In order to realize the automatic classification of the cupping test results of crescent edge of strip weld,a classification method based on lightweight network is de-signed.Firstly,data enhancement is used to expand the number of samples in the dataset,then the Grad-CAM algorithm is introduced to visualize the intermediate layer of the test model in the form of heat maps.Finally,a migration learning training method for the freez-ing feature extraction part is designed in conjunction with the visualization of the intermedi-ate layer of the MobileNet V3 network,and the four types of lightweight networks are tested in comparison,the experimental results show that the MobileNet V3 network based on migration learning has a better ability to classify the defects.
维普
万方数据