Search

In this paper, we proposed a more robust supervised hashing framework based on the Cauchy loss function and Supervised Discrete Hashing (SDH) called Robust Supervised Discrete Hashing (RSDH), which can learn a robust subspace consisted of binary codes. The Cauchy loss is used to measure the error between the label matrix and the product of the decomposed matrices. RSDH can not only reduce the outliers and noise of the hashing codes, but also achieve the more satisfactory retrieval effect. Image retrieval experiments demonstrate that RSDH performs better than the other hashing methods.

Interval-valued data are needed to manage either the uncertainty related to measurements, or the variability inherent to the description of complex objects representing group of individuals. A number of regression methods suitable to interval variables describing variability of complex objects are already available. However, less attention has been given to methods that, simultaneously, take into account the full interval information and are resistant to interval outlier observations, even with the frequent presence of atypical observations on interval-valued data sets. This paper proposes a new robust linear regression method for interval variables, where the presence of outliers either in the center or in the radius penalize both the center and the radius regression models. Moreover, the interval observations with outliers on both center and radius are more penalized than those observations with outliers only in the center (or in the radius). Besides, this paper provides a suitable iterative algorithm to estimate the parameters of the proposed method. The algorithm estimates the parameters of the center (or of the radius) model taking into account both information of the center and the radius. The convergence and time complexity of the iterative algorithm are also presented. Finally, the performance of the new method is compared with some previous robust regression approaches and evaluated on synthetic and real interval-valued data sets.

Recurrent broad learning system (RBLS) is an effective learning method for processing sequential data. By replacing the enhancement nodes of the broad learning system with recurrent structure, RBLS obtains the capacity to capture the dynamic characteristics of time series data. However, RBLS is derived under the minimum mean square error (MMSE) criterion, which is sensitive to outliers. Moreover, RBLS is insufficient for online sequential learning. To address these limitations, we propose a novel online robust echo state structure based RBLS (OR-ESBLS). In OR-ESBLS, kernel recursive maximum correntropy (KRMC) is introduced to both enhance the robustness and discover the nonlinear characteristics of feature nodes in an online manner. To reduce the heavy computational requirements caused by the kernel method, a Quasi-Monte Carlo (QMC) based Random Fourier Feature (RFF) is utilized for kernel approximation. Furthermore, we adopt the randomized sparse reservoir as the enhancement nodes of RBLS, which can much more efficiently capture dynamic information of the data in the sequential learning setting. Experiments on both synthetic and real-world datasets are reported. The results show that the proposed OR-ESBLS can provide superior performance in online sequential time series prediction.

Video engagement is important in online advertisements where there is no physical interaction with the consumer. Engagement can be directly measured as the number of seconds after which a consumer skips an advertisement. In this paper, we propose a model to predict video engagement of an advertisement using only a few samples. This allows for early identification of poor quality videos. This can also help identify advertisement frauds where a robot runs fake videos behind the name of well-known brands. We leverage on the fact that videos with high engagement have similar viewing patterns over time. Hence, we can create a similarity network of videos and use a graph-embedding model called DeepWalk to cluster videos into significant communities. The learned embedding is able to identify viewing patterns of fraud and popular videos. In order to assess the impact of a video, we also consider how the view counts increase or decrease over time. This results in a heterogeneous graph where an edge indicates similar video engagement or history of view counts between two videos. Since it is difficult to find labelled samples for ‘fraud’ video, we leverage on a one-class model that can determine ‘fraud’ videos with outlier or abnormal behavior. The proposed model outperforms baselines in F-measure by over 20 % .

Seeking reliable correspondences is a fundamental and significant work in computer vision. Recent work has demonstrated that the task can be effectively accomplished by utilizing a deep learning network based on multi-layer perceptrons, which uses the context normalization to deal with the input. However, the context normalization treats each correspondence equally, which will reduce the representation capability of potential inliers. To solve this problem, we propose a novel and effective Local-Global Self-Attention (LAGA) layer based on the self-attention mechanism, to capture contextual information of potential inliers from coarse to fine, and suppress outliers at the same time in processing the input. The global self-attention module is able to capture abundant global contextual information in the whole image, and the local self-attention module is used to obtain rich local contextual information in the local region. After that, to obtain richer contextual information and feature maps with stronger representative capacity, we combine global and local contextual information. The extensive experiments have shown that the networks with our proposed LAGA layer perform better than the original and other comparative networks in outdoor and indoor scenes for outlier removal and camera pose estimation tasks.

Multi-view clustering, which utilizes the ample information provided by multiple sources to obtain better performance, has attracted much attention. However, existing clustering algorithms either have no ability to offer confidence for each assignment or suffer from the disturbance of outliers. To address these problems, in this paper, we propose a novel multi-view fuzzy clustering method via transferring softmin to fuzzy models. To obtain fuzzy assignments, we utilize the softmin with temperature and further develop an efficient algorithm to solve the non-convex problem approximately. We also show another explanation for the algorithm from the aspect of the prior distribution of various views. Besides, we design a scalable robust loss function, which interpolates between l2-norm and the squared l2-norm, to enhance the robustness to outliers. Extensive experiments show the superiority of our model under different clustering metrics.

