Wavelet multiscale representation and morphological filtering for texture segmentation

Images are typically composed of structures or features that are scale-invariant in some sense over different spatial scales. The authors first study two scale invariant measures, fractal dimension and a D-dimensional measure both estimated by morphological filtering and used as textural features for segmenting natural textures. They show that they are independent parameters and describe different local characteristics of the irregular intensity surface of a textured image. Further, they demonstrate that the D-dimensional measure gives better clustering in feature space than fractal dimension, at least for distinguishing rural and urban textures. This simplifies segmentation efficiently. However this method suffers from what the authors call an edge effect and they point out that it is difficult to obtain anisotropic information. This leads to segmentation errors. The second part of the paper presents a novel framework which combines wavelet multiresolution decomposition and morphological filtering to effectively reject isolated edges, thus exploiting the assumption that urban texture is isotropic. Examples are given throughout