MGF-YOLO: A Lightweight Industrial Inspection Algorithm for Small Defects on Steel Surfaces

To address the issues of varying defect scales and background interference on steel surfaces, which result in low detection accuracy and slow speed in traditional YOLOv8 models, this paper proposes a new MGF -YOLO method for steel surface defect detection. By introducing the MobileNetV3-Small backbone network, the model's parameter count is reduced, enhancing computational speed. A new Dynamic-Ghost Conv (DGConv) is proposed by combining Dynamic Conv and Ghost Conv, improving the neck network's C2f module for more precise localization of fine textures and complex structural defects. Additionally, a weighted bidirectional feature pyramid network (BiFPN) is introduced to enhance the feature fusion network, reducing the probability of defect information loss and improving multi-scale feature recognition. To boost the accuracy and efficiency of detecting complex defects, a more focused Focaler-10 U loss function replaces the CIoU loss function, which improves convergence. Detection results on the NEU-DET dataset indicate that the MGF-YOLO model achieves an accuracy of 85.4%[email protected] and 81.42 FPS, demonstrating a high balance between accuracy and speed. This makes it better suited for defect detection needs in steel strip production.