Diffusion UAD: Unsupervised Diffusion Model Applied to Part Defect Detection

The image-based anomaly detection model is increasingly valued in practical industrial manufacturing due to its remarkable effectiveness and detection efficiency. However, part inspection methods based on traditional generative models are often limited in improving detection efficiency due to constraints on image reconstruction quality. An innovative DiffusionUAD model applicable to part inspection is proposed, comprising a reconstruction subnetwork and a discriminative subnetwork. This model represents the image reconstruction process through a diffusion model as a gradual denoising process from noise to restore normal images, thereby enhancing image reconstruction quality. Consequently, it can reconstruct anomalous regions disturbed by Gaussian noise into normal areas, effectively detecting and locating abnormal parts of components, overcoming the limitations of traditional generative models in image reconstruction quality. Additionally, adding noise based on the foreground improves the accuracy and fidelity of the model. Comprehensive evaluations against six methods on two common datasets demonstrate that DiffusionUAD outperforms state-of-the-art methods, proving its effectiveness and broad applicability.