Finite Element Method-YOLO: Lightweight Steel Strip Surface Defect Detection Algorithm

To improve the efficiency and accuracy of steel surface defect detection while reducing computational complexity, this paper proposes a lightweight detection algorithm—Finite Element Method-YOLO (FEM-YOLO). The algorithm aims to enhance defect detection performance by optimizing the network architecture and minimizing computational overhead. Methodologically, FEM-YOLO is based on the YOLOv8n architecture, incorporating a lightweight Feature Net network as the backbone for feature extraction. Through efficient parameter sharing and feature extraction mechanisms, the algorithm reduces its complexity. Additionally, FEM-YOLO innovatively combines Enhance Conv with the C2f module to form the C2f-Enhance module, thereby improving the representation of fine details and edges within the feature maps. To further enhance detection performance, a multi-path shared convolutional detection head is designed. This design significantly reduces the number of parameters through parameter sharing, thereby improving detection accuracy while maintaining the lightweight nature of the algorithm. The experimental results demonstrate that, on the NEU-DET enhanced dataset, the FEM-YOLO algorithm achieves a parameter count of 1.4 M, which is reduced to 43.7% of the baseline algorithm, with a computational complexity of 4.6 GFLOPs, 52.9% lower than the baseline. Furthermore, the FPS reaches 256. When employing the Focal_EIoU loss function, the mean Average Precision (mAP) reaches 83.3%, validating the algorithm’s efficiency and accuracy in steel surface defect detection.