A vision-based method for lap weld defects monitoring of galvanized steel sheets using convolutional neural network

Abstract

Zn vapour is easily generated on the surface by fusion welding galvanized steel sheet, resulting in the formation of defects. The present study develops a novel method for automatically detect typical undercuts, pores and burn-through defects from CCD using concepts of Convolutional Neural Network(CNN) and active vision. Four deep learning methods of training a defects classifier using 1) an AlexNet built from scratch, 2) a Visual Geometry Group 16 (VGG16) built from scratch, 3) the output features of the VGG16 network architecture previously trained on the general ImageNet dataset, 4) the convolutional block attention module adaptively refining the intermediate feature map of VGG16 are investigated. To detect the weld defects, laser streak images of different weld defects are collected and two types of data augmentation are used to reduce overfitting caused by the limited training dataset. In the method 1 and 2, the improvement of accuracy by data augmentation is gradually slow. VGG16 model in combination with data augmentation 2 achieve the state-of-the-art performance for defects detection in the galvanized steel sheets lap joint. Transfer learning methods using the output features of VGG16 and attention mechanism methods introducing the convolutional block attention module achieve improved accuracy in the case of a small number of training dataset. And all classifiers have a 100 % recognition rate for good welds. This shows that the method of combining CNN and laser vision is feasible, and this method can be used to record the number and location of typical defects in continuous welds.