A Method for Classification of Surface Defect on Metal Workpieces Based on Twin Attention Mechanism Generative Adversarial Network

Abstract

Aiming at the insufficient sample size and unbalanced category of the metal workpiece surface defect dataset in industrial production. But traditional image augmentation methods based on generative adversarial network (GAN) do not effectively correlate the dependency between local feature and global feature, and the loss function is not rational, which causes the problem of insufficient image diversity and poor quality. A relative mean generative adversarial network (TARGAN), driven by a twin attention mechanism, is proposed to generate high-quality defect images. This method employs the twin-attention mechanism that integrates the spatial attention module and the channel attention module. Moreover, it has validated the ability of the twin attention mechanism to improve the refinement of defective features by ablation study. Meanwhile, the loss function during training is improved from the absolute relationship judgment to the relative mean value of the different relationships to improve the diversity of the generated images. Compared with current GANs, this model is superior in generating high-quality and diverse metal workpiece surface defect images thus effectively improving the performance of the MobileNetV2-based classifier. By extended MGSD and NEU-UB based on this method, the accuracy of the defect classification model reached 99% and 98.25% respectively, which is 9.5% and 13.25% higher than the model without TARGAN expansion.