Steel surface small defect target detection based on improved YOLOv8

Abstract To address the issues of low precision and poor performance in small object detection with existing deep learning-based steel surface defect detection algorithms, an improved YOLOv8 network-based detection algorithm is proposed. The network incorporates a designed iRMB module that enhances multi-scale feature fusion through multi-layer feature extraction and residual structures, thereby improving the model’s ability to recognize small defects on steel surfaces. Additionally, the CReToNext module is integrated to replace the C2f module in the feature fusion layer, allowing for better handling of various types and complexities of defects. The SlideLoss function is utilized as the classification loss function, enhancing the detection capability for challenging targets. To validate the feasibility of the algorithm, multiple improved algorithms were compared, and ablation experiments were conducted to explore the effectiveness of each improved module. The experimental results show that the improved algorithm achieves 95.7% mean average precision (mAP) and 93.9% precision on the open NEU-DET dataset, which is better than the most advanced detection algorithm on NEU-DET and 6.3% and 6.2% higher than the baseline YOLOv8.