Unsupervised multiresolution texture segmentation using wavelet decomposition

This paper presents an application of 2D continuous wavelet transform (WT) to texture analysis. A complete set of wavelets is defined from an isotropic mother wavelet called the Mexican hat, allowing the decomposition of an image on different scales. Simple orientation texture features are computed on an analyzing window moving through the resulting images, in order to characterize each region of the image by a feature vector. A multiresolution clustering method, based on Kohonen´s (1982) self-organizing map, is then used to group regions into homogeneous areas, according to the similarity of their feature vectors. We show on two examples that the best image segmentation is achieved with a variable size window, depending on the scale of analysis