SIA-net: Structural information awareness network based on normal samples for surface defect detection

Surface defect detection is a challenging task in industrial manufacturing, and the detection method based on deep learning has become the mainstream trend in the industry. However, this method requires a large number of labeled datasets for network training, and it cannot detect new types of defects randomly generated in the actual production process, making it difficult to transfer to actual production. To address this problem, we propose a structural information awareness network (SIA-Net). It is constructed by the adaptive generative adjunctive network (AGAN) module, which can simulate the background style of defect-free samples to randomly embed more reasonable defects. Then, the model is trained to recover the defect embedded area and the defect area is inferred according to the difference between before and after image restoration. Furthermore, in order to avoid blurring the image structure information during feature extraction, the self-attention encoder (SE) and spatial awareness decoder (SD) modules are designed to aggregate the image structure information to generate the final prediction results. We selected four public datasets and specially developed a box defect dataset to verify its detection effect. Experimental results (mIoU/mPA) (Kolektor: 88.91%/89.55%, AITEX defect: 89.61%/91.46%, RSDDs: 86.89%/87.88%, MT defect: 88.36%/90.32%, box defect: 89.71%/90.57%) show that our proposed method clearly outperforms the current unsupervised detection methods.