A simple and reliable semi-supervised anomaly detection network for detecting crack in stamped parts

Stamped parts play a crucial role in industrial manufacturing, and it is particularly important to automatically inspect their surface cracks. Since crack is rare and diverse, supervised defect detection methods lack sufficient data and cannot achieve ideal results. Unsupervised anomaly detection algorithms, which do not require crack data, can identify unknown cracks. However, they tend to have high rates of missed detections and false positives when dealing with complex backgrounds in stamped parts. To address these problems, this paper proposes a network called simple and reliable semi-supervised anomaly detection, considering the presence of a small number of anomalous data in actual production. This network uses a large number of normal samples and a small number of anomalous samples to detect surface cracks in stamped parts. Firstly, a pre-trained feature extractor is used for feature extraction, coupled with a designed feature adaptation network to reduce domain bias. Secondly, by extracting normal features from normal images, adding noise to these normal features to generate abnormal features, and extracting abnormal features from abnormal images at multiple scales, a feature space is constructed. Finally, by training a simplified discriminator based on the constructed feature space, computational efficiency is enhanced, and the deployment process is simplified. In the experiments, we collaborated with a multinational company, using an actual production dataset for verification. The proposed algorithm can achieve the score of area under the receiver operating characteristic curve of 98.2% for detection and 97.9% for localization at a processing speed of 19 frames per second.