MS-YOLOv5s: An Improved YOLOv5s for the Detection of Imperceptible Defects on Steel Surfaces

The detection of surface defects in steel is crucial for maintaining high product quality standards and preventing financial losses due to inferior goods. However, the subtle imperfections that escape human observation present a significant challenge to existing algorithms. In response to this pressing need, our research introduces an advanced approach, MS-YOLOv5s, designed to better distinguish between background and defects. Firstly, our study presents a new neck module, the comprehensive position feature pyramid networks (CPFPN), which improves the precision of detecting barely noticeable flaws. This is achieved by using the spatial channel attention module (SCAM) on intermediate feature maps to retain more positional information from the original image. Moreover, this innovative method adopts multi-scale learning, dynamically adjusting the input image size during training to amplify the differences between defects and background. MS-YOLOv5s achieves 80.5% and 65.7% mean average precision (mAP) respectively on the NEU-DET and GC10-DET datasets, demonstrating robust performance across various scenarios and outperforming many methods in identifying defects on the steel surface.