Metal Surface Defect Detect Using Denoising Diffusion Probabilistic Model and Improved YOLOv8

Metal surface defect detection is extremely important in industrial production. However, the limited number of defect samples and imbalanced distribution of categories have always been challenges in this field. In this paper, we propose a new metal surface defect detection method using denoising diffusion probabilistic model (DDPM) and improved YOLOv8 network. Firstly, we adopt DDPM to augment metal surface defect images, enrich the dataset and samples distribution. Secondly, we modify the YOLOv8 network by introducing attention mechanisms and deformable convolution modules. Several contrastive experiments have been carried out on the NEU-DET metal defect dataset to verify the effectiveness of the proposed method. The contrastive experiments have shown that DDPM can augment the data effectively and improve the mean average precision by 3.4%. The attention mechanism can enhance the detection performance for various defect categories, and the deformable convolution can significantly improve the detection results for crazing and rolled-in scale. Simultaneously combining attention mechanisms and deformable convolution further improved crazing and rolled-in scale detection result but lowered other defects. The experimental results of this paper have positive significance for the study of metal surface defect detection.