We consider the hard-label based black-box adversarial attack setting which solely observes the target model’s predicted class. Most of the attack methods in this setting suffer from impractical number of queries required to achieve a successful attack. One approach to tackle this drawback is utilising the adversarial transferability between white-box surrogate models and black-box target model. However, the majority of the methods adopting this approach are soft-label based to take the full advantage of zeroth-order optimisation. Unlike mainstream methods, we propose a new practical setting of hard-label based attack with an optimisation process guided by a pre-trained surrogate model. Experiments show the proposed method significantly improves the query efficiency of the hard-label based black-box attack across various target model architectures. We find the proposed method achieves approximately 5 times higher attack success rate compared to the benchmarks, especially at the small query budgets as 100 and 250.

Deep metric learning aims to create a feature space where the projected samples have maximized inter-class and minimized intra-class distance. Most approaches employ only distance-based metrics to achieve this objective, but neglect other properties of the projections in the embedding space, such as their density, sparsity and presence of outliers. In this paper, we propose a novel density-based regulariser, LOFReg, designed to be used as a complement to previously proposed distance-based metric learning loss functions for the re-identification (ReID) and few-shot classification (FSC) tasks. Our method is based on the well-known, in anomaly detection literature, local outlier factor (LOF) algorithm, which estimates the local density deviation of a data point with respect to its neighbours. These measurements are used in our regularisation methodology to achieve an embedding space of evenly distributed samples and to increase the generalisation ability of the model compared to solely distance-based learning. Comprehensive experiments on four publicly available datasets for ReID and FSC, demonstrate consistent improvement against previously proposed metric learning loss functions. Particularly, our experiments show up to 5% improvement in ReID settings, up to 13% in FSC settings and up to 6% improvement against other state-of-art regularisers.

Traditional vehicle re-identification (ReID) approaches, based on visible spectrum data achieve high performance, but have limited capability of real-life applications, as they perform poorly under occluded visibility conditions, such as night-time and bad weather. In such cases, the use of infrared spectrum thermal imagery offers complementary and persistent information when the visual data contribution is inadequate. It is therefore highly beneficial to create a vehicle ReID framework that can exploit both modalities, if available, and is able to apply cross-modality matching, when ReID is required across single modal sensors. This is an extremely challenging task because the nature of the two data modalities induces high discrepancy between the two domains. In this paper we propose a robust end-to-end 2-stream vehicle ReID system that aims to solve the multi-modal and cross-modal ReID problem together by minimising the domain shift between infrared and visible distributions. Our framework consists of a shared network part, following the 2 independent streams, to extract shareable features, along with a domain alignment technique to narrow the gap between the two domains and inter-modality learning to address the cross-domain matching problem. The proposed system achieves state-of-the-art results on RGBN300 dataset, when both modalities are available at inference time. Moreover, our work is the first to explore the cross-modal settings for vehicle ReID and attempts to reduce the performance drop of the cross-modal scenario, when the query and the gallery images come from different modalities. We first measure the baseline cross-modal performance, and then prove that the proposed method improves up to 11% in mAP and 16% in Rank-1 score against the baseline.

We propose a new method for single-camera real-world 3-D human pose estimation. Our method uses multitask training together with iterative pose refinement using a novel conditional attention mechanism. For iterative pose refinement, the output of each convolutional layer is conditioned on the latest pose estimate, using a conditioned squeeze-and-excitation network architecture that incorporates novel feedback connections. Multitask training on both an in-the-wild 2-D pose dataset and a controlled 3-D pose dataset allows for real-world 3-D pose estimation without the need for a large-scale in-the-wild 3-D pose dataset, which is unavailable. Experiments are performed on several real-world datasets, as well as the Human 3.6 Million and HumanEva-I datasets, to show that the combined attention mechanism, iterative refinement scheme, and multitask training allow us to achieve robust and competitive performance with only a simple network architecture. In addition, we show that our method is efficient enough to run on commodity hardware, producing pose estimates in real time.

