Can defects classification based on improved SOFM neural network

Abstract

In order to realize automatically classifying can defects and improve the convergence speed and the classification accuracy of Self-Organizing Feature Map (SOFM) neural network, 5 improved measures are presented in this paper. They include using typical sample vector, introducing frequency sensitive factor, learning rate adaptive adjustment, selecting convergence criterion and searching winning neuron. Based on the convergence function of SOFM, 6 features of can defects are selected as classification indexes, including the area, anisotropic, roundness, compactness, aspect ratio and the average gray value. The sample data of can defects is classified into 3 categories, which are large edge collapse, pit and dot. The experimental result shows that, compared with classic SOFM method, the classification accuracy of the improved SOFM method is 96%, which is 12% higher than classic SOFM method, and the convergence speed is about 3 times as much as the speed of classic SOFM method. The method presented in this paper has been applied in the actual industrial production.