Image recognition is one of the most popular research problems in the field of computer vision.It becomes an urgent matter that how to extract effective visual representation from the large-scale and high-dimensional visual data and employ it to perform robust and efficient recognition with high accuracy.Generally,the performance of image recognition system is heavily dependent on the choice of data representation.A powerful discriminative data representation can greatly uncover the underlying information of the observed data,and also can clearly improve the performance of the image recognition system.This dissertation is based on the sparse representation and low-rank representation learning theory,and concentrates on extracting meaningful discriminative mid-level information to bridge the connections between low-level observed data and high-level semantic knowledge for effective classification and prediction.Moreover,we exploit the learned representation to discover and analyze the inherent patterns hidden in the data,and then improve the robustness and efficacy of the image recognition models.It is noteworthy that the discriminative representation learning systems should equip with the following three main characteristics:(i)extracting the most compact and prominent invariant features to achieve the best recognition results even for the simplest recognition classifier such as k-NN;(ii)eliminating noise and interference information from data to apply for multiple robust image recognition tasks;(iii)effectively reducing the dimensionality of the visual features to promote the efficiency of image recognition algorithms.This dissertation proposes some novel robust image representation learning models,which are performed on different recognition tasks showing their superior results as well as highly computational efficiency.Specifically,the major research innovations of this dissertation are as follows.To deal with the image preprocessing and recognition problem of axis-symmetrical objects,we propose a robust symmetrical representation learning(RSRL)model based on the axis-symmetrical structure of images.RSRL can automatically produce approximately symmetrical data representation according to the geometrically symmetrical structure of images.RSRL takes the face recognition problem as an example to explore the structurally symmetrical representation learning problem,which is applicable to image preprocessing for the axis-symmetrical object images and virtual dictionary learning for the sparse representation classifier.Specifically,by taking virtue of the axis-symmetrical property of face,RSRL exploits an gradient descent algorithm to iteratively update the left-and right-half face image vectors to generate approximately axis-symmetrical virtual face images,which have the following important merits: 1)RSRL is an effective face image preprocessing method,which can well alleviate the negative effect of heterogeneous illumination variations to increase the visual effects of face images;2)RSRL can automatically produce an approximately axis-symmetrical virtual face images,and aggressively mitigates the intra-class differences from illumination and viewpoint variations to overcome the drawbacks of the sparse representation classifier.As an unsupervised representation learning method,we performed extensive experiments on different face datasets for face image preprocessing and virtual dictionary learning to testify its effectiveness.To overcome the deficiency of the traditional sparse representation learning methods,we propose an effective block-diagonal low-rank representation(BDLRR)method.The elaborate BDLRR is formulated as the discriminative semantic constraints construction problem of shrinking the unfavorable representation from off-block-diagonal elements and strengthening the compact block-diagonal representation under the framework of low-rank representation.Specifically,to boost the incoherent power of the extra-class representation,BDLRR removes the negative representation from the off-block-diagonal components and conveys the positive representation to the block-diagonal structure at the same time,such that better discriminative data representations are obtained to eliminate the representation noise.Moreover,a constructed subspace structure is developed to enhance the coherence of the intra-class representation by simultaneously improving the self-expressive capabilities of training samples and further narrowing the representation gap between training and test samples based on the semi-supervised learning.To accommodate our method for large-scale problems,the out-of-sample extension is further explored how to deal with new data instances.To improve the robustness of the traditional regressive representation learning models,we propose an elastic-net regularized regressive representation learning(ENRRL)framework and a discriminative ENRRL(DENRRL)method.DENRRL incorporates the three-fold characteristics,i.e.compact projection matrix,discriminative regression targets and robust to errors in data,into one unified learning framework,which makes the learned representations discriminative enough for multiple image classification tasks.Specifically,DENRRL for the first time integrates the elastic-net regularization of singular values and distinctive regression targets construction into one unified regressive representation learning framework.The underlying characteristics of the elastic-net regularization of singular values are explicitly uncovered and analyzed.Moreover,by virtue of enlarging the margins of different classes,DENRRL interpolates the ?-dragging technique into the ENRRL framework to enlarge the margins of different classes and make the learned targets distinguishable.Finally,we propose an efficient algorithm to optimize the resulting low-rank minimization problem.To optimize the low flexibility of the existing graph based representation learning models,we propose a discriminative marginalized visual representation learning(MSRL)framework.MSRL jointly incorporates the flexible self-tuning marginal targets analysis,discriminative latent subspace construction and probabilistic geometric structure adaptation into one learning framework,such that the resulting data representations have obvious discriminative capabilities with the near-optimal margins.Specifically,MSRL directly constructs self-tuning regression targets from data with a preferable near-optimal margin constraint.Moreover,a probabilistic graphical structure adaptation is developed to preserve the local similarities and capture accurate underlying structures of data,which in turn guides the construction of marginal regression targets simultaneously.In addition,the regression results are further predicted in the discriminative latent subspace of data,which can capture the underlying correlation patterns.Extensive experiments demonstrate the discrimination and effectiveness of the learned visual representations when solving different recognition tasks.The experimental results demonstrates that MSRL outperforms the state-of-the-art data representation algorithms,and the efficiency of MSRL is further verified by the comparisons of the computational time.In summary,to improve the recognition performance and computation complexity of image recognition problems,we proposed a series of discriminative representation learning models,which are successfully applied for robust multiple image recognition tasks.Extensive experiments were performed on different publicly available image datasets to show their effectiveness and efficiency,and the experimental results demonstrated that improving the discriminability of data representation can clearly promote the robustness and generalization capabilities of image recognition algorithms. |