A Novel Defect Inspection Approach Based on Self-attention Convolutional Adversarial Auto-Encoder

Defect inspection based on computer vision has gradually replaced the traditional detection method, which is widely used in industry and greatly improving the efficiency of industrial production. Most defect inspection methods use supervised learning, which require vast amounts of label images for training. However, the complete defect inspection task is challenging due to some problems, such as diversity of defects, high similarity defects (e.g., low contrast, ambiguous boundary and small background difference) and large defect scale. To solve the above problems, a novel unsupervised defect inspection approach with Self-attention Convolutional Adversarial Auto-Encoder (SCAAE) is proposed in this paper, based on the encoder-decoder-discriminator structure. In this approach, we improve the encoder and decoder via the self-attention mechanism to enhance the feature extraction ability of convolutional autoencoder (CAE), and then the discriminator helps SCAAE impose the latent variable to cluster by a prior distribution. Finally, large number of experiments on four datasets demonstrate the effectiveness of SCAAE and outperforms state-of-the-art methods.