In this paper, we solve the problem of feature matching by designing an end-to-end network (called TSSN-Net). Given putative correspondences of feature points in two views, existing deep learning based methods formulate the feature matching problem as a binary classification problem. In these methods, a normalizer plays an important role in the networks. However, they adopt the same normalizer in all normalization layers of the entire networks, which will result in suboptimal performance. To address this problem, we propose a Two-step Sparse Switchable Normalization Block, which involves the advantage of adaptive normalization for different convolution layers from Sparse Switchable Normalization and robust global context information from Context Normalization. Moreover, to capture local information of correspondences, we propose a Multi-Scale Correspondence Grouping algorithm, by defining a multi-scale neighborhood representation, to search for consistent neighbors of each correspondence. Finally, with a series of convolution layers, the end-to-end TSSN-Net is proposed to learn correspondences with heavy outliers for feature matching. Our experimental results have shown that our network achieves the state-of-the-art performance on benchmark datasets.

The Gaussian process is a powerful statistical learning model and has been applied widely in nonlinear regression and classification. However, it fails to model multi-modal data from a non-stationary source since a prior Gaussian process is generally stationary. Based on the idea of the mixture of experts, the mixture of Gaussian processes was established to increase the model flexibility. On the other hand, the Gaussian process is also sensitive to outliers and thus robust Gaussian processes have been suggested to own the heavy-tailed property. In practical applications, the datasets may be multi-modal and contain outliers at the same time. In order to overcome these two difficulties together, we propose a mixture of robust Gaussian processes (MRGP) model and establish a precise hard-cut EM algorithm for learning its parameters. Since the exact solving process is intractable due to the fact that non-Gaussian probability density functions of the noises are adopted into the likelihood of the proposed model on the dataset, we employ a variational bounding method to approximate the marginal likelihood functions so that the hard-cut EM algorithm can be implemented effectively. Moreover, we conduct various experiments on both synthetic and real-world datasets to evaluate and compare our proposed MRGP method with several competitive nonlinear regression methods. The experimental results demonstrate that our MRGP model with the hard-cut EM algorithm is much more effective and robust than the competitive nonlinear regression models.

k-Means is a well-studied clustering problem that finds applications in many fields related to unsupervised learning. It is known that k-means clustering is highly sensitive to the isolated points (called outliers). Such outliers can significantly influence the final cluster configuration and should be removed to obtain quality solutions. In this paper, we study the k-means with outliers problem. Given a set of data points, the problem is to remove a set of outliers such that the k-means clustering cost of the remaining points is minimized. Designing efficient algorithms for this problem remains an active area of research due to its important role in dealing with noisy data. We consider a relaxed objective function and propose a local search algorithm for k-means with outliers. It is shown that the algorithm has better performance guarantees than previously implemented methods. In particular, it yields a constant-factor bi-criteria approximation solution to the problem. Moreover, we show experimentally that the algorithm performs much better than its provable guarantee and dominates other state-of-the-art methods for the problem.

This paper focus on analyzing graph connection of human joints for skeleton based video anomaly detection, which is more effective and efficient than those image-level reconstruction based or prediction based methods that may be affected by complex background. Specifically, we propose a spatial temporal graph convolutional networks based prediction network for skeleton based video anomaly detection. In other words, we build a normal graph describing graph connection of joints in normal data, where joints of abnormal events will be outliers of this graph. To our knowledge, this is the first work to apply graph convolutional networks on skeleton-based video anomaly detection. Experiments show that our proposed normal graph achieves the-state-of-art performance, compared to those image-level reconstruction-based or prediction-based methods, as well as RNN based methods upon joints.

Feature selection methods face new challenges in large-scale classification tasks because massive categories are managed in a hierarchical structure. Hierarchical feature selection can take full advantage of the dependencies among hierarchically structured classes. However, most of the existing hierarchical feature selection methods are not robust for dealing with the inevitable data outliers, resulting in a serious inter-level error propagation problem in the following classification process. In this paper, we propose a robust hierarchical feature selection method with a capped l 2 -norm (HFSCN), which can reduce the adverse effects of data outliers and learn relatively robust and discriminative feature subsets for the hierarchical classification process. Firstly, a large-scale global classification task is split into several small local sub-classification tasks according to the hierarchical class structure and the divide-and-conquer strategy, which makes it easy for feature selection modeling. Secondly, a capped l 2 -norm based loss function is used in the feature selection process of each local sub-classification task to eliminate the data outliers, which can alleviate the negative effects outliers and improve the robustness of the learned feature weighted matrix. Finally, an inter-level relation constraint based on the similarity between the parent and child classes is added to the feature selection model, which can enhance the discriminative ability of the selected feature subset for each sub-classification task with the learned robust feature weighted matrix. Compared with seven traditional and state-of-art hierarchical feature selection methods, the superior performance of HFSCN is verified on 16 real and synthetic datasets.

Low-rank approximation of matrices plays an important role in many application scenarios, including image denoising. This paper introduces a new low-rank approximation method named minimum unbiased risk estimate formulation of 2DPCA (MURE-2DPCA). MURE-2DPCA starts by considering the problem of estimating noise-free matrices from observations, and can exhibit robustness to outliers. In the case of a single data matrix constructed with Gaussian vectors, we find that the optimal dimension of the principal subspace can be determined automatically from the data itself. Based on MURE-2DPCA, a three-step algorithm is developed for color image denoising. Experiments demonstrate the ability and efficiency of our algorithm in achieving the denoising task.

