate Agentable, we constructed two datasets: AgentSet, which consists of 84 real world Agent projects, and AgentTest, which contains 78 Agent projects specifically designed to include various types of defects. Our evaluation shows that Agentable achieves a precision of 88.79% on the real-world agent dataset and a recall of 91.03% on the manually labeled defect dataset. Furthermore, our analysis identifies 889 defects in real-world agent projects, highlighting the prevalence of these issues in practice.

Surface normal estimation is a fundamental task in point cloud processing and plays a crucial role in downstream applications. Existing methods typically extract features from local neighborhoods or patches, followed by surface fitting or direct regression to predict normals. However, the scale ambiguity in determining the optimal neighborhood hinders effective extraction of geometric information, making normal estimation for unstructured point clouds with significant density variations particularly challenging. To address this challenge, we propose APN-Net, an adaptive perception network for point cloud normal estimation. Specifically, we design the Graphical Information Self-perception (GIS) module, which provides an implicit manner for region partitioning and expands the receptive field, enabling automatic extraction of both local geometric details and global structural information, while alleviating the scale ambiguity in determining the optimal neighborhood. Moreover, to capture complex geometric details, we introduce the Adaptive Graph Convolution (AGC) module, which employs adaptive kernels to model relationships among points across different semantic regions, thereby enabling richer feature representation. Extensive experiments on both synthetic and real-world scanned datasets demonstrate that APN-Net achieves superior performance in unoriented normal estimation, particularly for point clouds with significant density variations.

Decentralized Applications (DApps) serve as the gateway to utilizing blockchain technology. As their prevalence continues to grow, DApps are becoming increasingly interconnected. For instance, a DApp does not need to manage the prices of various tokens internally, as it can retrieve this information from other DApps that provide more up-to-date data. However, such deep reliance also introduces more attack surfaces, posing greater risks to both DApps and their users. In this paper, we refer to the security threat arising from the interdependence of DApps as Insecure State Dependency (ISD). Public reports indicate that ISD has led to losses exceeding 340 million USD.Existing ISDs are mostly found by extensive manual auditing and lucky incidents, as automated discovery of such issues is extremely difficult. More specifically, it is by no means trivial to (1) achieve precise data tracking in the intertwined and invisible interactions of DApps, (2) obtain fine-grained semantic information in low semantic bytecode. In this paper, we propose a novel framework, called InsFinder, for detecting ISD in DApps. Specifically, InsFinder consists of three unique modules to overcome the aforementioned challenges. (1) InsFinder employs dynamic cross-DApp taint analysis to achieve accurate multi-source data tracking in heavily coupled DApp interactions. (2) InsFinder uses source mapping to map bytecode identifiers into meaningful source code, such as variable names or statements, enabling a deeper understanding of bytecode. (3) InsFinder implements fine-grained access control and static analysis for ISD entry point detection. Evaluation on a manually annotated dataset with 93 real-world ISDs shows that InsFinder successfully detects 72 of them, achieving a precision of 84.7% and a recall of 77.4%. Furthermore, InsFinder successfully uncovers 165 previously unreported ISDs across 122 DApp projects. These ISDs collectively impact over 2 million USD.

Federated Learning (FL) faces challenges due to straggler clients that impede timely parameter uploads, potentially leading to suboptimal global model performance. Existing approaches using synchronous and asynchronous communication suffer from long waiting times or convergence issues. We propose Fed-OGD, a novel asynchronous FL method addressing the straggler problem through gradient orthogonalization. Our approach innovatively frames the straggler issue using catastrophic forgetting theory, viewing stragglers as instances of the global model “forgetting” to aggregate their parameters. Fed-OGD introduces an Orthogonal Gradient Descent (OGD) technique that caches straggler gradients and orthogonalizes the difference between these and current active client gradients. By projecting active gradients onto straggler orthogonal bases and subtracting the resulting components, we obtain orthogonalized gradients guiding the model towards optimality. We provide theoretical convergence guarantees and demonstrate Fed-OGD’s effectiveness through extensive experiments. Our method achieves state-of-the-art performance across multiple datasets among SOTA FL baselines, with notable improvements in non-IID (non-Independent and identically distributed) scenarios: there are few main categories with many samples while other categories hold few samples in a client. Fed-OGD achieves that 16.66% increase in accuracy on CIFAR-10, and significant gains on CIFAR-100 (5.37%), Tiny-ImageNet (38.51%), and AG_NEWS (16.30%).

