A Powerful Object Detection Network for Industrial Anomaly Detection
Yifei Ge, Lin Meng · 2024
The comprehensive task of industrial anomaly detection involves accurately classifying and precisely locating each anomaly, presenting significant practical challenges. Achieving high accuracy in both classification and localization is often difficult in a large-size image, as it requires methods to effectively balance the demands of both aspects. To overcome the challenges, this study proposes a robust object detection network YOLO-RS. In detail, we utilize the improved CSPdarknet as the backbone architecture of YOLO-RS for feature extraction. Meanwhile, this research develops a powerful neck framework of YOLO-RS for feature fusion. Furthermore, the YOLO-RS is implemented on the PCB public dataset to evaluate the proposal. Experimental results show that the YOLO-RS surpasses other state-of-the-art object detection models, achieving 93.44% mAP with only 10.15 M model parameters and reaching a fast inference speed of 73.15 FPS. These demonstrate the proposal.