Simultaneous feature selection and clustering is a major challenge in unsupervised learning. In particular, there has been significant research into saliency measures for features that result in good clustering. However, as datasets become larger and more complex, there is a need to adopt a finer-grained approach to saliency by measuring it in relation to a part of a model. Another issue is learning the feature saliency and advanced model parameters. We address the first by presenting a novel Gaussian mixture model, which explicitly models the dependency of individual mixture components on each feature giving a new component-based feature saliency measure. For the second, we use Markov Chain Monte Carlo sampling to estimate the model and hidden variables. Using a synthetic dataset, we demonstrate the superiority of our approach, in terms of clustering accuracy and model parameter estimation, over an approach using a model-based feature saliency with expectation maximisation. We performed an evaluation of our approach with six synthetic trajectory datasets obtaining an average clustering accuracy of 97 percent. To demonstrate the generality of our approach, we applied it to a network traffic flow dataset obtaining an accuracy of 93 percent for intrusion detection. Finally, we performed a comparison with state-of-the-art clustering techniques using three real-world trajectory datasets of vehicle traffic. Our approach achieved an average clustering accuracy of 96 percent compared to 77-95 percent for the other techniques. In conclusion, for the datasets considered, component based feature saliency measures gave improved clustering over those based on whole models.

We present DANdroid, a novel Android malware detection model using a deep learning Discriminative Adversarial Network (DAN) that classifies both obfuscated and unobfuscated apps as either malicious or benign. Our method, which we empirically demonstrate is robust against a selection of four prevalent and real-world obfuscation techniques, makes three contributions. Firstly, an innovative application of discriminative adversarial learning results in malware feature representations with a strong degree of resilience to the four obfuscation techniques. Secondly, the use of three feature sets; raw opcodes, permissions and API calls, that are combined in a multi-view deep learning architecture to increase this obfuscation resilience. Thirdly, we demonstrate the potential of our model to generalize over rare and future obfuscation methods not seen in training. With an overall dataset of 68,880 obfuscated and unobfuscated malicious and benign samples, our multi-view DAN model achieves an average F-score of 0.973 that compares favourably with the state-of-the-art, despite being exposed to the selected obfuscation methods applied both individually and in combination.

Semantic models derived from visual information have helped to overcome some of the limitations of solely text-based distributional semantic models. Researchers have demonstrated that text and image-based representations encode complementary semantic information, which when combined provide a more complete representation of word meaning, in particular when compared with data on human conceptual knowledge. In this work, we reveal that these vision-based representations, whilst quite effective, do not make use of all the semantic information available in the neural network that could be used to inform vector-based models of semantic representation. Instead, we build image-based meta-embeddings from computer vision models, which can incorporate information from all layers of the network, and show that they encode a richer set of semantic attributes and yield a more complete representation of human conceptual knowledge.

Researchers have recently demonstrated that tying the neural weights between the input look-up table and the output classification layer can improve training and lower perplexity on sequence learning tasks such as language modelling.Such a procedure is possible due to the design of the softmax classification layer, which previous work has shown to comprise a viable set of semantic representations for the model vocabulary, and these these output embeddings are known to perform well on word similarity benchmarks.In this paper, we make meaningful comparisons between the input and output embeddings and other SOTA distributional models to gain a better understanding of the types of information they represent.We also construct a new set of word embeddings using the output embeddings to create locally-optimal approximations for the intermediate representations from the language model.These locally-optimal embeddings demonstrate excellent performance across all our evaluations.

Identifying vehicles across a network of cameras with non-overlapping fields of view remains a challenging research problem due to scene occlusions, significant inter-class similarity and intra-class variability. In this paper, we propose an end-to-end multi-level re-identification network that is capable of successfully projecting same identity vehicles closer to one another in the embedding space, compared to vehicles of different identities. Robust feature representations are obtained by combining features at multiple levels of the network. As for the learning process, we employ a recent state-of-the-art structured metric learning loss function previously applied to other retrieval problems and adjust it to the vehicle re-identification task. Furthermore, we explore the cases of image-to-image, image-to-video and video-to-video similarity metric. Finally, we evaluate our system and achieve great performance on two large-scale publicly available datasets, CityFlow-ReID and VeRi-776. Compared to most existing state-of-art approaches, our approach is simpler and more straightforward, utilizing only identity-level annotations, while avoiding post-processing the ranking results (re-ranking) at the testing phase.