Smart contracts are programs running on blockchains, managing substantial volumes of wealth stored within the blockchain platforms. To safeguard these assets, developers design and implement access control policies. However, incomplete and incorrect access control policies allow malicious attackers to gain unauthorized access and exploit additional assets. Previous tools for detecting access control vulnerabilities in smart contracts rely on predefined patterns, specifications, or mining access control policies from historical transactions. However, these methods are constrained due to their predetermined nature and the diversity and complexity of smart contracts.In this paper, we presentASTRO, a new framework employing code similarity to detect access control vulnerabilities in smart contracts. In contrast to prior approaches that heavily rely on predefined, vulnerable code samples,ASTROdetects whether a target contract has access control vulnerabilities by comparing it against a database of audited contracts. Moreover, to mitigate the impact of language-specific features (e.g., diverse conditional statements and modifiers) and writing style characteristics, we integrate pruning and normalization techniques. We evaluateASTROon a total of 22 smart contracts with assigned access control CVEs and those attacked because of access control vulnerabilities from the past two years. Evaluation results demonstrate that, compared to state-of-the-art tools (i.e., AChecker, SpCon),ASTROsurpasses all tools in recall and achieves an improvement in recall by at least 2.8 times. In addition,ASTROachieves a precision of 78.33% on a dataset consisting of real-wild contracts. Furthermore,ASTROsuccessfully identified 19 exploitable vulnerable contract that can be used to directly gain access to the contract’s permissions and obtain benefits.

Continual learning of new concepts is essential in artificial intelligence systems, especially in the field of image classification, which necessitates incremental knowledge acquisition without catastrophic forgetting. An inherent problem in current methodologies is the stability-plasticity dilemma or excessive storage overhead. In this paper, we propose a novel method called Adversarial Reprogramming Visual Prompts (ARVP) for Class-Incremental Learning (CIL). ARVP employs an adaptive adversarial reprogramming strategy for visual prompts, strategically downsampling background pixels while preserving discriminative pixels, thus maintaining sample discriminability without compromising quality. Furthermore, a Bop-based formulation and an end-to-end training approach are implemented to enhance the learning efficiency and effectiveness of the model. The optimization of ARVP is divided into two stages, focusing on adding visual prompts and downsampling images. Experimental results on CIFAR-100 and Food-101 benchmarks show that integrating ARVP with existing CIL approaches leads to performance enhancements and using the prompted exemplars can achieve a new state-of-the-art CIL accuracy. Further analysis reveals that ARVP remains effective even with reduced old sample usage.

Rainfall, as a ubiquitous atmospheric phenomenon, causes visual image degradation through rain streaks, which in turn hinders human vision and degrades the performance of subsequent computer vision algorithms. In order to further improve the performance of visual algorithms in harsh rain environments, image rain removal technology has become a hot topic in this field. In computer vision, the integration of the Vision Transformer (ViT) marks an important progress, and its transfer application to the image rain removal task has achieved remarkable success. However, the application of transformers in this task still faces challenges. While they excel at modelling long-range dependencies, transformers are less adept at modelling local contexts for image restoration tasks. Moreover, the internal transmission of features within dense transformer layers often needs more finesse. In this paper, we tackle these shortcomings by incorporating convolutional mechanisms within both the self-attention components of the transformer blocks and the subsequent feed-forward networks. We refine the feature flows amalgamated between recursive transformer blocks by utilizing feature fusion strategies. To further refine the high-frequency textural details of images, we employ frequency domain contrastive learning techniques to augment the delineation of contrastive sample information, ensuring that the restored image closely resembles the clear image in terms of texture structure while maintaining a distinction from the rain image. Extensive quantitative and qualitative experimental results show that the proposed deraining network surpasses mainstream methods on public datasets.

