FOD-YOLO: A Steel Surface Defect Detection Model Based on Multi-Dimensional Dynamic Convolution

In complex industrial environments, surface defects of steel are often obscured by highly textured backgrounds, while subtle cracks and low contrast further aggravate difficulty of detection, posing significant challenges to existing automated inspection methods. For this purpose, addressing problems of high miss detection rate and insufficient feature extraction, we propose FOD- YOLO, a detection method for steel surface defects built on YOLOv8n model. By introducing Focus module into backbone network to replace traditional strided convolution layers, lossless downsampling is achieved through space-to-depth transformation, effectively preserving fine- grained feature information and details of subtle defects, and to cope with multi-scale variations in morphology of steel surface defects, we design C2f_OD module, which integrates omni-dimensional dynamic convolution, capable of dynamically adjusting parameters across spatial, channel, and convolutional kernel dimensions of feature responses, thereby enhancing adaptive extraction capability for irregular defects. To strengthen neck architecture, we integrate DySample dynamic upsampling operator, effectively replacing traditional upsampling techniques, and optimize reconstruction process of feature pyramid through content- aware point sampling mechanism, thereby reducing semantic loss during feature fusion process. Comparative experiments conducted on NEU-DET dataset demonstrate that proposed model consistently outperforms YOLOv8n baseline model, achieving 4.0percent improvement on mAP@50 metric and 2.8percent improvement on mAP@50-95 metric, with the miss detection rate also being significantly reduced, highlighting superior reliability of proposed method in defect detection under complex background noise conditions.