DBRNet: Dual-Branch Real-Time Segmentation NetWork for Metal Defect Detection

Metal surface defect detection is an important task for quality control in industrial production processes, and the requirements for accuracy, and running speed are becoming increasingly high. However, maintaining the realization of real-time surface defect segmentation remains a challenge due to the complex edge details of metal defects, inter-class similarity, and intra-class differences. For this reason, we propose Dual-branch Real-time Segmentation NetWork (DBRNet) for pixel-level defect classification on metal surfaces. First, we propose the Low-params Feature Enhancement Module (LFEM), which improves the feature extraction capability of the model with fewer parameters and does not significantly reduce the inference speed. Then, to solve the problem of inter-class similarity, we design the Attention Flow-semantic Fusion Module (AFFM) to effectively integrate the high-dimensional semantic information into the low-dimensional detail feature map by generating flow-semantic offset positions and using global attention. Finally, the Deep Connection Pyramid Pooling Module (DCPPM) is proposed to aggregate multi-scale context information to realize the overall perception of the defect. Experiments on NEU-Seg, MT, and Severstal Steel Defect Dataset show that the DBRNet outperforms the other state-of-the-art approaches in balance accuracy, speed, and params. The code is publicly available at https://github.com/fffcompu/DBRNet-Defect.