Steel Defect Detection Based on Improved YOLOv8n

This paper proposes the 2SC-YOLOv8n object detection model to address issues such as small surface defect sizes in steel, complex backgrounds, and insufficient detection accuracy of existing models. First, the Backbone network of YOLOv8n is improved by integrating a lightweight downsampling module Context Guided Block (CGB) with the original network structure, enabling the module to capture local features, surrounding context, and overall information. Additionally, an attention mechanism Shuffle Attention (SA) is incorporated, which combines shuffling operations with channel and spatial attention mechanisms, effectively enhancing the network’s feature extraction capability. Finally, the attention mechanism SENetV2 is integrated into the Neck network of YOLOv8n to optimize channel and global features through the Squeeze-and-Excitation (SaE) module, thereby improving classification accuracy. Comparative and ablation experiments conducted on the public NEU-DET dataset demonstrate that the 2SC-YOLOv8n model achieves a mean average precision (mAP) of 79.8% at an intersection over union (IoU) threshold of 0.5, representing a 1.3% improvement over the YOLOv8n model. These findings confirm that the 2SC-YOLOv8n model enhances the accuracy of steel surface defect detection and reduces the miss detection rate, making it more suitable for the requirements of real-world industrial production.