Classification of Arc Welding Joints Images Based on Convolutional Neural Network

摘要

A vision inspection system for the arc welding joints in the present industrial production is needed for overcoming certain difficulties such as high light reflection and low contrast. The defects of the arc welding joints are complicated and divevisiform, which make them difficult to extract the Region of Interest (ROI) and classify the images by the method of feature values comparison. A classifier based on Convolutional Neural Network (CNN) is built to learn the features in the images and classify them. The CNN architecture and the training parameters are optimized. As a result, the classifier demonstrates an overall accuracy of 95.93% from test sets, which is much better than the accuracy of manual vision inspection.