Accelerometer signals play a critical role in many fields, for example navigation and vehicle safety. However, uncontrollable factors such as defective equipment and harsh environments make the signals recorded by sensors contain a large amount of noise, which poses a great challenge. Most existing methods are based on traditional signal processing and often face problems of incomplete noise reduction or signal distortion after denoising. In this paper, we propose a data-driven denoising method for accelerometer signals (ADNet) based on the Wave_U_Net network architecture, incorporating Multi-Head Attention mechanism and Spatial Attention mechanism. This enhances the network’s feature extraction capability and accelerates its convergence speed. Meanwhile, the attention mechanism focuses the model’s attention on the clean signal’s feature information, improving the network’s fitting capability to the signal distribution. Therefore, ADNet can achieve more thorough denoising while reducing signal distortion after denoising. Finally, we conduct experiments on Walking speed dataset and field experiment dataset to verify the effectiveness of ADNet. The experimental results show that ADNet outperforms other baseline models in terms of Mean Square Error, Mean Absolute Error, Root Mean Square Error, and Signal-to-Noise Ratio.

An image enhancement method, which is solving the issue of detail loss caused by the inability of existing image enhancement methods to balance brightness and saturation in Wireless Capsule Endoscope (WCE) low-light environment, is proposed. Firstly, we design a multi-scale fast guided filter to estimate the illumination component and utilize the OTSU method to determine the function parameters based on the grayscale information of the illumination component. Secondly, we construct a brightness enhancement function based on the Weber-Fechner law to achieve brightness enhancement of the V component image. At the same time, we designed the brightness enhancement coefficient and combined with Haar wavelet to operate the S component image to balance the brightness and saturation of the WCE enhanced image. Finally, the image enhancement result is obtained by merging the channels and converting to the RGB color space. Comprehensive experimental results show that compared with existing methods, our proposed method improves the mean, standard deviation and information entropy evaluation criteria by 18.2, 5.81 and 0.26 respectively. Furthermore, the feature point detection and matching numbers of the enhanced images increased by an average of 59.3% and 32.9% respectively. Moreover, the effectiveness of this method is further verified through the improvement of experimental results of single-image depth estimation accuracy.

Limited data usually cause deep neural networks to hold poor performance after training, and many generative models are proposed to synthesize data to improve the performance of models. However, existing models ignore capturing the small defect details (e.g., features and locations), resulting in that most models cannot augment the Defect Location Sensitive Data (DLS data) in which the ratio of object size to the image size is small (e.g., 20%) and the locations of the defects are only on the object. In this paper, we propose a new augmentation model, named Defect Location Sensitive data augmentation GAN (DLS-GAN), to address DLS data augmentation problem. First, we modify the vanilla generator with two Encoder-Decoder models, and view the limited masked-images masked by labeling the defect-free pixels while remaining the defect pixels in defect images and many defect-free images as the input of the two models. The extracted feature map from the first Encoder-Decoder model provides the defect features and location information; the second one extracts the features of defect-free images, and integrates the two different features with a designed Defect Feature Transfer Module to synthesize images with desired defects. Second, we employ two discriminators to estimate the scores of both distribution matching degree and defect similarity between real data and generated ones. With the two modifications, we design a new loss function, and then prove that it makes our model get converged. Last, we conduct extensive experiments to demonstrate the significant performance improvement and generalizability of DLS-GAN on different types of DLS datasets. The experimental results show that our DLS-GAN outperforms the SOTA generative models in terms of synthesizing high quality images with desired defects.Note to Practitioners—Automated defect image detectors play an important role in the field of automated manufacturing. Training a detector with superior detection performance usually requires a large number of samples. However, it is difficult to collect many defect samples in practice. Although existing generative methods can synthesize realistic-like images, they cannot generate the Defect Location Sensitive Data (DLS Data) which refer to the samples that the defects appear at the specific locations in product objects, resulting in the synthesized images invalid. This paper proposes a new defect image generation model called DLS-GAN to address this problem, and validates its performance in different real-world industrial datasets ranging from DLS Data to Non-DLS Data. Such generated images can be adopted as useful resources for improving the detection performance of automated detector.

