Defect Detection for Steel Strip Surface Based on SG-YOLOv5

The steel strip is widely used in the industrial sector as a crucial raw material. However, various forms of damages can occur on the surface of steel strips during production and usage, necessitating the quick and accurate detection of these defects. Current detection methods are faced with challenges such as large model parameters, low detection accuracy, and slow detection speed. To address these challenges, this paper proposes the SG-YOLOv5 algorithm. Firstly, we introduce the SIMAM non-parametric attention mechanism at the output end of the backbone to enhance the feature extraction capability and stability of the detection algorithm without introducing additional parameters. Additionally, we utilize GSconv lightweight convolution kernels and a lightweight module named VGSCSP, which is optimized based on GSconv, to reduce the complexity of the backbone network. The SG-YOLOv5 algorithm significantly reduces the model parameters, accelerates the detection speed, and improves the accuracy of the model. The final model achieves a reduction in parameters by 15.26%, an increase in detection accuracy by 1.8%, and a frame rate of 35.5 frames per second (fps).