Research on CNN-based intelligent recognition method for negative images of weld defects

Abstract The existing technology of automatic classification and recognition welding negative images by computer is difficult to achieve a multiple defect while maintaining high accuracy, the developed model image cannot meet actual needs field. Therefore, convolutional neural network (CNN)-based intelligent algorithm for weld defects proposed, B/S (Browser/Server) architecture feature database combined with CNN established subsequently, which converted from migration learning method. It makes full use big data simplifies development process, so that has better generalization ability training accuracy 97.18% achieved after training. results comparison experiments traditional algorithms show CNN-based defective negatives an 92.31% dichotomous defects, significantly than algorithm, effectively improving achieving multi-category recognition. At same time, method was combining segmentation applied in field good results, further verifying feasibility images.