Texton construction using anisotropic diffusion filters

Abstract

A novel and efficient anisotropic diffusion filters method for texton construction is presented in this paper. Recently, texton based texture analysis has been an intensely researched topic especially using Gabor filters output as features for texton construction. The redundancy of Gabor filters makes texton based methods highly time costing. Considering the time cost reducing problem, anisotropic diffusion (AD) filters are introduced in feature extraction for texton construction. Because anisotropic diffusion makes lots of similar features effectively merge together, this method largely reduces feature vectors and time cost on the basis of reliable texton construction. Because of the nonlinearity of nonlinear diffusion, the widely used histogram comparison technique - Earth move distance (EMD) - is not effective for this framework, an appropriate method is proposed instead. Experiments demonstrate the better quality of our method.