A comparison study of Gabor and log-Gabor wavelets for texture segmentation

Texture analysis is a significant challenge for computer vision but not yet solved. Since image textures possess spatial continuity at both local and global scales, they are widely used to perform tasks such as object segmentation or surface analysis. Based on the fact that the human visual system (HVS) can segment textures robustly, many texture segmentation schemes use biological models. Modern theories about the functioning of the HVS lead us think that the visual process takes advantage of image redundancy at different scales and orientations, therefore it can be modeled by a bank of Gabor wavelets. Despite the fact that Gabor wavelets optimize the theoretical limit of joint resolution between space and frequency domain, they do not have zero-mean, which induces a DC component in the coefficient of any frequency band. In addition, they do not have a uniform coverage of the frequency domain. These drawbacks may cause errors in the extraction of the appropriate texture features. In this paper, we present a modification of log-Gabor wavelets that allows eliminate DC component. They can yield a fairly uniform coverage of the frequency domain in an octave scale scheme and preserve redundancy at the same time. We analyzed performance of both Gabor and log-Gabor wavelets using a modification of Jain's unsupervised texture segmentation algorithm [1] and we compared results using confusion matrices.