Neural network texture classifiers using direct input co-occurrence matrices

A visual comparison between co-occurrence matrices representing nine different natural texture classes is described. Based on these comparisons, matrices of greatly reduced sizes are utilized directly without computing secondary features, to train a feedforward neural classifier to distinguish among the various texture classes. It is shown that recognition rates higher than those obtained using a selected set of conventional features can be obtained. A practical application concerning the classification of wear particles by their surface texture is also discussed. It is shown that high recognition rates can be achieved in distinguishing among samples belonging to five different wear particle classes.<>