Many subspace learning methods are implemented on a matrix of sample data. For multi-dimensional data, these methods have to convert data samples into vectors in advance, which often destroys the inherent spatial structure of the sample data. In this paper, we propose a robust tensor low-rank sparse representation (TLRSR) method that can directly perform subspace learning on three-dimensional tensors. Firstly, the dual constraints of low-rankness and sparseness make the representation tensor effectively capture the global structure and local structure of sample data, respectively. Secondly, in order to deal with outliers and noise, we adopt the tensor l 2 , 1 -norm to characterize the noise of tensor composed of multiple samples. Thirdly, the denoised tensor instead of the original tensor is used as the dictionary to find the low-rank sparse representation tensor. Finally, an iterative update algorithm is proposed for the optimization of TLRSR, compared with the state-of-the-art methods, clustering on face images and denoising on real images verify the good performance of our proposed TLRSR in tensor subspace learning.

In the past years, most works have focused on designing an indigenous network architecture to advance progress in classification, but another potential opportunity for improvement, that is, research on classification losses, is underdeveloped. Although some new losses have been proposed, most of them either are variants of softmax loss or should combine with softmax loss. Hence, the inherent deficiencies of softmax loss, such as sensitiveness, class-balanced restriction, closed-set limitation, non-scale-invariance and incoordination between the intraclass distance and interclass distance, cannot be completely overcome. In light of this, we pave a new way to design a loss that has no relation to softmax loss and can avoid its weaknesses. We also propose an efficient algorithm to optimize the new loss that can circumvent computing the complicated gradients of a fraction, and the convergence is theoretically ensured. Extensive experimental results on benchmark datasets demonstrate that the new loss is competitive with state-of-the-art losses for classification. Additionally, other specially designed experiments show that the new loss is also effective at handling class-imbalanced problems, is robust in addressing outliers and can discover samples of unseen classes in open-set cases.

The self-expressiveness property of the data, i.e., each sample is linearly represented as a combination of other samples, has recently aroused much attention in the community of data mining and machine learning and shown great promising for subspace clustering. However, real-world data are usually contaminated by noise and outliers. The true similarity between samples directly learned from the original data may deviate from the intrinsic structure of the data, and subsequent clustering results will be severely affected. Hence naively taking a corrupted dictionary, i.e., data itself, will not always obtain the desired clustering performance. To address the above issues, this paper proposes a robust principal component analysis induced self-representation clustering method via adaptively learning the similarity graph from the clean data in the self-representation framework, where the original data are recovered, and the representation coefficients are obtained simultaneously in a unified framework. Specifically, we jointly integrate a nonnegative constraint and a distance regularization into the proposed framework, which guarantees the learned affinity matrix can simultaneously capture the global and local structure of data. Experimental results on both synthetic data and several famous multimedia datasets demonstrate that the proposed method performs much better against state-of-the-arts.

Seeking reliable correspondences between pairwise images is non-trivial in feature matching. In this paper, we propose a novel network, called the Spatial-Channel Self-Attention Network (SCSA-Net), to capture abundant contextual information of correspondences for obtaining reliable correspondences and estimating accurate camera pose of the matching images. In our proposed SCSA-Net, we introduce two types of attention modules, i.e., the spatial attention module and the channel attention module. The two types of attention modules are able to capture complex global context of the feature maps by selectively aggregating mutual information in the spatial dimension and channel dimension, respectively. Meanwhile, we combine the outputs of two modules to generate rich global context and obtain feature maps with strong representative ability. Our SCSA-Net is able to effectively remove outliers, and simultaneously estimate accurate camera pose between pairwise images. These reliable correspondences and camera pose are vital for many computer vision tasks, such as SfM, SLAM and stereo matching. The tremendous experiments on outlier removal and pose estimate tasks have shown the better performance improvements of our SCSA-Net over current state-of-the-art methods on both outdoor and indoor datasets. Especially, our SCSA-Net outperforms the recent state-of-the-art OANet++ by 5.55% mAP5°on unknown outdoor datasets. Code is available at https://github.com/x-gb/SCSA-Net.

Anomaly detection methods differ in their time complexity, sensitivity to data dimensions, and their ability to detect local/global outliers. The recently proposed algorithm FiRE is a ‘sketching’ based linear-time algorithm for identifying global outliers. This work details FiRE.1, an extended implementation of FiRE that fares well on local outliers as well. We provide an extensive comparison with 18 state-of-the-art anomaly detection algorithms on a diverse collection of 1000 annotated datasets. Five different evaluation metrics have been employed. FiRE.1’s performance was particularly remarkable on datasets featuring a large number of local outliers. In the sequel, we propose a new “outlierness” criterion to infer the local or global identity of outliers.

The mean square error (MSE), as the most commonly used cost function for auto-encoder, is sensitive to outliers or impulsive noises in real-world application, which may misguide the training process. At the same time, stacked auto-encoder(SAE) is indeed a totally fully connected network, the parameters exponentially increase as the nodes and layers increase, which may cause over-fitting, huge computational complexity and storage overhead. So the robustness and sparseness problem of the auto-encoder need to be further investigated. In this paper, we develop a robust and structural sparsity stacked auto-encoder with L21-norm loss function and regularization (LR21-SAE). Our L21-norm loss function can alleviate the negative impact of outlier samples, thus show superior robust performance. Our L21-norm regularization can enforce some rows/columns of weight matrix shrink to zero entirely, thus promote to learn sparse features and choose compact network. We have validated our LR21-SAE model on several common datasets. Experimental results show that LR21-SAE is significantly robust to outlier noises for real-world data, it also can get sparse nodes connection deep neural network with notable less number of parameters than what the original non-sparse network has, while maintain outstanding performance.

