Unsupervised Industrial Products Anomaly Detection Via Adaptive Complementary Channels and Complete Span Enhancement

Industrial products anomaly detection plays an important role in intelligent manufacturing process. Unsupervised anomaly detection has been widely popular in recent years, but only using normal samples to train the model will lead to over-generalization of the model. Based on this, a novel anomaly detection algorithm with adaptive complementary channels and complete span enhancement (ACCE) is proposed. Based on the differences in product defect rates, the ACCE system constructed a three-level structural anomaly model and a diversified logical anomaly model for data enhancement, thereby significantly improving the matching degree between generated anomalies and real anomalies. Meanwhile, in order to fully explore the potential value of generating abnormal data, ACCE proposed a teacher–student collaborative training framework with maximum diffusion loss, strengthened the discrimination boundary of abnormal features, and achieved more accurate segmentation performance. Furthermore, a secondary de-differentiation adaptive complementary channel selection module was designed on the student network, integrating frequency domain de-differentiation, local-global cascade LSTM (LGC-LSTM), and the adaptive channel weighting mechanism, effectively enhancing the network’s ability to repair defect locations. Additionally, through the dual optimization of the fusion mechanism and attention fine-tuning, the teacher network has a stronger feature adaptability in the anomaly detection task. A large number of experiments on three mainstream datasets MVTec AD, MVTec 3D-AD and VisA demonstrate the validity and generalization ability of the proposed method. In addition, a real-world application of PCB board inspection validates the practicality of the model.