PUB - Publikationen an der Universität Bielefeld

Texture discrimination plays an important role in image segmentation tasks. We propose a model of visual texture perception which involves a biological motivated feature extraction mechanism and the application of a neuro-dynamical model, the Competitive Layer Model (CLM). After giving an introduction to the "Gestalt"-approach of perceptual grouping, we will describe the architecture of the fully connected recurrent CLM and propose an algorithm to efficiently simulate the dynamics of the system. Based upoin Gabor filters we present a detailed description of how to gain a reliable representation of texture features from images. Further stages of processing include an early nonlinearity and the spatial pooling of the texture features. A single interaction function between these features which involves the two principles of similarity and proximity then allows the usage of the CLM for image segmentation. We show that our model of texture perception can be applied to a remarkable range of images known from pychophysical studies and that the results are in good consistence with human introspection.