Somatic tumors have a high-dimensional, sparse, and small sample size nature, making cancer subtype stratification based on somatic genomic data a challenge. Current methods for improving cancer clustering performance focus on dimension reduction, integrating multi-omics data, or generating realistic samples, yet ignore the associations between mutated genes within the patient-gene matrix. We refer to these associations as gene mutation structural information, which implicitly includes cancer subtype information and can enhance subtype clustering. We introduce a novel method for cancer subtype clustering called SIG(Structural Information within Graph). As cancer is driven by a combination of genes, we establish associations between mutated genes within the same patient sample, pair by pair, and use a graph to represent them. An association between two mutated genes corresponds to an edge in the graph. We then merge these associations among all mutated genes to obtain a structural information graph, which enriches the gene network and improves its relevance to cancer clustering. We integrate the somatic tumor genome with the enriched gene network and propagate it to cluster patients with mutations in similar network regions. Our method achieves superior clustering performance compared to SOTA methods, as demonstrated by clustering experiments on ovarian and LUAD datasets.

Spatio-temporal action detection networks, which need to simultaneously extract and fuse spatial and temporal features, often result in existing models becoming bloated and difficult to run in real-time and deploy on edge devices. This paper introduces an efficient and real-time spatio-temporal action detection model, YOWOv3. This model uses efficient 3D and 2D backbone networks to separately extract spatial and spatial-temporal features from sequential information. A lightweight spatio-temporal feature fusion module, designed by deeply integrating convolution and self-attention mechanisms, further enhances the extraction of spatio-temporal features. We refer to this module as the CFACM (Channel Fusion & Attention Convolution Mix) module. Our approach not only outperforms the latest efficient spatio-temporal action detection models in terms of lightness, reducing the model size by 24% compared to the latter, but also improves the mAP accuracy on the UCF101-24 dataset by 1.35%, while maintaining excellent speed performance, thus achieving a balance between accuracy and speed. Furthermore, existing models often use 3D convolutions to extract temporal information, which may be limited on certain devices, such as Apple’s M series processors. To mitigate the potential issue of 3D convolution operators not being supported during edge deployment of spatio-temporal action detection models, we employ a spatio-temporal shift module containing only 2D convolutions. This enables the model to acquire temporal information and inject the obtained temporal features into multi-level spatio-temporal feature extraction models. This not only liberates the model from the constraints of 3D convolution operations but also enhances the model’s balance between accuracy and speed. This results in state-of-the-art performance in lightweight networks using only 2D convolutions.

Mode collapse has been a persisting challenge in generative adversarial networks (GANs), and it directly affects the applications of GAN in many domains. Existing works that attempt to solve this problem have some serious limitations: models using optimal transport (OT) strategies (e.g., Wasserstein distance) lead to vanishing or exploding gradients; increasing the number of generators can cause several generators focusing on the same mode; and approaches that modify the loss also do not satisfactorily resolve mode collapse. In this article, we reduce mode collapse by formulating it as a Monge problem of OT map. We show that the Monge problem can be transformed to the distribution transformation problem in GAN, and a rectified affine neural network can be considered as a measurable function. In this way, we propose Monge GAN that uses this measurable function to transform the generated data distribution into the original data distribution. We utilize the Kantorovich formulation to obtain the OT cost, which is regarded as the OT distance between the two distributions. Finally, we conduct extensive experiments on both image and numerical datasets to validate our Monge GAN in reducing model collapse.

Due to the contradiction between limited bandwidth and huge transmission parameters, federated Learning (FL) has been an ongoing challenge to reduce the model parameters that need to be transmitted to server in clients for fast transmission. Existing works that attempt to reduce the amount of transmitted parameters have limitations: 1) the reduced number of parameters is not significant; 2) the performance of the global model is limited. In this paper, we propose a novel method called Fed-KGF that significantly reduces the amount of model parameters while improving the global model performance. Our goal is to reduce those transmitted parameters by reducing the number of convolution kernels. Specifically, we construct an incomplete model with a few representative convolution kernels, and propose Kernel Generation Function (KGF) to generate other convolution kernels to render the incomplete model to be a complete one. We discard those generated kernels after training local models, and solely transmit those representative kernels during training, thereby significantly reducing the transmitted parameters. Furthermore, there is a client-drift in the traditional FL because of the averaging method, which hurts the global model performance. We innovatively select one or few modules from all client models in a permutation way, and only aggregate the uploaded modules rather than averaging all modules to reduce client-drift, thus improving the global model performance and further reducing the transmitted parameters. Experimental results on both non-Independent and Identically Distributed (non-IID) and IID scenarios for image classification and object detection tasks demonstrate that our Fed-KGF outperforms SOTA FL models.

