HMD2: Texture Surface Defect Recognition With Histogram Matching and Two Stage Denoising Using a Diffusion Model

Texture surface defect recognition plays a key role in ensuring both the aesthetic and functional quality of products, and its importance is further emphasized when considering the impact of defects on product performance and safety. However, existing defect recognition methods predominantly rely on image segmentation techniques, which have limitations in accurately detecting complex texture patterns and subtle defects. To address these challenges, this paper proposes a novel defect recognition method based on a Diffusion model utilizing a U-Net backbone network. The proposed approach introduces a loss function that effectively captures changes in color distribution not only in defect areas but also in normal regions, by incorporating histogram matching. Additionally, the method includes a process that removes high-frequency noise by retaining only low-frequency components. By combining these techniques, the proposed method is designed to significantly improve defect detection accuracy without the need for complex segmentation processes, enabling defect detection at the whole image level. The performance of the proposed method was evaluated using the texture categories of the MVTec-AD, resulting in an average AUC of 0.977 across all texture categories. These results demonstrate the effectiveness of the proposed method in detecting defects on complex texture surfaces by leveraging color distribution and low-pass filtering.