Short-term traffic flow forecasting at isolated points is a fundamental yet challenging task in many intelligent transportation systems. We present a novel long short-term memory (LSTM) network enhanced by temporal-aware convolutional context (TCC) blocks and a new loss-switch mechanism (LSM) to carry out this task. Compared with conventional recurrent neural networks (RNN) or LSTM networks, the proposed network can capture much more distinguishable temporal features and effectively counteracting noise and outliers for more accurate prediction. The proposed TCC blocks, leveraging dilated convolution, produce an enlarged receptive field in temporal contexts, and formulate a temporal-aware attention mechanism to learn the complicated and subtle temporal features from the traffic flows. We further cascade multiple TCC blocks in the network to learn more temporal features at different scales. To deal with the noise and outliers, we propose a novel loss-switch mechanism (LSM) by combining the traditional mean square error loss and the generalized correntropy induced metric (GCIM), which is capable of effectively counteracting non-Gaussian disturbances. The whole network is trained in an end-to-end manner guided by the loss-switch mechanism. Extensive experiments are conducted on two typical benchmark datasets and the experimental results corroborate the superiority of the proposed model over state-of-the-art methods.

Many semi-supervised learning methods have been developed in recent years, especially graph-based approaches, which have achieved satisfactory performance in the practical applications. There are two points that need to be noticed. Firstly, the quality of the graph directly affects the final classification accuracy. However, graph-based algorithms mostly use k-Nearest Neighbor to construct the graph. And the directly constructed graph is inaccurate due to outliers and erroneous features in the data. Secondly, the amount of labeled data is a small part of all data. It cannot be guaranteed that all categories of data are included in the labeled data and the labels of data are not totally correct in practice. To address the aforementioned problems, we propose a new graph-based semi-supervised method named ALGSSL via adaptive Laplacian graph. In the algorithm, we adaptively update the graph to reduce the sensitiveness of the construction of initial graph. Meanwhile, we use the regularization parameters to set confidence on existing labels, which can reduce the impact of the error labels on the result and discover the new category. Experiments on three toy datasets and nine benchmark datasets demonstrate the proposed method can achieve good performance.

With the rapid emergence of data that can be described by different feature sets or different “views”, multi-view subspace clustering has attracted considerable research attention. To uncover the common latent structure shared by multiple views, existing models usually impose the sparse or/and low-rank constraint on the coefficients of each view data, and use Frobenius norm or l 1 -norm based metric to measure the residuals of multi-view data. However, the intuition behind the sparse or low-rank regularization is implicit. Besides, the Frobenius norm or l 1 -norm based metric is suitable to handle either Gaussian noise or sparse noise, which are very sensitive to larger noise or outliers. When the data is contaminated by large noise or densely corrupted, performance of existing models is degraded dramatically. In this article, we propose a novel multi-view subspace clustering method to provide superior robustness against large noise or outliers embedded in multi-view data. Our method adopts a more direct and intuitive block diagonal regularization to preserve the underlying structure of each view, and meantime introduces the cauchy loss function to deal with large noise. The derived consensus representation matrix can effectively preserve the underlying common structure of multi-view data and be robust to large noise and data corruptions. Experimental results show that our method outperforms the state-of-the-arts on both synthetic and real-world benchmark datasets.

Weakly-supervised multi-label learning has emerged as a hot topic more recently. Most existing methods deal with such problem by learning from the data where the label assignments are incomplete while the feature information is ideal. However, in many real applications, due to the influence of occlusion, illumination and low-resolution, the acquired features are often noisy, which may reduce the robustness of the learning model. In this paper, to overcome the above shortcoming, we propose a novel weakly-supervised multi-label learning framework called WML-LSC, where the low-rank and sparse constrain schemes are jointly incorporated to capture the desired feature information. Specifically, we first decompose the observed feature matrix into an ideal feature matrix and an outlier matrix. Considering that similar instances usually share similar visual characteristics, we constrain the ideal feature matrix to be low-rank. Meanwhile, a reasonable assumption is that the noise is sparse compared with the feature matrix, which leads outlier matrix to be sparse. In addition, a linear self-recovery model is adopted to reconstruct the incomplete label assignment matrix by exploiting label correlations. Finally, the desired model is trained on the ideal feature matrix and the refined label matrix. Extensive experimental results demonstrate that our proposed method can achieve superior and comparable performance against state-of-the-art methods.

Nonnegative matrix factorization (NMF) has attracted more and more attention due to its wide applications in computer vision, information retrieval, and machine learning. In contrast to the original NMF and its variants, this paper proposes a novel unsupervised learning framework, called robust structured nonnegative matrix factorization (RSNMF) which respects both global and local structures of the data space. Specifically, to learn a discriminative representation, RSNMF explores both the global structure via considering the data variance and the local structure via exploiting the data neighborhood. To well address the problem of noise and outliers, it imposes joint L2,1-norm minimization on both the loss function of NMF and the regularization of the basis matrix. The geometric structure and the joint L2,1-norm are formulated as an optimization model, which is solved by the proposed iterative algorithm. Finally, the convergence of RSNMF is analyzed theoretically and empirically. The experimental results on real-world data sets show the effectiveness of our proposed algorithm in comparison to state-of-the-art algorithms.