Color-tone represents the prominent color of an image, and training generative adversarial nets (GAN) to change color-tones of generated images is desirable in many applications. Advances such as HistoGAN can manipulate color-tones of generated images with a target image. Yet, there are challenges. Kullback-Leibler (KL) divergence adopted by HistoGAN might bring the color-tone mismatching, because it is possible to provide infinite score to a generator. Moreover, only relying on distribution estimation also produces images with lower fidelity in HistoGAN. To address these issues, we propose a new approach, named dynamic weights GAN (DW-GAN). We use two discriminators to estimate the distribution matching degree and details' similarity, with Laplacian operator and Hinge loss. Laplacian operator can help capture more image details, while Hinge loss is deduced from mean difference (MD) that could avoid the case of infinite score. To synthesize desired images, we combine the loss of the two discriminators with generator loss and set the weights of the two estimated scores to be dynamic through the previous discriminators' outputs, given that the training signal of a generator is from a discriminator. Besides, we innovatively integrate the dynamic weights into other GAN variants (e.g., HistoGAN and StyleGAN) to show the improved performance. Finally, we conduct extensive experiments on one industrial Fabric and seven public datasets to demonstrate the significant performance of DW-GAN in producing higher fidelity images and achieving the lowest Frechet inception distance (FID) scores over SOTA baselines.

We study the problem of classifying different cooking styles, based on the recipe. The difficulty is that the same food ingredients, seasoning, and the very similar instructions result in different flavors, with different cooking styles. Existing methods have limitations: they mainly focus on homogeneous data (e.g., instruction or image), ignoring heterogeneous data (e.g., flavor compound or ingredient), which certainly hurts the classification performance. This is because collecting enough available heterogeneous data of a recipe is a non-trivial task. In this paper, we present a new heterogeneous data augmentation method to improve classification performance. Specifically, we first construct a heterogeneous recipe graph network to represent heterogeneous data, which includes four main-stream types of heterogeneous data: ingredient, flavor compound, image, and instruction. Then, we draw a sequence of augmented graphs for Semi-Supervised learning through multinomial sampling. The probability distribution of sampling depends on the Cosine distance between the nodes of graph. In this way, we name our approach as Multinomial Sampling Graph Data Augmentation (MS-GDA). Extensive experiments demonstrate that MS-GDA significantly outperforms SOTA baselines on cuisine classification and region prediction with the recipe benchmark dataset. Code is available at https://github.com/LiangzheChen/MS-GDA .

Balancing image local detail enhancement with brightness enhancement has been a challenge. The images captured by wireless capsule endoscopy (WCE) are low-light and unclear. To this end, we propose an adaptive enhancement method for WCE images. Firstly, we use the guided filter to filter and smooth the WCE images to approximate its illumination component, and then the reflection component is obtained by decomposing it based on the Retinex model. Secondly, an adaptive Sigmoid function is obtained according to the positive correlation between the just-noticeable difference (JND) threshold of the illumination component and the gain parameter of the Sigmoid function, to adaptively enhance the illumination component, and then based on the Retinex model fusion with the reflection component. Finally, we combine with Gamma correction algorithm to enhance the contrast of the above results. Experimental results show that the proposed method can adaptively enhance the overall effect and local details of WCE images; and the feature extraction and matching effects are better than the classical enhancement algorithms, with an average increase of 88.8% and 59.1% respectively.

3D convolution networks play an important role in extracting spatiotemporal features in video action recognition. However, it usually brings a large number of paramters, which results in deployment difficulty in edge devices with limited memory space. Although lightweight 3DCNNs can reduce the mode size significantly, it causes a serious loss of accuracy. This paper proposes a novel approach to reduce the model size while preserves accuracy by combining lightweight networks with structural reparameterization. To reduce the model size, we propose 3D-DBB module, based on 2D Diverse Branch Block(DBB). Furthermore, we propose three structures based on 3D-DBB: (1) 3D depthwise convolution (called 3D-DBB-DepthWise), (2) the 3D pointwise convolution (called 3D-DBB-PointWise), and (3) reparameterizable depthwise separable structure (called DP3DBB), which is the concatenation of the two previous structures. We design and compare the effect of two different replacements for replacing depthwise separable structures in lightweight networks. Our method achieves 93.33% with only 0.42% loss in accuracy when the model size is only 1/50 of that of 3D-ResNeXt101.