Steven Derby, Paul Miller, Barry Devereux. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

In this paper, we propose a video-based person re-identification system for wide area tracking based on a recurrent neural network architecture. Given short video sequences of a person, generated by a tracking algorithm, our video re-identification algorithm links these tracklets in full trajectories across a network of non-overlapping cameras in an open-world scenario. In our system, features are first extracted from each frame using a convolutional neural network. Then, a recurrent layer combines information across time-steps. The features from all time-steps are finally combined using temporal pooling to give an overall appearance feature for the complete sequence. Our system is trained to perform re-identification using a Siamese network architecture. Experiments are conducted on the iLIDS-VID and PRID-2011 video re-identification data sets as well as in the DukeMTMC multi-camera tracking data set.

Distributional models provide a convenient way to model semantics using dense embedding spaces derived from unsupervised learning algorithms.However, the dimensions of dense embedding spaces are not designed to resemble human semantic knowledge.Moreover, embeddings are often built from a single source of information (typically text data), even though neurocognitive research suggests that semantics is deeply linked to both language and perception.In this paper, we combine multimodal information from both text and image-based representations derived from state-of-the-art distributional models to produce sparse, interpretable vectors using Joint Non-Negative Sparse Embedding.Through indepth analyses comparing these sparse models to human-derived behavioural and neuroimaging data, we demonstrate their ability to predict interpretable linguistic descriptions of human ground-truth semantic knowledge.

In this paper, we propose a novel android malware detection system that uses a deep convolutional neural network (CNN). Malware classification is performed based on static analysis of the raw opcode sequence from a disassembled program. Features indicative of malware are automatically learned by the network from the raw opcode sequence thus removing the need for hand-engineered malware features. The training pipeline of our proposed system is much simpler than existing n-gram based malware detection methods, as the network is trained end-to-end to jointly learn appropriate features and to perform classification, thus removing the need to explicitly enumerate millions of n-grams during training. The network design also allows the use of long n-gram like features, not computationally feasible with existing methods. Once trained, the network can be efficiently executed on a GPU, allowing a very large number of files to be scanned quickly.

Person reidentification involves recognizing a person across nonoverlapping camera views, with different pose, illumination, and camera characteristics. We propose to tackle this problem by training a deep convolutional network to represent a person's appearance as a low-dimensional feature vector that is invariant to common appearance variations encountered in the reidentification problem. Specifically, a Siamese network architecture is used to train a feature extraction network using pairs of similar and dissimilar images. We show that the use of a novel multitask learning objective is crucial for regularizing the network parameters in order to prevent overfitting due to the small size of the training data set. We complement the verification task, which is at the heart of reidentification, by training the network to jointly perform verification and identification and to recognize attributes related to the clothing and pose of the person in each image. In addition, we show that our proposed approach performs well even in the challenging cross-data set scenario, which may better reflect real-world expected performance.

In this paper we propose a novel recurrent neural network architecture for video-based person re-identification. Given the video sequence of a person, features are extracted from each frame using a convolutional neural network that incorporates a recurrent final layer, which allows information to flow between time-steps. The features from all timesteps are then combined using temporal pooling to give an overall appearance feature for the complete sequence. The convolutional network, recurrent layer, and temporal pooling layer, are jointly trained to act as a feature extractor for video-based re-identification using a Siamese network architecture. Our approach makes use of colour and optical flow information in order to capture appearance and motion information which is useful for video re-identification. Experiments are conduced on the iLIDS-VID and PRID-2011 datasets to show that this approach outperforms existing methods of video-based re-identification.

Most traditional data mining algorithms struggle to cope with the sheer scale of data efficiently. In this paper, we propose a general framework to accelerate existing algorithms to cluster large-scale datasets which contain large numbers of attributes, items, and clusters. Our framework makes use of locality sensitive hashing to significantly reduce the cluster search space. We also theoretically prove that our framework has a guaranteed error bound in terms of the clustering quality. This framework can be applied to a set of centroid-based clustering algorithms that assign an object to the most similar cluster, and we adopt the popular K-Modes categorical clustering algorithm to present how the framework can be applied. We validated our framework with five synthetic datasets and a real world Yahoo! Answers dataset. The experimental results demonstrate that our framework is able to speed up the existing clustering algorithm between factors of 2 and 6, while maintaining comparable cluster purity.

In this paper we explore ways to address the issue of dataset bias in person re-identification by using data augmentation to increase the variability of the available datasets, and we introduce a novel data augmentation method for re-identification based on changing the image background. We show that use of data augmentation can improve the cross-dataset generalisation of convolutional network based re-identification systems, and that changing the image background yields further improvements.

