Self-Supervised Fast Texture Defect Detection Based on Salient Object Detection

Texture defect detection is an essential technology in large-scale industrial abnormal detection. Currently, researchers in the defect detection field primarily calculate the anomaly scores of images and perform threshold segmentation to obtain defect detection results. This method generally yields good detection results, but precise boundaries are challenging to obtain due to the similarity of anomaly scores in the edge regions of images. Additionally, there is a growing interest in accelerating defect detection methods to meet real-world application scenarios. To address these challenges, this paper introduces salient object detection methods for texture defect detection. We define the defect regions in defect images as salient regions, effectively capturing the overall image through multiscale learning. This achieves precise localization of defects (salient objects). Moreover, data augmentation is used to artificially synthesize defect images, resolving the salient object detection method’s reliance on strong supervised information and achieving self-supervised defect detection. To better distinguish features between normal and salient regions, we incorporate a contrastive learning module aimed at enhancing feature differentiation. The effectiveness of our proposed method is demonstrated by experimental results on the MVTec texture dataset, achieving a high frame rate exceeding 30 FPS on the Nvidia GeForce GTX 2080ti GPU.