An evaluation of different network models in machine vision applications

Recently texture segmentation with neural networks has received much interest in fields like remote sensing, medical imaging and autonomous vehicles. We propose to use this approach to improve state-of-the-art machine vision systems. In this paper we present new expermimental data to evaluate the performance of different features as well as different neural network models in a segmentation task. We compare features calculated from second-order statistics (coocurrence features). We investigate different neural network classifers, i. e. multilayerd perceptron and restricted coulomb energy model, as well as different conventional classifiers, i.e. minimum distance and nearest neighbour. A real world example of texture segmentation is the detection of defects on industrial treated surfaces.