MaskCore: Towards Precise Semantic Segmentation in Industrial Anomaly Detection

In modern manufacturing, a significant challenge in applying deep learning to industry surface defect detection is the scarcity of defect samples and the diversity of abnormal types. This paper introduces MaskCore, a semi-supervised anomaly detection algorithm for industry surface defect detection. MaskCore addresses the current challenges of image feature lacking industrial semantic bias, under-represented local features, and coarse anomaly segmentation. First, self-supervised learning extracts features with industrial image semantic bias through feature adaptor. Second, semantic patch extraction enhances representation at object level. Lastly, the anomaly score is generated based on pixel-level anomaly segmentation via SAM (Segment anything model). We curated an industrial lens dataset, PhoneLens, with smaller surface anomalies surpasses that of available benchmark datasets. Our results show state-of-the-art performance on MVTec AD benchmark dataset, with 97.6% classification AUROC and 97.4% segmentation AUROC. On PhoneLens dataset, MaskCore outperforms the benchmark with 88.3% classification AUROC and 94.7% segmentation AUROC, highlighting its efficacy in industrial anomaly detection.