It is a prevailing trend that the user-generated content (UGC) on social media is shifting toward mobile video and micro content. Sentiment analysis and emotion recognition extract the opinions expressed in UGC and are important to understand the fast-growing mobile micro-videos. Although extensive research efforts have been devoted to it, most of the existing studies pre-processed the video before extracting sentiment features, pre-sampling as an example. Due to the rather unstructured nature of user-generated videos and the sparsely expressed emotions, sampling before sentiment analysis can remove visual contents important to understand the video. In this paper, a novel unsupervised sentiment key frame extraction (SKFE) model based on low-rank and sparse representation is proposed. Sparsity is the important characteristic distinct to a video frame. The low-rank constraint is helpful to improve the robustness to noise and outliers. The l 2 , 1 -norm sparsity is adopted to improve the robustness via different indexes running through the feature dimensions and among the video frames. To better characterize the relationship between the local features and alleviate the sensitiveness of sparse coding, a regularization term based on the Laplacian matrix is introduced to preserve the consistency of sparse codes for similar local features. The experimental results on publicly available datasets demonstrate the effectiveness of the proposed SKFE model.

Low-rank representation (LRR) is a very competitive technique in various real-world applications for its powerful capability in discovering latent structure of noisy or corrupted data set. However, traditional low-rank models treat each data point and feature equally so that noisy data cannot be detected and suppressed effectively and have obvious deterioration in performance, especially in heavy noisy scenario. In this paper, to address this problem, we develop a method of feature and sample scaling for low rank representation. The importance of data points and their features in both feature and sample spaces are considered, as such, clean data points and noisy data points and their features can be distinguished. In addition, based on the observation that noisy data points are usually deviated far away from the principal projection of the data set, a cosine similarity metric between data vector and the principal projection vector is developed to measure the importance of each sample. Applying our method into two classical low rank models such as Low Rank Representation (LRR) and Bilinear Factorization (BF), we can learn better low-rank structure of clean data, while the outliers or missing data being suppressed. Extensive experimental results on ORL, COIL20 and video surveillance, demonstrate that our proposed method can outperform state-of-the-art low rank methods in image clustering tasks with various levels of corruptions, especially in a heavy noisy scenario.

Video summarization shortens a lengthy video into a succinct version, whose challenges mainly originate from the difficulties of discovering the inherent relations between the original video and its summary, meanwhile minimizing the semantic information loss. Supervised approaches, especially those in deep learning framework, have demonstrated their effectiveness in video summarization. However, these approaches mainly focus on one of the challenges, and seldom pay close attention to both challenges simultaneously. To this end, we propose to pay close attention to this deficiency by incorporating the ideas of both the encoder-decoder attention and semantic preserving loss in a deep Seq2Seq framework for video summarization. Moreover, we also introduce Huber loss to replace the popular mean square error loss to enhance the robustness of the model to outliers. Extensive experiments on two benchmark video summarization datasets demonstrate that the proposed approach consistently outperforms the state-of-the-art ones.

Seeking reliable correspondences has been always the key of achieving image matching task in computer vision. The recent work LFGC aims to achieve this goal by training a binary correspondence classifier, which infers the probabilities of correspondences being inliers and regresses the relative camera pose via learning the input correspondence data. Specially, the LFGC employs multi-layer perceptrons to extract features and context normalization to model underlying contextual relation of correspondences. Nevertheless, the contextual relation obtained by context normalization is easily interfered by outliers, which limits the accuracy and robustness of the correspondence classifier. To address this limitation, based on the observation that the contextual relation in spatial neighborhood is generally well preserved on most scenarios, this paper proposes an improved strategy, i.e., maximizing local neighborhood consensus to establish robust neighborhood context. Specifically, we construct the neighborhood for correspondences and preserve neighbors’ features similar to their medoids, where the neighborhood information is embedded. Subsequently, we combine this neighborhood structure with multi-layer perceptrons to classify the correspondences as inliers or outliers, simultaneously potential inliers are employed to regress the relative camera pose encoded by the essential matrix. Qualitative and quantitative experimental results on benchmark datasets demonstrate that the improved network can effectively promote performance in outlier removal and camera pose estimation comparing with the baseline.

In this paper, the H∞ filtering problem is addressed for a class of nonlinear stochastic discrete-time systems subject to Round-Robin protocol, measurement outliers, randomly occurring parameter uncertainties and randomly occurring nonlinearities. First of all, the considered system is influenced by the parameter uncertainties, nonlinear dynamics and external disturbances. Next, mutually independent Bernoulli-distributed white sequences are employed to describe the phenomena of randomly occurring parameter uncertainties and nonlinearities, respectively. Besides, the Round-Robin protocol is introduced to prevent the occurrence of data collisions during the signal transmission. Moreover, a filter with adaptive saturation is constructed to weaken the impact of the measurement outliers. Through constructing a suitable Lyapunov function, sufficient conditions are established to ensure that the error system dynamics is exponentially mean-square stable and satisfies the H∞ performance. Subsequently, the filter gains are designed by solving a set of linear matrix inequalities. Finally, a simulation example is given to show the effectiveness of the filtering scheme proposed in this paper.

