Huaijiang Sun

Label propagation based on collaborative representation for face recognition

Recently, collaborative representation (CR) has been shown to produce impressive performance on face recognition. However, the performances of CR depend on the number of labeled training samples for each class. When the labeled training samples per class are insufficient, CR would perform inaccurately and correspondingly degrades the final recognition performance. To solve this problem, in this paper, we introduce the CR into semi-supervised learning and propose a novel semi-supervised label propagation approach based on collaborative representation. Based on the subspace assumption that samples of the same class lie in the same subspace, each labeled sample can be well represented by the unlabeled samples of the same class. Our algorithm exploits a large amount of unlabeled samples which contain much more useful information as a dictionary to represent labeled samples, and propagates the label information from labeled data to unlabeled data. Thus, the information of unlabeled data can be effectively explored in our method, which can further improve the performance of collaborative representation with limited labeled training samples. Furthermore, we introduce our label propagation into other semi-supervised learning algorithm to further improve its, recognition performance. Experimental results are presented to demonstrate the efficacy of the proposed method.

Human motion recovery jointly utilizing statistical and kinematic information

Human motion data that are captured by the markers attached to an actors body have been widely used in many areas. However, occlusion caused by the actors body or clothing might make several markers missing for a period of time during the capture process, which highlights the need for motion recovery in the human motion capture process. In recent years, low-rank matrix completion and sparse coding have been used in many data-driven motion recovery methods. However, applying them directly to recover missing data is not effective because low rank is only a basic statistical property of human motion. In addition, the dictionary is usually learned and used in a complete feature space, while human motion must be recovered from an incomplete feature space. Moreover, low-rank matrix completion and sparse coding take advantage only of the statistical property and ignore another important property, i.e., the kinematic property of human motion. Inspired by coupled dictionary learning, we modify the traditional dictionary learning process and propose a new process for the special task of motion recovery. The new recovery process jointly utilizes statistical and kinematic information. Within the proposed method, we first learn a dictionary from a large number of complete-incomplete training frame pairs, to preserve the statistical information of motion data. Then, with the smoothness constraint and the bone-length constraint which take the kinematic information into recovery process, we recover motions using sparse representations of incomplete frames and a learned dictionary through an optimization model. Additionally, we employ two gradient-based optimization algorithms for dictionary learning and motion recovery. Extensive experiment results and comparisons with four other state-of-the-art methods demonstrate the effectiveness of the proposed method.

Multiple Kernel Sparse Representation-Based Orthogonal Discriminative Projection and Its Cost-Sensitive Extension

Sparse representation-based classification (SRC) has been developed and shown great potential for real-world application. Based on SRC, Yang et al. devised an SRC steered discriminative projection (SRC-DP) method. However, as a linear algorithm, SRC-DP cannot handle the data with highly nonlinear distribution. Kernel sparse representation-based classifier (KSRC) is a non-linear extension of SRC and can remedy the drawback of SRC. KSRC requires the use of a predetermined kernel function and selection of the kernel function and its parameters is difficult. Recently, multiple kernel learning for SRC (MKL-SRC) has been proposed to learn a kernel from a set of base kernels. However, MKL-SRC only considers the within-class reconstruction residual while ignoring the between-class relationship, when learning the kernel weights. In this paper, we propose a novel multiple kernel sparse representation-based classifier, and then we use it as a criterion to design a multiple kernel sparse representation-based orthogonal discriminative projection method. The proposed algorithm aims at learning a projection matrix and a corresponding kernel from the given base kernels such that in the low dimension subspace the between-class reconstruction residual is maximized and the within-class reconstruction residual is minimized. Furthermore, to achieve a minimum overall loss by performing recognition in the learned low-dimensional subspace, we introduce cost information into the dimensionality reduction method. The solutions for the proposed method can be efficiently found based on trace ratio optimization method. Extensive experimental results demonstrate the superiority of the proposed algorithm when compared with the state-of-the-art methods.Sparse representation-based classification (SRC) has been developed and shown great potential for real-world application. Based on SRC, Yang et al. devised an SRC steered discriminative projection (SRC-DP) method. However, as a linear algorithm, SRC-DP cannot handle the data with highly nonlinear distribution. Kernel sparse representation-based classifier (KSRC) is a non-linear extension of SRC and can remedy the drawback of SRC. KSRC requires the use of a predetermined kernel function and selection of the kernel function and its parameters is difficult. Recently, multiple kernel learning for SRC (MKL-SRC) has been proposed to learn a kernel from a set of base kernels. However, MKL-SRC only considers the within-class reconstruction residual while ignoring the between-class relationship, when learning the kernel weights. In this paper, we propose a novel multiple kernel sparse representation-based classifier, and then we use it as a criterion to design a multiple kernel sparse representation-based orthogonal discriminative projection method. The proposed algorithm aims at learning a projection matrix and a corresponding kernel from the given base kernels such that in the low dimension subspace the between-class reconstruction residual is maximized and the within-class reconstruction residual is minimized. Furthermore, to achieve a minimum overall loss by performing recognition in the learned low-dimensional subspace, we introduce cost information into the dimensionality reduction method. The solutions for the proposed method can be efficiently found based on trace ratio optimization method. Extensive experimental results demonstrate the superiority of the proposed algorithm when compared with the state-of-the-art methods.

