Research On Image Recognition Based On Sparse Representation And Low-rank Matrix Recovery

With the development of computer vision and pattern recognition, image recognition technology has made remarkable development in the past decades. It has been widely used in the field of scientific research and industrial production. However, the images acquired in actual are often affected by large noise, occlusion, illumination changes and other conditions, which increase the difficulty of computer recognition. In addition, with the development of information collection technology, the amount of images to be processed by computer are increasing, image dimension is also getting higher and higher, these situations place a burden on computer processing and storage. How to extract effective and compact information from large amounts of data becomes a problem that people need to solve.In recent years, with the development of compressed sensing theory, sparse representation algorithm and low-rank matrix recovery algorithm have gradually become research hotspot because of their strong anti-interference ability, robustness and generalization ability. Sparse representation classifier (SRC) greatly improves the efficiency of image recognition and the robustness to noise due to the use of sparse characteristics, but the classifier often lacks robustness when training images are polluted by noise. In addition, alignment requirements of training images of the classifier is also very high. The emergence of low rank matrix recovery theory offers new ideas to solve the problems which sparse representation classifier is facing. The paper considers using low rank matrix recovery to deal with the training images, remove the influence of noise, at the same time solve the problem of training images unaligned, provide good training dictionary for sparse representation classifier, which is the main problem to be solved in this paper. The paper mainly studies the image recognition method based on the combination of the two algorithms. Through the analysis and comparison of the advantages and disadvantages of some algorithms at home and abroad, some improved methods are proposed. The main work and innovation are as follows.(1) The image recognition algorithms in this paper are proposed mainly against non-ideal conditions, that is, the training samples and test samples are affected by one or more factors, such as illumination changes, partial occlusion, noise pollution and image misalignment etc. Using the low-rank matrix recovery algorithm to decompose the training samples into low-rank matrix and sparse error matrix and then putting effective information into the sparse representation classifier to acquire the recognition result of target image is the main idea of this paper.(2) In the case that the training sample images are contaminated by noise, the low-rank matrix recovery algorithm is used to effectively recover the low rank structure of the training images. Since the image features extracted by Gabor wavelet is more suitable for image representation, the paper introduces Gabor wavelet transform, and constructs a more compact feature dictionary using Gabor features. In this paper, the Gabor feature extraction of low rank structure obtained by low-rank matrix recovery algorithm is carried out. Furtherly, by using the principal component analysis method, the transformation matrix of the Gabor feature of low-rank structure can be acquired. Then the Gabor feature of training and test samples are projected on the same linear subspace through this transformation matrix. Finally, in the linear subspace, sparse coding based on Gabor feature dictionary of the training samples and the test sample of Gabor feature vectors are calculated. Through the simulation experiment on face database, it is proved that the proposed method is feasible, and has good classification ability and strong robustness.(3) SRC classifier has higher alignment requirements for training database, but the images collected in practical application may not only be effected by illumination change, occlusion noise, but also have the problem of image misalignment. In order to solve this problem, the paper proposes a new low-rank matrix recovery model which can not only decompose the sparse error structure and low rank structure effectively, but also align images automatically. The proposed model is simultaneously joined the transformation matrix for aligning image and the Fisher criteria for enhancing the discriminant of the low rank structure. Images are automatically aligned by the RPCA function with transform matrix, which solves the problem that the sparse representation classifier is highly required for the alignment of the image database. The Fisher criterion applied on the low rank matrix reduces the within-class scatter of the low rank matrix, while increasing its between-class scatter. These improvements provide a powerful guarantee for sparse representation classification. Furthermore, the low rank matrix and sparse error matrix are combined together to form the over-complete dictionary. Then the sparse linear combination of test sample on the over-complete dictionary and the class associated reconstruction error of test sample are calculated. The experimental results show that the proposed method has achieved satisfactory results in non-ideal situation.