In this paper we propose a graph stream clustering algorithm with a unified similarity measure on both structural and attribute properties of vertices, with each attribute being treated as a vertex. Unlike others, our approach does not require an input parameter for the number of clusters, instead, it dynamically creates new sketch-based clusters and periodically merges existing similar clusters. Experiments on two publicly available datasets reveal the advantages of our approach in detecting vertex clusters in the graph stream. We provide a detailed investigation into how parameters affect the algorithm performance. We also provide a quantitative evaluation and comparison with a well-known offline community detection algorithm which shows that our streaming algorithm can achieve comparable or better average cluster purity.

In this paper we extend the minimum-cost network flow approach to multi-target tracking, by incorporating a motion model, allowing the tracker to better cope with long term occlusions and missed detections. In our new method, the tracking problem is solved iteratively: Firstly, an initial tracking solution is found without the help of motion information. Given this initial set of track lets, the motion at each detection is estimated, and used to refine the tracking solution. Finally, special edges are added to the tracking graph, allowing a further revised tracking solution to be found, where distant track lets may be linked based on motion similarity. Our system has been tested on the PETS S2.L1 and Oxford town-center sequences, outperforming the baseline system, and achieving results comparable with the current state of the art.

We introduce a novel dual-stage algorithm for online multitarget tracking in realistic conditions. In the first stage, the problem of data association between tracklets and detections, given partial occlusion, is addressed using a novel occlusion robust appearance similarity method. This is used to robustly link tracklets with detections without requiring explicit knowledge of the occluded regions. In the second stage, tracklets are linked using a novel method of constraining the linking process that removes the need for ad-hoc tracklet linking rules. In this method, links between tracklets are permitted based on their agreement with optical flow evidence. Tests of this new tracking system have been performed using several public datasets.

This paper introduces a momentum-like regularisation term for the region-based Mixture of Gaussians framework. Momentum term has long been used in machine learning, especially in backpropagation algorithms to improve the speed of convergence and subsequently their performance. Here, we prove the convergence of the online gradient method with a momentum term and apply it to background modelling by using it in the update equations of the region-based Mixture of Gaussians algorithm. It is then shown with the help of experimental evaluation on both simulated data and well known video sequences that these regularised updates help improve the performance of the algorithm.

This paper presents an event recognition framework, based on Dempster-Shafer theory, that combines evidence of events from low-level computer vision analytics. The proposed method employing evidential network modelling of composite events, is able to represent uncertainty of event output from low level video analysis and infer high-level events with semantic meaning along with degrees of belief. The method has been evaluated on videos taken of subjects entering and leaving a seated area. This has relevance to a number of transport scenarios, such as onboard buses and trains, and also in train stations and airports. Recognition results of 78% and 100% for four composite events are encouraging.

The Richardson-Lucy algorithm is one of the most important in image deconvolution. However, a drawback is its slow convergence. A significant acceleration was obtained using the technique proposed by Biggs and Andrews (BA), which is implemented in the deconvlucy function of the image processing MATLAB toolbox. The BA method was developed heuristically with no proof of convergence. In this paper, we introduce the heavy-ball (H-B) method for Poisson data optimization and extend it to a scaled H-B method, which includes the BA method as a special case. The method has a proof of the convergence rate of O(K(-2)), where k is the number of iterations. We demonstrate the superior convergence performance, by a speedup factor of five, of the scaled H-B method on both synthetic and real 3D images.

We address the problem of multi-target tracking in realistic crowded conditions by introducing a novel dual-stage online tracking algorithm. The problem of data-association between tracks and detections, based on appearance, is often complicated by partial occlusion. In the first stage, we address the issue of occlusion with a novel method of robust data-association, that can be used to compute the appearance similarity between tracks and detections without the need for explicit knowledge of the occluded regions. In the second stage, broken tracks are linked based on motion and appearance, using an online-learned linking model. The online-learned motion-model for track linking uses the confident tracks from the first stage tracker as training examples. The new approach has been tested on the town centre dataset and has performance comparable with the present state-of-the-art.

Modelling pixels using mixture of Gaussian distributions is a popular approach for removing background in video sequences. This approach works well for static backgrounds because the pixels are assumed to be independent of each other. However, when the background is dynamic, this is not very effective. In this paper, we propose a generalisation of the algorithm where the spatial relationship between pixels is taken into account. In essence, we model regions as mixture distributions rather than individual pixels. Using experimental verification on various video sequences, we show that our method is able to model and subtract backgrounds effectively in scenes with complex dynamic textures.

