Algorithm for Detecting Defects on Steel Surfaces Based on DMO-YOLO

To address the issue of low detection accuracy in steel surface defect detection tasks, this study proposes an improved YOLOv8n algorithm named DMO-YOLO. Firstly, the algorithm introduces the diverse branching block convolution (DBB) module into the backbone network, which aims to improve the detection effect and at the same time, it can reduce the model parameters. Secondly, some $\mathbf{C 2 f}$ modules in the backbone network have been replaced with C2f_ODConv modules that incorporate full-dimensional dynamic convolution (ODConv), which enhances the model’s focus on multidimensional features and improves the detection accuracy. Thirdly, the traditional SPPF module has been replaced by the Focal-Modulation mechanism, which can focus attention on different spatial scales, thereby better addressing the challenges posed by complex scenes. Finally, the Multi-Scale Dilated Attention (MSDA) is added to the neck network to make the model more flexible to focus on defective regions. Experiments show that the DMO-YOLO algorithm achieves a mAP of $84.3 \%$ on the NEU-DET dataset, Compared with the YOLOv8n algorithm, the [email protected], [email protected]:.95 are improved by 3.7% and $0.5 \%$, respectively.