Texture image retrieval using compact texton co-occurrence matrix descriptor

Statistical approaches and structural approaches have been extensively investigated in texture analysis of content-based image retrieval whereas little work has integrated them. This paper puts forward an effective texture descriptor-compact texton co-occurrence matrix (CTCM), interweaving the universal local structural textons and statistical co-occurrence matrix seamlessly. First, the CTCM transforms the gray image into a texton index image quickly, utilizing our proposed compact local binary pattern, which generates universal and compact texton dictionary with linear computation complexity. Then the low dimensional texton co-occurrence matrix (TCM) is achieved conveniently. At last, the histogram of TCM is constructed, which represents both the implicit statistical characteristics of individual textons and the explicit statistical characteristics of spatial texton pairs. Experimental results on Brodatz demonstrate that the retrieval performance of CTCM is superior over that of state-of-the-art methods, including the Gabor, local binary pattern and gray-level co-occurrence matrix; furthermore, the CTCM with linear computation complexity is overwhelmingly faster than the Gabor.