Abstract: Industrial defect image samples serve as fundamental data for industrial product defect detection,classification,and grading.To address the current challenges in industrial defect inspection,which include difficulties in target detection under complex environments and insufficient sample quantities resulting in poor feature extraction, a pre-trained autoencoder generative adversarial network was proposed.Pre-trained autoencoder was used to replace the generator network of the basic Generative Adversarial Network (GAN),facilitating better integration of data features by guiding the generator network.An encoder-decoder loss function was redesigned to replace the adversarial loss function of GAN by incorporating target image features.Experimental validation was conducted using a dataset of steel coil end-face defects.Experimental results indicate that after the improved GAN data augmentation,the mean Average Precision mAP_0.5 increased by a maximum of 0.118,while the precision for single-class defect detection increased by a maximum of 0.138.

Key words: Generative Adversarial Network(GAN), industrial image generation, pre-trained autoencoder, defect detection