Multi-agent surrounding is a collaborative task that uses multiple agents to surround a stationary or moving target. Multi-agent surrounding has a wide range of applications, such as area monitoring of unmanned ships, environmental monitoring, and exploration of unknown environments. Existing work pay attention to the case of one-to-one surrounding of targets by agents, but there is a lack of consideration for the case where agents are not one-to-one with the target. In this paper, we propose the concept of virtual target, which is used as a mediator to realize the generic multi-agent surrounding a target. The main idea is to surround the actual target with the virtual target, while the agents surround the virtual target, where the generation of the virtual target is based on any given surrounding graph and random sampling, and the performance of the surrounding is ensured by the virtual target control algorithm and the agent controlling algorithm. The Lyapunov stability analysis and simulation results show that the proposed approach can make the virtual target fit the actual target effectively, and the agents can surround the actual target with the shape of the virtual target effectively.

Imbalanced data cause deep neural networks to output biased results, and it becomes more serious when facing extremely imbalanced data regarding the outliers with tiny size (the ratio of the outlier size to the image size is around 0.05%). Many data argumentation models are proposed to supplement imbalanced data to alleviate biased results. However, the existing augmentation models cannot synthesize tiny outliers, which make the generated data unavailable. In this article, we propose a new augmentation model named extremely imbalanced data augmentation generative adversarial nets (EID-GANs) to address the extremely imbalanced data augmentation problem. First, we design a new penalty function by subtracting the outliers from the cropped region of generated instance to guide the generator to learn the features of outliers. After this, we combine the output value of the penalty function with the generator loss to jointly update the generator’s parameters with backpropagation. Second, we propose a new evaluation approach that adopts two outlier detectors withk-fold cross-validation to assess the availability of generated instances. We conduct extensive experiments to demonstrate the significant performance improvement of EID-GAN on two extremely imbalanced datasets, which are the industrial Piston and the Fabric datasets, and one general imbalanced dataset, i.e., the public DAGM dataset. The experimental results show that our EID-GAN outperforms the state-of-the-art (SOTA) augmentation models on different imbalanced datasets.

The generative adversarial network (GAN) is usually built from the centralized, independent identically distributed (i.i.d.) training data to generate realistic-like instances. In real-world applications, however, the data may be distributed over multiple clients and hard to be gathered due to bandwidth, departmental coordination, or storage concerns. Although existing works, such as federated learning GAN (FL-GAN), adopt different distributed strategies to train GAN models, there are still limitations when data are distributed in a non-i.i.d. manner. These studies suffer from convergence difficulty, producing generated data with low quality. Fortunately, we found that these challenges are often due to the use of a federated averaging strategy to aggregate local GAN models' updates. In this article, we propose an alternative approach to tackling this problem, which learns a globally shared GAN model by aggregating locally trained generators' updates with maximum mean discrepancy (MMD). In this way, we term our approach improved FL-GAN (IFL-GAN). The MMD score helps each local GAN hold different weights, making the global GAN in IFL-GAN getting converged more rapidly than federated averaging. Extensive experiments on MNIST, CIFAR10, and SVHN datasets demonstrate the significant improvement of our IFL-GAN in both achieving the highest inception score and producing high-quality instances.

For static scenes with multiple depth layers, the existing defocused image deblurring methods have the problems of edge ringing artifacts or insufficient deblurring degree due to inaccurate estimation of blur amount, In addition, the prior knowledge in non blind deconvolution is not strong, which leads to image detail recovery challenge. To this end, this paper proposes a blur map estimation method for defocused images based on the gradient difference of the boundary neighborhood, which uses the gradient difference of the boundary neighborhood to accurately obtain the amount of blurring, thus preventing boundary ringing artifacts. Then, the obtained blur map is used for blur detection to determine whether the image needs to be deblurred, thereby improving the efficiency of deblurring without manual intervention and judgment. Finally, a non blind deconvolution algorithm is designed to achieve image deblurring based on the blur amount selection strategy and sparse prior. Experimental results show that our method improves PSNR and SSIM by an average of 4.6% and 7.3%, respectively, compared to existing methods. Experimental results show that our method outperforms existing methods. Compared with existing methods, our method can better solve the problems of boundary ringing artifacts and detail information preservation in defocused image deblurring.

