This study presents an unsupervised anomaly detection approach for identifying foreign objects on conveyor belt systems in industrial environments. The Reverse Distillation for Anomaly Detection (RD4AD) model, trained solely on normal iron ore images,...
This study presents an unsupervised anomaly detection approach for identifying foreign objects on conveyor belt systems in industrial environments. The Reverse Distillation for Anomaly Detection (RD4AD) model, trained solely on normal iron ore images, detects anomalies by measuring feature reconstruction discrepancies between teacher and student networks. The trained model was quantized to INT8 and deployed on the Qualcomm QCS6490 platform to evaluate inference performance. Experimental results show that QCS6490-based heterogeneous computing achieved 95.9% AUROC-Image and 21.50 fps, demonstrating real-time performance on edge devices.