Dual-flow feature enhancement network for robust anomaly detection in stainless steel pipe welding

Runlong Cao, Jianqi Zhang, Yun Shen, Huanhuan Zhou, Peiying Zhou, Guowei Shen, Zhengwen Xia, Ying Zang, Qingshan Liu, Wenjun Hu

Abstract: Anomaly detection in industrial environments, particularly in complex applications like weld anomaly detection in stainless steel pipes, presents substantial challenges. The aim of this paper is to propose a novel approach to address these challenges. We introduce the dual-flow feature enhancement network (DFE), which leverages the dual stream adaptive illumination enhancement module (DSIM) to enhance defective samples, improving feature representation. The enhanced and original images are processed in parallel, and the features from both streams are fused to create a more comprehensive representation. We conduct experiments on two datasets: the widely used MVTec AD dataset and the newly collected stainless steel pipe anomaly detection (SSP AD) dataset. The results show that our method achieves state-of-the-art performance on both datasets, significantly improving the robustness and accuracy of weld anomaly detection. Our approach demonstrates strong potential for practical applications in real-world industrial scenarios. The source code, data, and pretrained models are available at https://github.com/RainloongCao/DFE-master.

External IDs:dblp:journals/vc/CaoZSZZSXZLH25