A Surface Defect Detection Method Based on Image Guided Conditional Diffusion Model

Surface defect detection aims to identify local image regions that significantly deviate from the normal appearance. In recent years, surface defect detection using generative models such as GANs and VAEs has shown promising results. However, due to the training instability of GANs and the suboptimal generation performance of VAEs, especially in scenarios with complex backgrounds and structures, they often exhibit coarse reconstruction of normal areas and introduce many false positives. Therefore, this paper proposes a defect detection model based on a classifier-guided conditional diffusion model. The approach for defect detection in this paper involves constructing an image generation network using a conditional diffusion model. The model takes defect images with class labels as conditional inputs to repair the defective regions in the input samples. The input samples and the repaired samples are simultaneously fed into the Improved Local Binary Pattern (ICLBP) network for local image contrast to detect defects and accurately identify the defective regions. During the model training phase, positive samples are inputted, and through a data augmentation module, images are classified and labeled. The defect generation module is then utilized to randomly generate defect images, making the training samples more representative of real-world scenarios. Experimental results on the DAGM 2007 dataset demonstrate that the proposed surface defect detection method achieves superior image reconstruction quality and higher detection accuracy when combined with Improved Local Binary Pattern (ICLBP).