Training-free counter propagation neural network for pattern recognition of fabric defects

We present an application of a Training-free counter propagation network (TFCPN) to detect fabric defects. The TFCPN, which is a modification of Hecht-Nielsen's counter propagation network (CPN), learns through a simple recording algorithm devoid of any training, while retaining the topology of the CPN model. The mathematical justification for the modification is also presented. Four kinds of fabric defects-neps, broken ends, broken picks, and oil stains-most likely to be found during weaving are considered for recognition by the network. Results show that fabric defects such as these inspected by means of image recognition in accordance with the TFCPN agree approximately with initial expectations. The CPN reported in this paper is training-free, and it can learn complicated textile design problems.