Human Motion Segmentation via Robust Kernel Sparse Subspace Clustering

Studies on human motion have attracted a lot of attentions. Human motion capture data, which much more precisely records human motion than videos do, has been widely used in many areas. Motion segmentation is an indispensable step for many related applications, but current segmentation methods for motion capture data do not effectively model some important characteristics of motion capture data, such as Riemannian manifold structure and containing non-Gaussian noise. In this paper, we convert the segmentation of motion capture data into a temporal subspace clustering problem. Under the framework of sparse subspace clustering, we propose to use the geodesic exponential kernel to model the Riemannian manifold structure, use correntropy to measure the reconstruction error, use the triangle constraint to guarantee temporal continuity in each cluster and use multi-view reconstruction to extract the relations between different joints. Therefore, exploiting some special characteristics of motion capture data, we propose a new segmentation method, which is robust to non-Gaussian noise, since correntropy is a localized similarity measure. We also develop an efficient optimization algorithm based on block coordinate descent method to solve the proposed model. Our optimization algorithm has a linear complexity while sparse subspace clustering is originally a quadratic problem. Extensive experiment results both on simulated noisy data set and real noisy data set demonstrate the advantage of the proposed method.

Keyframe Extraction for Human Motion Capture Data Based on Joint Kernel Sparse Representation

Human motion capture data, which are used to animate animation characters, have been widely used in many areas. To satisfy the high-precision requirement, human motion data are captured with a high frequency (120 frames/s) by a high-precision capture system. However, the high frequency and nonlinear structure make the storage, retrieval, and browsing of motion data challenging problems, which can be solved by keyframe extraction. Current keyframe extraction methods do not properly model two important characteristics of motion data, i.e., sparseness and Riemannian manifold structure. Therefore, we propose a new model called joint kernel sparse representation (SR), which is in marked contrast to all current keyframe extraction methods for motion data and can simultaneously model the sparseness and the Riemannian manifold structure. The proposed model completes the SR in a kernel-induced space with a geodesic exponential kernel, whereas the traditional SR cannot model the nonlinear structure of motion data in the Euclidean space. Meanwhile, because of several important modifications to traditional SR, our model can also exploit the relations between joints and solve two problems, i.e., the unreasonable distribution and redundancy of extracted keyframes, which current methods do not solve. Extensive experiments demonstrate the effectiveness of the proposed method.

Image compressive sensing via Truncated Schatten-p Norm regularization

Low-rank property as a useful image prior has attracted much attention in image processing communities. Recently, a nonlocal low-rank regularization (NLR) approach toward exploiting low-rank property has shown the state-of-the-art performance in Compressive Sensing (CS) image recovery. How to solve the resulting rank regularization problem which is known as an NP-hard problem is critical to the recovery results. NLR takes use of logdet as a smooth nonconvex surrogate function for the rank instead of the convex nuclear norm. However, logdet function cannot well approximate the rank because there exists an irreparable gap between the fixed logdet function and the real rank. In this paper, Truncated Schatten-p Norm regularization, which is used as a surrogate function for the rank to exploit the benefits of both schatten-p norm and truncated nuclear norm, has been proposed toward better exploiting low-rank property in CS image recovery. In addition, we have developed an efficient iterative scheme to solve the resulting nonconvex optimization problem. Experimental results have demonstrated that the proposed algorithm can significantly outperform the existing state-of-the-art image CS methods. Graphical abstractIllustrations of Truncated Schatten-p Norm regularization based CS approach (CS-TSPN). First, obtain an estimate image from sensing matrix and measurements. Second, for each reference patch, group similar patches in its neighborhood. Third, apply TSPN constraints to each group matrix. Then reconstruct the image form these improved group matrices and sensing matrix.ź HighlightsTruncated Schatten-p Norm regularization has been proposed for CS image recovery.ADMM can efficiently solve the resulting complicated optimization problem.CS-TSPN can significantly reduce the required sampling measurements.CS-TSPN can achieve the superior performance compared with other CS methods.