In intelligent video surveillance systems, scalability (of the number of simultaneous video streams) is important. Two key factors which hinder scalability are the time spent in decompressing the input video streams, and the limited computational power of the processor. This paper demonstrates how a combination of algorithmic and hardware techniques can overcome these limitations, and significantly increase the number of simultaneous streams. The techniques used are processing in the compressed domain, and exploitation of the multicore and vector processing capability of modern processors. The paper presents a system which performs background modeling, using a Mixture of Gaussians approach. This is an important first step in the segmentation of moving targets. The paper explores the effects of reducing the number of coefficients in the compressed domain, in terms of throughput speed and quality of the background modeling. The speedups achieved by exploiting compressed domain processing, multicore and vector processing are explored individually. Experiments show that a combination of all these techniques can give a speedup of 170 times on a single CPU compared to a purely serial, spatial domain implementation, with a slight gain in quality.

A multi-camera system is described for 3-D localisation of subjects within a confined space. In particular, we present a novel neighbourhood association algorithm to solve the problem of associating detections in multiple camera views with subjects. To evaluate our approach, experiments were conducted using multiple view video sequences of up to four subjects simulating typical passenger behaviour on a bus. ROC curves were generated for three different versions which showed that for smaller values of the neighbourhood radius parameter, the system tended to over-estimate the number of subjects. However, increasing the radius reduced the over-estimation from 60% to 5%.

Mixture of Gaussians (MoG) modelling [13] is a popular approach to background subtraction in video sequences. Although the algorithm shows good empirical performance, it lacks theoretical justification. In this paper, we give a justification for it from an online stochastic expectation maximization (EM) viewpoint and extend it to a general framework of regularized online classification EM for MoG with guaranteed convergence. By choosing a special regularization function, l1norm, we derived a new set of updating equations for l1regularized online MoG. It is shown empirically that l1regularized online MoG converge faster than the original online MoG.

We introduce a mechanism for improving the result of elastic registration for simultaneously acquired low&contrast fluorescent mi croscopy images. B&spline basis functions are commonly used for modelling nonrigid image deformations. We present a novel approach that improves the result of B&spline based registration process. This approach monitors the behaviour of local similarity measure values during optimization in order to recognize and exclude the pixels which drive the optimization into local optima. These are low intensity pixels, mostly in background regions but also in data areas, where there is no or little useful image structure to guide registration. Such pixels play a fallacious role in the registration process due to their high frequency content and pseudorandom contribution in the overall similarity measure and its derivatives. We show that excluding such pixels during the optimization considerably improves the registration result. This approach can reduce the average registration error up to several pixels and reduce the number of iterations down to one&third of the current implementations.

This paper mainly addresses the problem of age classification. Image fe atures can be extracted using a difference of Gaussian filter followed by Radon tran sform. The relevance and importance of these features are determined in a scaling support vector machine classifier, where zero weights are assigned to irrelevant varia bles. To enhance the quality of feature selection, we introduce entropy estimation to the scaling c lassifier. Experimental results demonstrate that the proposed algorithm leads to better recognition accuracy than the state of the art.

This paper presents Yagada, an algorithm to search labelled graphs for anomalies using both structural data and numeric attributes. Yagada is explained using several security-related examples and validated with experiments on a physical Access Control database. Quantitative analysis shows that in the upper range of anomaly thresholds, Yagada detects twice as many anomalies as the best-performing numeric discretization algorithm. Qualitative evaluation shows that the detected anomalies are meaningful, representing a combination of structural irregularities and numerical outliers.

The Intelligent Sensor Information System (ISIS) is described. ISIS is an active CCTV approach to reducing crime and anti-social behavior on public transport systems such as buses. Key to the system is the idea of event composition, in which directly detected atomic events are combined to infer higher-level events with semantic meaning. Video analytics are described that profile the gender of passengers and track them as they move about a 3-D space. The overall system architecture is described which integrates the on-board event recognition with the control room software over a wireless network to generate a real-time alert. Data from preliminary data-gathering trial is presented.

