MSAENet: a multi-scale attention edge network for defect detection in strip steel surface defects

In recent years, surface defect detection technology for strip steel based on salient object detection (SOD) has attracted growing attention from researchers. However, the task remains challenging due to diverse defect types, complex patterns, and low contrast between defects and background surfaces. Existing methods often fail to effectively segment various subtle defects under noise interference and low-contrast conditions. Therefore, we propose a novel multi-scale attention edge network (MSAENet) for strip steel surface defects. Specifically, to explore semantic cues in high-level features, we propose a multi-scale attention module that effectively captures and utilizes multi-scale information. In addition, we design a multi-scale calibration module to enhance extraction of complex low-level edge information through a differential pooling mechanism. Furthermore, we develop a cross-scale interaction module to integrate cross-scale feature representations by exploiting complementary dependencies between network layers. Finally, we design an edge refinement module to further refine the semantic edge response. Extensive experiments on public ESDIs-SOD datasets show that our method outperforms 21 baseline methods. Quantitatively, MSAENet achieves state-of-the-art performance, with an S α of 0.9045, an M of 0.0173, an E ξ max of 0.9731, an F β max of 0.9016, demonstrating significant improvements in defect segmentation accuracy under challenging conditions. The source code is available at https://github.com/CDUTJ102/MSAENet .