In this paper, we propose a precise Iterative Closest Point (ICP) algorithm which is a RGB information supplemented registration method based on correntropy measurement, which can overcome noise and outliers to complete precise point cloud registration. Firstly, in order to achieve accurate point set registration of single structure geometry point cloud with a large number of planes information rather than geometry information, we add the RGB information to our initial method as information compensation, which greatly ensure our algorithm could achieve accurate registration. Secondly, due to the noise and outliers existed in RGB-D data can easily cause deviations in registration, in order to achieve precise registration in RGB-D data, we introduced the correntropy measurement method into the initial model. Experimental results on our scene dataset demonstrate the proposed algorithm can achieve precise point set registration in RGB-D scene data well.

Given a set of 2-dimensional (2D) scattering points, obtained from the edge detection process, the aim of ellipse fitting is to construct an elliptic equation that best fits the scattering points. However, the 2D scattering points may contain some outliers. To address this issue, we devise a robust ellipse fitting approach based on two analog neural network models, Lagrange programming neural network (LPNN) and locally competitive algorithm (LCA). We formulate the fitting task as a nonsmooth constrained optimization problem, in which the objective function is an approximated l0-norm term. As the LPNN model cannot handle non-differentiable functions, we utilize the internal state concept of LCA to avoid the computation of the derivative at non-differentiable points. Simulation results show that the proposed ellipse fitting approach is superior to several state-of-the-art algorithms.

Template matching is an important solution for the object detection in instance-level. Many kinds of matching methods have been utilized to find the template position in the target image. But most of them ignore the matching angle, which causes the result can’t satisfy the detection accuracy when the object has a big in-plane rotation. In this paper, we propose a robust method for template matching with large angle localization. The basic idea is to iteratively search the corresponding patch-features and updating the template location with rotation transformation. The method firstly employs the intensity centroid to rectify local patches as rotation-invariant features. And then, Best-Buddies Pairs (BBPs) are extracted to find corresponding features between template and target images. To further enhance the robustness against outliers, a robust objective function is presented to register features based on the Maximum Correntropy Criterion and optimized for the transformation with translation and rotation parameters. Experimental results demonstrate the effectiveness and robustness of the proposed method.

The robust identification problem of the linear parameter varying (LPV) systems with output data corrupted by outliers is considered in this paper. The local identification approach is used, and the LPV model is obtained by interpolating the local models with an exponential weighting function. In order to handle outliers that could occur in industrial processes, the corresponding probabilistic model is established with the process noise assumed to be mixture Laplace distributed, then the formulas to iteratively update the unknown model parameters and noise-free output are derived under the Variational Bayesian (VB) framework, which approximates the required posteriors and could avoid high dimensional integrals. One numerical example and a practical chemical process are employed to verify the efficacy of the developed algorithm.

This paper reports the usage of the occurrence vector provided by the PAELLA algorithm in the context of robust regression. PAELLA was originally conceived as an outlier detection and data cleaning technique. A novel approach is to use this algorithm not for discarding outliers but to generate information related to the reliability of the observations recorded in the dataset. This approach proves to provide successful results when compared to traditional common practice such as outlier removal. A set of experiments using a contrived difficult artificial dataset are described using both neural networks and classical polynomial fitting. Finally, a successful comparison of our approach to two state-of-the-art algorithms proves the benefits of using the PAELLA algorithm in the context of robust regression.

Robust Principal Component Analysis (RPCA) aiming to recover underlying clean data with low-rank structure from the corrupted data, is a powerful tool in machine learning and data mining. However, in many real-world applications where new data (i.e., out-of-samples) in the testing phase can be unseen in the training procedure, (1) RPCA which is a transductive method can be naturally incapable of handing out-of-samples, and (2) violently applying RPCA into this applications does not explicitly consider the relationships between reconstruction error and low-rank representation. To tackle these problems, in this paper, we propose a Double Robust Principal Component Analysis to deal with the out-of-sample problems, which is termed as DRPCA. More specifically, we integrate a reconstruction error into the criterion function of RPCA. Our proposed model can then benefit from (1) the robustness of principal components to outliers and missing values, (2) the bridge between reconstruction error and low-rank representation, (3) low-rank clean data extraction from new datum by a linear transform. To this end, extensive experiments on several datasets demonstrate its superiority, when comparing with the state-of-the-art models, in several clustering and low-rank recovery tasks.

Riemannian sparse coding methods are attracting increasing interest in many computer vision applications, relying on its non-Euclidean structure. One such recently successful task is image set classification by the aid of Grassmann Manifolds, where an image set can be seen as a point. However, due to irrelevant information and outliers, the probe set may be represented by misleading sets with large sparse coefficients. Meanwhile, it is difficult for a single subspace to cover changes within an image set and the hidden structure among samples is relaxed. In this paper, we propose a novel Grassmann Locality-Aware Group Sparse Coding model (GLGSC) that attempts to preserve locality information and take advantage of the relationship among image sets to capture the inter and intra-set variations simultaneously. Since the contributions of different gallery subspaces to the probe subspace should vary in importance, we then introduce a novel representation adaption term. In addition, a kernelised version of GLGSC is proposed to handle non-linearity in data. To reveal the effectiveness of our algorithm over state-of-the-art, several classification tasks are conducted, including face recognition, object recognition and gesture recognition.

Because of the mechanism of TLS system, noise, outliers, various occlusions, varying cloud densities, etc. inevitably exist in the collection of TLS point clouds. To achieve automatic TLS point cloud registration, many methods, based on the hand-crafted features of keypoints, have been proposed. Despite significant progress, the current methods still face great challenges in accomplishing TLS point cloud registration. In this paper, we propose a multi-scale neural network to learn local shape descriptors for establishing correspondences between pairwise TLS point clouds. To train our model, data augmentation, developed on pairwise semi-synthetic 3D local patches, is to extend our network to be robust to rotation transformation. Then, based on varying local neighborhoods, multi-scale subnetworks are constructed and fused to learn robust local features. Experimental results demonstrate that our proposed method successfully registers two TLS point clouds and outperforms state-of-the-art methods. Besides, our learned descriptors are invariant to translation and tolerant to changes in rotation.

