Self-Attention-Based Adversarial Autoencoder Network for Surface Defect Detection

Surface defect inspection is critically necessary and essential to guarantee high product qualities in the manufacturing industry. In this paper, we propose a self-attention-based adversarial autoencoder (A-3) network that can realize the classification, reconstruction, and segmentation of products for surface defect inspection. To train the A-3 network, defect generation methods in color and deformation are first proposed to construct pseudo defect samples, and then a perceptual loss function combining structural similarity with localization map highlighting the important regions is proposed. Also, an edge consistency loss function is proposed to boost the quality of reconstruction images. Furthermore, two input-to-output modes (defect-free to defect-free and defect to defect-free) are used for training, which result in robust reconstruction images to determine whether the input is defective or not. Extensive experimental results in the MVTec AD dataset proved the efficiency and effectiveness of the proposed A-3 network.