The use of single and dual-camera approaches to locating a subject in a 3-D cluttered space is investigated. Specifically, we investigate the case where the lower portion of the body may be occluded, e.g., by a chair on a bus. Experiments were conducted involving eleven subjects moving along a pre-designated route within a cluttered space. For each time instant the position of each subject was manually estimated and compared to that produced automatically. The dual camera approach was found to give significantly better performance than the single camera approach. It was found that inaccurate bounding of the lowest part of the subject, due to occlusion, led to localisation errors in range as large as 10m for the latter. Using the side bounds of the detected object, which were found to be robust, accurate azimuth estimates can be obtained for a single camera. The dual-camera approach exploits the greater degree of accuracy in azimuth to estimate the range through triangulation, giving average localisation errors of 40cm over the space of interest.

Demand for bus surveillance is growing due to the increased threats of terrorist attack, vandalism and litigation. However, CCTV systems are traditionally used in forensic mode, precluding an in-time reaction to an event. In this paper, we introduce a real-time event composition framework which can support the instant recognition of emergent events based on uncertain or imperfect information gathered from multiple sources. This framework deploys a rule-based reasoning component that can infer malicious situations (composite events) from a set of correlated atomic events. These are recognized by applying analytic algorithms to the multimedia contents of bus surveillance data. We demonstrate the significance and usefulness of our framework with a case study of an on-going bus surveillance project.

This paper describes the anatomy of a pilot surveillance system with a speech-based interface for content-based retrieval of video data. The proposed system relies on an ontology-based information sharing architecture and lets components of the system communicate among each other through TCP/IP communication channels. The aim of developing the pilot system was to explore dependencies between image analysis, event detection, video annotation, and speech- based retrieval of the video content in the context of a broader spoken dialogue system.

A change detection framework which fuses both spatial and temporal data using fuzzy if-then rules is presented. Temporal data is used on a per-pixel basis to monitor the sequence for changes by employing a fuzzy codebook model. Spatial data is gathered using a fuzzy multithresholding algorithm that decomposes the RGB color space into three color pair histograms. This system is found to be robust to noise and allows the algorithm to process successfully even when the underlying sequences result in under-segmentation of the spatial data.

A motion detection framework is presented that integrates spatial and temporal data by decomposition of the feature space. Temporal data is employed to monitor frame changes on a per-pixel basis, whilst the spatial data is used to provide region-based segmentation of image frames. The thresholded difference image from the change detection is then fused with the segmented image. It is found that this framework greatly improves robustness to noise, while retaining and, in some cases, improving the sensitivity to regions of interest. Further, the decomposition reduces the time required to run due to the reduction in computational complexity.

An automated unsupervised technique, based upon a Bayesian framework, for the segmentation of low light level imagery is proposed. Primarily, Mixture Modelling is used to provide a baseline estimate. This estimate is then refined to consider spatial correlations using Markov Random Field (MRF) Modelling. The technique has been implemented assuming low-light level Poisson statistics, and the results compared to the more widely used assumption of Gaussian statistics. Investigations revealed the Poisson technique quantitatively outperforms the Gaussian technique for synthetic low light imagery, both before and after avalanche multiplication, via the multiplication register of a Low-light Level Charge Coupled Device (L3CCD). The technique was then applied to the task of segmenting a biomedical dataset obtained from a L3CCD. Qualitative results were promising, again showing improvement over the Gaussian technique. 1

An independent moving objection detection system is described which uses a colour background model for a nonstationary sensor. Over time the RGB distribution of a static point in a scene, observed from a static or moving camera, is cylindrical rather than spherical and is modelled as such. The background model is recursively updated over time and is robustly aligned to the most recent frame. Discrepancies between the current observed frame and the background model are considered foreground events and are determined using a classification framework which combines per-pixel colour and brightness differences. In principle the colour (RGB) images provide greater discriminative power than greyscale (brightness). However for several sequences tested, which are of road scenes from an airborne sensor, it was found that the colour information did not improve the classification of foreground pixels significantly.

We present a novel visual representation, called local co-occurring patterns (LCPs), which consists of characteristic local features and the statistical co-occurance relations between them. The LCPs can be discovered using an associate rule mining algorithm. Experiments show that LCPs widely exist in a large image corpus, and are more discriminant than individual local features in visual categorization tasks such as subcategory and face recognition. Furthermore, state-of-the-art categorization performance was achieved on two test data-sets.