Multi-view clustering (MVC) methods are effective approaches to enhance clustering performance by exploiting complementary information from multiple views. One main disadvantage of most existing MVC methods is that the corresponding optimization problems are non-convex and thus local optimal solutions are usually obtained. To alleviate this issue, we propose a novel multi-view clustering method equipped with self-paced learning, which first learns the MVC model with easy examples and then progressively considers complex ones from each view. In addition, a soft weighting scheme of self-paced learning is designed to further reduce the negative impact from outliers and noises. Furthermore, to consider the importance of different views, we develop an auto-weighted technique to automatically assign weights to views. The proposed model only needs the number of clusters as input and can be easily solved by an alternating optimization paradigm. Experimental results on various benchmark data sets demonstrate the effectiveness of the proposed model.

Person re-identification (re-id) task has attracted a lot of attention because of the excellent performance powered by the Convolutional Neural Network (CNN). However, video-based person re-id is still challenging and far to be solved. On the one hand, the sequence contains complementary information but also more noise information. On the other hand, the training for video is still mainly based on classifying a single frame via its person identity. The representation of the video is generated by aggregating each frame feature. All these will cause that the robustness of video feature is not adequate at the training stage, and the model is easy to be misled by noise information existing in frames. In order to alleviate the difficulty of training video-based re-id, we propose a novel loss named Weighted Triple-Sequence Loss (WTSL) to optimize the video-based feature and reduce the impact of outliers. Further more, we design a Spatial Transformed Partial Network (STPN) coordinated with jointly optimizing image-level and video-level features to generate more robust representation. Extensive experiments show that our algorithm outperforms the state-of-the-art results and achieves 82.2%, 95.2%, and 85.9% rank-1 accuracy on three popular video-based benchmarks: iLIDS-VID, PRID2011, and MARS, respectively.

Subspace clustering (SC) has been widely applied to segment data drawn from multiple subspaces. However, for sequential data, a main challenge in subspace clustering is to exploit temporal information. In this paper, we propose a novel robust sequential subspace clustering approach with a l1-norm temporal graph. The l1-norm temporal graph is designed to encode the temporal information underlying in sequential data. By using the l2 norm, it can enforce well temporal similarity of neighboring frames with a sample-dependent weight, and mitigate the effect of noises and outliers on subspace clustering because large errors mixed in the real data can be suppressed. Under assumption of data self-expression, our clustering model is put forward by further integrating the classical Sparse Subspace Clustering and the l1-norm Temporal Graph (SSC-L1TG). To solve the proposed model, we introduce a new efficient proximity algorithm. At each iteration, the sub-problem is solved by proximal minimization with closed-form solution. In contrast to the alternating direction method of multipliers (ADMM) employed in most existing clustering approaches without convergence guarantee, the proposed SSC-L1TG is guaranteed to converge to the desired optimal solution. Experimental results on both synthetic and real data demonstrate the efficacy of our method and its superior performance over the state-of-the-art methods.

The explosive growth of video data has raised new challenges for many video processing tasks such as video browsing and retrieval, hence, effective and efficient video summarization (VS) is urgently demanded to automatically summarize a video into a succinct version. Recent years have witnessed the advancements of sparse representation based approaches for VS. However, video frames are analyzed individually for keyframe selection in existing methods, which could lead to redundancy among selected keyframes and poor robustness to outlier frames. Due to that adjacent frames are visually similar, candidate keyframes often occur in temporal blocks, in addition to sparse presence. Therefore, in this paper, the block-sparsity of candidate keyframes is taken into consideration, by which the VS problem is formulated as a block sparse dictionary selection model. Moreover, a simultaneous block version of Orthogonal Matching Pursuit (SBOMP) algorithm is designed for model optimization. Two keyframe selection strategies are also explored for each block. Experimental results on two benchmark datasets, namely VSumm and TVSum datasets, demonstrate that the proposed SBOMP based VS method clearly outperforms several state-of-the-art sparse representation based methods in terms of F-score, redundancy among keyframes and robustness to outlier frames.

This research proposes a 1-norm regularized linear programming nonparallel hyperplane support vector machine (LNSVM) model to solve binary classification problems and enhance the robustness performance. Numerous nonparallel support vector machine (SVM) models have been studied with outstanding performance on classification tasks. However, most nonparallel SVM models require two independent models to determine hyperplanes. In addition, due to the involvement of the 2-norm terms, traditional SVM models may suffer from the lack of robustness to outliers and irrelevant features. Therefore, the LNSVM model is proposed by reformulating a typical nonparallel SVM model through the 1-norm regularization. By applying the exterior penalty theory, the proposed LNSVM model is converted to the dual exterior penalty problem, which is solved by the Newton–Armijo algorithm. The essential differences that distinguish the LNSVM model from other nonparallel SVM models are: (1) Different from typical nonparallel SVM models, which solve two quadratic programming (QP) problems, the proposed LNSVM model determines two nonparallel hyperplanes simultaneously by solving a single linear programming (LP) model; (2) The robustness performance of the proposed LNSVM model has been enhanced to tolerate noisy data through the involvement of 1-norm loss function, which can also eliminate redundant features by generating sparse solution during the training procedure. The performance of the proposed LNSVM model is tested through a comparison with state-of-art SVM-based classifiers using a synthetic dataset and 11 practical benchmark datasets. The experimental results show the superiority of the proposed LNSVM model, by achieving better classification performance regarding accuracy, sensitivity, specificity, and removing redundant features synchronously.

