Industrial Image Anomaly Localization Method based on Reverse Knowledge Distillation

Anomaly localization holds significant importance in the intricate production processes within smart manufacturing systems. Given that acquiring normal data is com paratively straightforward while abnormal data tends to be scarce, this study emphasizes unsupervised anomaly detection techniques centered on reconstruction. It introduces a novel framework that leverages image segmentation and reverse knowledge distillation for the detection and localization of a nomalies in moving objects. This framework is structured in to two distinct phases: component identification and anomaly diagnosis. The component identification phase, utilizing in stance segmentation methods, concentrates on key areas for inspection, ensuring the visual integrity of the components under scrutiny. In the anomaly diagnosis phase, the method of component anomaly diagnosis through reverse knowledge distillation facilitates the unsupervised reconstruction of tem plate images for anomaly diagnosis. This approach adeptly addresses the diversity of component types and their anomalies. An evaluation conducted on a dataset comprising images of scraper transmission chains demonstrates the framework’ s robustness in complex settings and its efficacy in detecting anomalies in transmission components.