Abstract Android apps have played important roles in daily life and work. To meet the new requirements from users, the apps encounter frequent updates, which involves a large quantity of code commits. Previous studies proposed to apply Just‐in‐Time (JIT) defect prediction for apps to timely identify whether the new code commits can introduce defects into apps, aiming to assure their quality. In general, high‐quality features are benefits for improving the classification performance. In addition, the number of defective commit instances is much fewer than that of clean ones, that is the defect data is class imbalanced. In this study, a novel compositional model, called KPIDL, is proposed to conduct the JIT defect prediction task for Android apps. More specifically, KPIDL first exploits a feature learning technique to preprocess original data for obtaining better feature representation, and then introduces a state‐of‐the‐art cost‐sensitive cross‐entropy loss function into the deep neural network to alleviate the class imbalance issue by considering the prior probability of the two types of classes. The experiments were conducted on a benchmark defect data consisting of 15 Android apps. The experimental results show that the proposed KPIDL model performs significantly better than 25 comparative methods in terms of two effort‐aware performance indicators in most cases.

Data usually resides on a manifold, and the minimal dimension of such a manifold is called its intrinsic dimension. This fundamental data property is not considered in the generative adversarial network (GAN) model along with its its variants; such that original data and generated data often hold different intrinsic dimensions. The different intrinsic dimensions of both generated and original data may cause generated data distribution to not match original data distribution completely, and it certainly will hurt the quality of generated data. In this study, we first show that GAN is often unable to generate simulation data, holding the same intrinsic dimension as the original data with both theoretical analysis and experimental illustration. Next, we propose a new model, called Hausdorff GAN, which removes the issue of different intrinsic dimensions and introduces the Hausdorff metric into GAN training to generate higher quality data. This provides new insights into the success of Hausdorff GAN. Specifically, we utilize a mapping function to map both original and generated data into the same manifold. We then calculate the Hausdorff distance to measure the difference between the mapped original data and the mapped generated data, toward pushing generated data to the side of original data. Finally, we conduct extensive experiments (using MNIST, CIFAR10, and CelebA datasets) to demonstrate the significant performance improvement of the Hausdorff GAN in achieving the largest Inception Score and the smallest Frechet inception distance (FID) score as well as producing diverse generated data at different resolutions.

Time series classification (TSC) is widely used in various real-world applications such as human activity recognition, smart city governance, etc. Unfortunately, due to different reasons, only part of time series could be collected which may obviously degrade the performance of time series classifiers. To alleviate this problem, time series augmentation aims to generate synthetic time series by learning useful features from collected time series. As the popular generative model, generative adversarial networks (GAN) is regarded as a promising model for time series augmentation. However, applying GAN to the time series data suffers from a challenge in which the generated instances hold low quality but the model has gotten saturation. In this paper, for time series augmentation, we proposed TSA-GAN which is a robust GAN model with a self-adaptive recovering strategy to solve this problem. On 85 datasets of the UCR 2015 archive, our proposed TSA-GAN helps time series classifiers achieve performance improvements ranging from 8.3% to 12.5%, which is far better than the baseline.

Due to the dynamic characteristics of news and user preferences, personalized recommendation is a challenging problem. Traditional recommendation methods simply focus on current reward, which just recommend items to maximize the number of current clicks. And this may reduce users' interest in similar items. Although the news recommendation framework based on deep reinforcement learning preciously proposed (i.e, DRL, based on deep Q-learning) has the advantages of focusing on future total rewards and dynamic interactive recommendation, it has two issues. First, its exploration method is slow to converge, which may bring new users a bad experience. Second, it is hard to train on off-line data set because the reward is difficult to be determined. In order to address the aforementioned issues, we propose a framework named DEN-DQL for news recommendation based on deep Q-learning with double exploration networks. Also, we develop a new method to calculate rewards and use an off-line data set to simulate the online news clicking environment to train DEN-DQL. Then, the well trained DEN-DQL is tested in the online environment of the same data set, which demonstrates at least 10% improvement of the proposed DEN-DQL.