Greedy algorithm (GA) is an efficient sparse representation framework with numerous applications in machine learning and computer vision. However, conventional GA methods may fail when applied to grossly corrupted data because they iteratively estimate the sparse signal using least squares regression, which is sensitive to gross corruption and outliers. In this paper, we propose a modal regression based greedy algorithm referred as MROMP (modal regression based orthogonal matching pursuit) to robustly learn the sparse signal from corrupted measurements. Unlike previous GA methods, MROMP is based on sparse modal regression, which has decent robustness to heavy-tailed noise and outliers. To efficiently optimize MROMP, we devise two half-quadratic based algorithms with guaranteed convergence. Our another two contributions are leveraging MROMP to develop a robust subspace clustering method to cluster data lying in a union of subspaces, and a robust pattern classification method to recognize data into the class that they belong to, respectively. The experimental results on both simulated and real datasets demonstrate the efficacy and robustness of MROMP for sparse signal recovery, data clustering and classification, especially for grossly corrupted data.

Owing to the requirements of various product grades or operation conditions, most of industrial processes work with multiple modes. Gaussian mixture regression (GMR) is one of the most widely adopted methods to develop inferential sensors for these processes. However, outliers exist widely due to data points that are incorrectly observed, recorded, or imported and are very hard to be completely recognized and removed. These outliers render the predictive performance of the GMR-based inferential sensors quite disappointing. Aiming at resolving this problem, we propose a robust inferential sensing method based on variational Bayesian Student’s-t mixture regression (VBSMR). In the VBSMR, we first explicitly consider the dependency of quality variables on process variables, where the Bayesian regularization is enabled to find the regression coefficients, and the Student’s-t distributions are introduced for handling outliers. Subsequently, a computationally efficient parameter learning procedure for the VBSMR using the variational Bayesian expectation maximization (VBEM) technique is developed, where the optimal number of mixing components can be automatically determined. Experiments conducted on both numerical and practical industrial examples are provided to demonstrate the availability and flexibility of the developed inferential sensor.

Steganographer detection is to detect culprit users, who attempt to hide confidential information with steganography, among many innocent users. By incorporating the knowledge of true embedding probability map that illustrates the probability distribution of embedding messages in the corresponding image, content-adaptive steganography and steganalysis have made great progress. Unfortunately, true embedding probability map is inappropriate for steganographer detection method due to the significant challenges that the steganographic algorithm and the embedding payload are usually unknown in the task of steganographer detection. In this paper, we propose a novel content-adaptive selective steganographer detection method incorporated with learning-based embedding probability estimation. The embedding probability estimation is first formulated as a pixel-wise segmentation and recognition problem and is integrated into multi-class dilated residual learning model to extract the discriminative features. In the end, the steganographer is identified by local factor outlier with the selective strategy. Extensive experiments demonstrate that the estimated embedding probability map shows robustness against different steganographic algorithms and different payloads. From our experiments, we also find that the proposed content-adaptive selective steganographer detection framework integrated by the estimated embedding probability map achieves low detection error rates in both spatial and frequency domains.

Unsupervised outlier detection for categorical data is important and essential for broad applications in various domains. The complex interactions between attributes and the relevance of attributes make it a stem challenge. Existing methods, including patterns-based and couplings-based methods, either fail to capture the complex interactions or cannot handle the diverse attributes well. In this paper, we propose a novel Neural Probabilistic Outlier Detection method for categorical data, called NPOD. We present a new log-bilinear neural model to learn the categorical distributions, and we observe that the inliers and the outliers can be well separated according to learning loss. Based on this basic observation, we give both empirical and theoretical analysis and present a new neural network architecture that captures the interactions of attributes. Moreover, the discriminative information is used in the proposed bias training process to make the inliers and the outliers more separable. Lastly, to distinguish relevance of attributes, two indicators are proposed for computing ensemble outlier score to get a reliable result. Experimental results show that NPOD significantly outperforms the state-of-the-art competitors on 12 real-world data sets in terms of AUC and P@k.

Many real world data, which we deal with today, often have very high dimensions. These high-dimensional data can be seen as collections of data points from a union of low-dimensional subspaces. Subspace clustering, one solution to the high-dimensional data problem, refers to a method by which a set of data points is divided into multiple clusters by finding multiple subspaces that fit each cluster. Most existing subspace clustering approaches construct an affinity matrix using the self-representation model, which can propagate disturbing noisy information because even noisy data points are used to represent the data; then, they perform spectral clustering on the obtained affinity matrix, which contains the irrelevant information about noise or outliers. This paper proposes a novel subspace clustering method based on the structured sparse PCA-based dictionary learning. Our proposed method learns the reduced dimensional dictionary and coefficient matrices using the structural information as well as sparsity of data. Then, the affinity matrix is constructed from inner product of the learned dictionary coefficient vectors, which shows the correlation among data points. The experimental results on three benchmark datasets verify that the proposed method outperforms state-of-the-art subspace clustering methods.