A new method for multi-texture segmentation using neural networks

A simple method for texture-based segmentation of colored images combining the Hotelling transform and an unsupervised neural network is presented. The method separates the image into n regions representing n possible existing textures. The Hotelling transform is used to improve the color contrast and achieve uncorrelated color components. Auto-contrast is then applied to get a better color spreading. Finally, a set of features extracted from the transformed image is submitted to a competitive neural network for training. Once finished, the neural network parameters are used to label the image regions according to their textures.