Research On Image Classification Algorithm Based On Low Rank And Sparse Representation

With the popularity of cameras and surveillance as well as the massive growth of digital images, image classification is becoming an active research topic. As we know, the key technology of image classification is feature extraction and representation. Researchers also pay attention to the data dimension reduction methods to avoid the curse of dimensionality. With the rapid development of Sparse Representation and Low Rank Representation, various kinds of image classification algorithms based on them have been proposed, which show strong vitality. Recently, inspired by discriminant analysis method, transfer learning and graph theory, researchers have great interest in classification algorithms of combining them with Low Rank and Sparse Representation theory. Although the rapid development of image classification technology, it also present several aspects which lead to the difficulty of image recognition. So, there is still a large space for development.The main research of this paper focuses on the image classification based on sparse and low rank representation theory. This paper aims to obtain a more efficient representation of feature information and improve the classification performance. The main contributions of this paper are as follows:(1) This paper gives a comprehensive analysis of research background and significance, sparse and low rank representation theory as well as the application on image classification.(2) This paper introduces some traditional classifiers, the basic theory referring to the image classification methods and several classification algorithms based on sparse and low rank representation in detail.(3) This paper proposes a novel sparse learning algorithm, called locality preserving projection with sparse penalty (SpLPP). SpLPP performs locality preserving projection in the LASSO framework. Combined with sparse and low rank representation, the proposed algorithm merges feature selection and dimensional reduction into one analysis, which indicates that the proposed algorithm can be performed in either supervised and unsupervised tasks. SpLPP aims to keep the local structure and the discrimination ability simultaneously, which improves the performance of image classification efficiently.(4) This paper applies SpLPP to face recognition and do experiments on two kinds of database to test its property of representation and classification respectively. Considering the problem about small sample, SpLPP may not be applied directly. Therefore, this paper presented its regularized version RSpLPP to make the algorithm more applicable.Experiments on synthetic and Frey database show that the proposed algorithms in this paper are competitive compared with the state of the art methods. SpLPP can keep the local structure and the discrimination ability simultaneously which improve its representation ability. Meanwhile, experiments on UCI and USPS database show that SpLPP performs well on image classification.