A high-precision and real-time lightweight detection model for small defects in cold-rolled steel

In cold-rolled steel strip processing, small defects like linear scratches and pits are challenging to detect with existing methods, adversely affecting product quality. To address this issue, we propose SC-YOLOv8, a lightweight and real-time surface defect detection model based on You Only Look Once version 8 nano (YOLOv8n), optimized for identifying small defects on steel strip surfaces. We introduce an adaptive feature extraction module using Dynamic Snake Convolution (DSConv) to enhance detection accuracy for small and elongated targets while maintaining a lightweight model structure. Additionally, we incorporate the Convolutional Block Attention Module (CBAM) to improve feature fusion and reduce information loss, refining feature maps and enhancing the representation of small defects without significantly increasing computational complexity. We utilize the Cold-Rolled Strip Defects Dataset (CR7-DET) which includes seven defect categories and comprises 4140 images with 11,020 annotated object boxes. The proposed SC-YOLOv8 model achieved a mean Average Precision (mAP) of 87.3%, outperforming most current state-of-the-art detection algorithms on CR7-DET. The model has a parameter size of 3.53M and GFLOPs of 8.4. In experimental evaluations, SC-YOLOv8 achieved 398 Frames Per Second (FPS) on the server and 74.7 FPS in laboratory tests simulating an industrial production line, with a total processing delay of 28 ms per frame, fully meeting real-time detection requirements. The lightweight design and high efficiency of this model position it well for adapting to future increases in production speed, aligning with the evolving demands of the steel manufacturing industry.