TCvBsISM: Texture Classification via B-Splines-Based Image Statistical Modeling

This paper presents an image statistical modeling-based texture classification (TC) approach via the Bayesian-driven B-splines probability density estimation of the image textural surface appearance (ITSA), termed TCvBsISM. It approximates the probability density functions (pdfs) of the marginal distribution and joint distribution, involving the global organization and the locally structural layout of local homogeneous patches in the texture surface, respectively, of both the image raw pixel space and the filter response space, by the linear combination of B-spline basis functions (BsBFs) for ITSA feature characterization. The corresponding linear weighting coefficients (LWCs) are determined by an entropy-based optimization criterion with a prior smooth constraint over the LWCs. By leveraging the B-spline-based pdf modeling, distinctive ITSA structural features of texture images are characterized by the LWCs of the pre-defined BsBFs, which are then embedded in an integrated statistical feature dictionary learning, texture pattern representation, and discrimination model to perform TC. Extensive confirmative and comparative experiments on three different texture databases and one natural environmental scene database demonstrate that the proposed TCvBsISM is very promising, especially when the images of different texture patterns appear to be quite similar with the limited training samples. The effects of various parameters on TCvBsISM, such as the choice of the filter bank, the size of the image statistical feature dictionary, as well as the number of BsBFs, are also discussed.