Texture Classification Method Based On Sparse Representation

Texture is important information in natural images, since it reflects the inherent nature of object’s surface. In the field of texture analysis, describing the texture effectively is one of the key technologies. However, the various categories and contents of textures increase the difficulty of texture description. Sparse representation is a hot spot in the field of signal processing, signal can be simply represented in an effective way by sparse representation. Based on the sparse representation theory, this thesis aimed at introducing sparse representation into texture description. Through building sparse model for texture images, we achieved a simple description for the complicated time domain texture, and applied this model into texture classification successfully. The main research contents and innovation points of this thesis can be summarized as below:1. Based on the sparse representation theory, we built sparse model for texture images. Texture is complicated, and contains lots of high frequency components, thus make it difficult to realize the corresponding sparse decomposition. So, this thesis researched adaptive dictionary learning methods for achieving the best sparse representation of texture images. In global texture description, large computation would be caused in sparse representation. In order to solve this problem, we researched patch-based texture description, which could transform local texture description to global texture description effectively,2. Researched a texton dictionary learning method based on sparse model of texture. Under this kind of dictionary, we could get a one to several reflection relationship between image patch and dictionary atoms through sparse decomposition. Using this relationship, we realized texture classification successfully. In order to reduce computation time, the proposed method using the coherence between image patch and atoms to cut the redundant dictionary. Under the sub-dictionary, sparse decomposition becomes efficiently.3. For the sake of highlighting the class information of texture images in sparse domain, and enhancing the characteristic ability of the dictionary, we researched a discriminative dictionary learning method and applied it into texture classification. In this method, a feature transform matrix was fused into the dictionary learning objective function. Hence, the dictionary and feature transform matrix could be trained at the same time. The trained feature transform matrix could transform the sparse coefficients into discriminative feature space, then classification could be realized.