COMPARISON OF SHAPE-FEATURES FOR THE CLASSIFICATION OF WEAR PARTICLES

Wear particle shapes are divided into four classes: Regular; irregular Circular and Elongated. They have been classified here using back-pro pagation neural networks which have been trained using different see o f rotation-, scale-and translation-invariant shape features derived fr om particle boundaries. The features include: Fourier coefficients bas ed on either boundary curvature analysis or XY co-ordinates of boundar y points; statistical moments of the curvature distribution including standard deviation, skewness and kurtosis; and two general shape descr iptions, aspect ratio and roundness. In order to evaluate the performa nces of the features, a series of tests have been carried out on a wea r particle database, and the results are compared. The boundary-curvat ure-based Fourier descriptors produce a shape classifier with the high est performance. The neural network trained by the Fourier features de rived from the boundary data provides a slightly lower classification rate which is similar to that achieved using three statistical moments combined with the two general shape features. (C) 1997 Elsevier Scien ce Ltd. All rights reserved.