Masking-Based Distinguishing Learning of Marginally Qualified Data for Memory Bank Defect Detection

In industrial scenarios, good parts produced often don't fully comply with standard specifications. Some samples might have abnormal conditions like water stains and overexposure. But since this deviation is within the predefined qualified range, they're still considered qualified. We call these samples marginally qualified samples. One specific challenge in our work is the model's accurate identification ability for marginally qualified samples: generating embeddings and doing outlier detection using only nominal (non-defective) example images. However, in actual industrial scenarios, many marginally qualified samples differ from the example images. Their inconsistent features will affect the final detection accuracy. Traditional anomaly detection methods would judge marginally qualified samples as anomalies. In this paper, we've expanded on embedding-based anomaly detection and proposed MDL (Masking-based Distinguishing Learning of Marginally Qualified Data). It models and sorts the nominal block features into the real non-defective feature memory bank and the marginally qualified sample feature memory bank to accurately judge marginally qualified samples as normal. Additionally, we've introduced marginally qualified data supplementation. This can accurately generate a large number of marginally qualified samples and nominal samples to simulate the distribution of real industrial normal samples. MDL can be integrated into any embedding-based anomaly detection framework and has achieved advanced performance in both anti-interference detection and localization. When using MDL, an image-level anomaly detection AUROC score of up to 96.93% has been achieved on the widely used MVTec AD benchmark synthetic images. We've further demonstrated the effectiveness of MDL, even for marginally qualified samples.