Texture Segmentation Using Local Energy in Wavelet Scale Space

Abstract

Wavelet transforms are attracting increasing interest in computer vision because they provide a mathematical tool for multiscale image analysis. In this paper, we show that i) the subsampled wavelet multiresolution representation is translationally variant; and ii) a wavelet transform of a signal generally confounds the phase component of the analysing wavelet associated with that scale and orientation. The importance of this observation is that commonly used features in texture analysis also depend on this phase component. This not only causes unnecessary spatial variation of features at each scale but also makes it more difficult to match features across scales. In this paper, we propose a complete 2D decoupled local energy and phase representation of a wavelet transform. As a texture feature, local energy is not only immune to spatial variations caused by the phase component of the analysing wavelet, but facilitates the analysis of similarity of across scales. The success of the approach is demonstrated by experimental results for aerial Infrared Line Scan (IRLS), satellite, and Brodatz images.