Image Classification Based On Low-rank,Sparse Decomposition And Relationship Between Groups

As the way of automatically understanding, recognizing and managing images, image classification has become an important issue in the field of computer vision. In recently years, image classification based on sparse representation has shown promising prospects. This paper takes the application of sparse representation in image classification as topic, and studies two major parts:dictionary learning and coding process.This paper uses low-rank, sparse matrix decomposition to deal with local features matrix of images, and obtains low-rank part and sparse part. The former represents homogeneousness, correlation for local features, and the latter represents their diversity. The algorithm learns dictionaries on both parts respectively, and then combines two dictionaries as a whole codebook to encode original features. Experiments demonstrate dictionary obtained by this method is more discriminative, and can better represent original features.Traditional sparse coding treats local features separately, ignoring interrelation between them. Group sparse coding makes descriptors in a group have uniform sparse pattern. This paper studies group sparse coding, and find it is important to rationally constitute groups. Only making similar features as a group can reduce difference of intra-class, and increases effect of representations. Based on these thoughts, this paper divide into groups by Ncut, which considers spatial neighborhood of descriptors when clustering, and makes local descriptors as a group as similar as possible.Group sparse coding considers interrelation between local features, which makes descriptors in a group have uniform sparse pattern, however, it still deals with groups separately. This paper further takes into account interrelation between groups, which adds a regularization term into object function of group sparse coding, making similar groups have similar coding, and preserving similarity accordance between descriptors and their codes. The algorithm further reduces difference of intra-class, which enhances feature representations of image, and increases performance of classification.