In recent years, Convolutional Neural Networks (CNN s) have been widely applied in various applications due to its powerful learning capability. However, its lack of explainability hinders its further usage in tasks requiring high reliability. Therefore, interpretability technique is the key to the application and deployment of CNN models. As a typical interpretability technique for CNN, Class Activation Map (CAM) utilizing the gradient based weights and activation map is widely applied to traditional CNN models for offering visual interpretability. However, the activation map adopted by CAM cannot loyally quantify the relevance between input samples and activation values. Hence, in this paper, we propose a new interpretability approach called Salience-CAM employing salience scores to accurately measure the relevance between input samples and activation values. To evaluate the effectiveness of Salience-CAM, comprehensive experiments are conducted on 6 selected time series datasets. By leveraging an evaluation algorithm proposed in this paper, the experimental results show that our proposed Salience-CAM outperforms the baseline by discovering more discriminative features.

Manipulating facial images results in negative influences on the social association, with deep generative models. Although many detection methods have been proposed, they have either designed sophisticated neural networks that lack enough interpretability, or found defects specific to one manipulation method. To address this issue, we propose a new approach to explore the defects of fake facial images after wavelet transform and call it GFS (Grayscale Frequency Statistics). First, we utilize Haar wavelet transformation to decompose the image into low-frequency approximation, horizontal detail, vertical detail, and diagonal detail. The GFS of real and fake images exhibit different distribution and forms in these four subbands. We qualitatively analyze these differences and quantify them as weights. Then, these four subband images are used to train four CNNs respectively, and the obtained detection results also verify the differences in GFS. After that, we combine the prediction results of the four CNNs and the corresponding weights to further improve the detection performance. We conduct extensive experiments on 11 datasets generated by various facial manipulation methods, and the superior results show the effectiveness of our proposed approach. Our findings indicate that the fake images generated by the current facial manipulation methods cannot simulate real images in wavelet-domain.

Clustering is an important unsupervised learning method with serious challenges when data is sparse and high-dimensional. Generated clusters are often evaluated with general measures, which may not be meaningful or useful for practical applications and domains. Using a distance metric, a clustering algorithm searches through the data space, groups close items into one cluster, and assigns far away samples to different clusters. In many real-world applications, the number of dimensions is high and data space becomes very sparse. Selection of a suitable distance metric is very difficult and becomes even harder when categorical data is involved. Moreover, existing distance metrics are mostly generic, and clusters created based on them will not necessarily make sense to domain-specific applications. One option to address these challenges is to integrate domain-defined rules and guidelines into the clustering process. In this work we propose a GAN-based approach called Catalysis Clustering to incorporate domain knowledge into the clustering process. With GANs we generate catalysts, which are special synthetic points drawn from the original data distribution and verified to improve clustering quality when measured by a domain-specific metric. We then perform clustering analysis using both catalysts and real data. Final clusters are produced after catalyst points are removed. Experiments on two challenging real-world datasets clearly show that our approach is effective and can generate clusters that are meaningful and useful for real-world applications.

As a typical task of time series mining, Time Series Classification (TSC) has attracted lots of attention from both researchers and domain experts due to its broad applications. To get rid of costly hand-crafting feature engineering process, deep learning techniques are applied for automatic feature extraction, which shows competitive or even better performance compared with state-of-the-art TSC solutions. However, on time series datasets presenting complex patterns, neither 1-Nearest-Neighbour classifier nor deep learning models are capable of achieving satisfactory classification accuracy which motivates us to explore new time series representations to help classifiers further improve the classification accuracy. In this paper, by building the binary distribution tree, an approach to time series classification based on deep learning models using new representations is proposed. By conducting comprehensive experiments over 6 most challenging time series datasets and comparing experimental results of the same classifier using the proposed representation or not, the potential of the proposed approach to enhancing time series classification accuracy is validated with a bunch of helpful findings.