Compressive sensing via nonlocal low-rank tensor regularization

The aim of Compressing sensing (CS) is to acquire an original signal, when it is sampled at a lower rate than Nyquist rate previously. In the framework of CS, the original signal is often assumed to be sparse and correlated in some domain. Recently, nonlocal low-rank regularization (NLR) approach has obtained the-state-of-the-art results in CS recovery which exploits both structured sparsity of similar patches and nonconvexity of rank minimization. However, it still suffers from two problems. First, the NLR approach can not preserve the original geometrical structure of image patches and ignores the relationship between pixels because it deals with the vector form of image patches and the matrix form of patch groups for simplicity. Second, logdet () can not well approximate the rank which is used as a surrogate function for the rank in NLR, because it is a fixed function and the optimization results by this function essentially deviate from the real solution of original minimization problem. In this paper, we propose a nonlocal low-rank tensor regularization (NLRT) approach toward exploiting the original structural information of image patches and structured sparsity of similar patches. We also exploit the use of Schatten p-norm as a nonconvex relaxation for the tensor rank. To further improve the computational efficiency of the proposed algorithm, we have developed a fast implementation utilizing the alternative direction multiplier method technique. Experimental results have demonstrated that the proposed NLRT approach significantly outperforms existing state-of-the-art CS algorithms for image recovery.

Kernel collaborative representation based dictionary learning and discriminative projection

Sparse representation based classification (SRC) has been developed and shown great potential due to its effectiveness in various classification tasks. But how to determine appropriate features that can best work with SRC remains an open question. Based on SRC and dimensionality reduction (DR) techniques, a simultaneous discriminative projection and dictionary learning method (DSRC) is proposed. However, as a linear algorithm, DSRC cannot handle the data with highly nonlinear distribution. Recently research has shown that the collaborative representation mechanism is more important to the success of SRC. Motivated by these concerns, in this paper, we propose a novel kernel collaborative representation based classifier (KCRC), and then we use it as a criterion to design a kernel collaborative representation based dictionary learning and discriminant projection method (KDL-DP). The proposed method aims at learning a projection matrix and a dictionary such that in the low dimension subspace the between-class reconstruction residual of a given data set is maximized and the within-class reconstruction residual is minimized. Extensive experimental results validate the superiority of the proposed approach when compared with the state-of-the-art methods.

Kernel dictionary learning based discriminant analysis

A feature learning and kernel dictionary learning method is proposed.Adopt the joint learning technique to learning the projection and dictionary simultaneously.The learned features which can fit the dictionary learning well.The sparse coding of the data can be easily obtained, and the reconstruction error can be reduced.The proposed method can achieve better performances in the projected space. Sparse representation based classification (SRC) has been successfully applied in many applications. But how to determine appropriate features that can best work with SRC remains an open question. Dictionary learning (DL) has played an import role in the success of sparse representation, while SRC treats the entire training set as a structured dictionary. In addition, as a linear algorithm, SRC cannot handle the data with highly nonlinear distribution. Motivated by these concerns, in this paper, we propose a novel feature learning method (termed kernel dictionary learning based discriminant analysis, KDL-DA). The proposed algorithm aims at learning a projection matrix and a kernel dictionary simultaneously such that in the reduced space the sparse representation of the data can be easily obtained, and the reconstruction residual can be further reduced. Thus, KDL-DA can achieve better performances in the projected space. Extensive experimental results show that our method outperforms many state-of-the-art methods.

Blind compressive sensing using block sparsity and nonlocal low-rank priors

Abstract Without knowing the sparsity basis, Blind Compressive Sensing (BCS) can achieve similar results with those Compressive Sensing (CS) methods which rely on prior knowledge of the sparsity basis. However, BCS still suffers from two problems. First, compared with block-based sparsity, the global image sparsity ignores the local image features and BCS approaches based on it cannot obtain the competitive results. Second, since BCS only exploits the weaker sparsity prior than CS, the sampling rate required by BCS is still very high in practice. In this paper, we firstly propose a novel blind compressive sensing method based on block sparsity and nonlocal low-rank priors (BCS-BSNLR) to further reduce the sampling rate. In addition, we take alternating direction method of multipliers to solve the resulting optimization problem. Experimental results have demonstrated that the proposed algorithm can significantly reduce the sampling rate without sacrificing the quality of the reconstructed image.