Abstract: This paper treats texture image as a linear superposition of a group of fundamental micro-patterns. The topographic independent component analysis is employed first to learn the image bases, and those mixing bases are then transformed into micro-patterns, according to the generative model. Two examples of applications of image bases are illustrated finally. One is to extract the fundamental visual elements from micro-patterns, and the other is texture segmentation by using the learned unmixing bases as filters. Experiments on mosaic texture images show that the segmenting results are satisfactory.