Despite many studies, texture characterization is still a challenging issue. In this paper, we propose a multiresolution salient point approach based on the intermittency index of a discrete wavelet transform (DWT). Each directional subband (LH, HL, and HH) of the DWT is first transformed into a salient point image (SPI) by using a threshold of the intermittency indices of its wavelet coefficients. We then propose two new texture features, namely, salient point density (SPD) and salient point distribution nonuniformity (SPDN) computed in the neighborhood of every pixel of the SPI. SPD characterizes the coarseness, while SPDN quantifies the distribution of texture primitives. A similar formulation to chi-square statistics is used to compute the SPDN feature. We thus obtain a set of feature images, which are then applied to the well-known K-means algorithm for the unsupervised segmentation of texture images. Experimental results with a standard texture (Brodatz) and natural images demonstrate the robustness and potentiality of the proposed features compared to a wavelet energy (WE) or local extrema density (LED).
展开▼
