Context-aware defect sample generation using conditional diffusion model for surface defect inspection

Surface defect inspection is a critical task in manufacturing processes to ensure product quality and operational safety. Despite the successful application of deep learning in various industrial domains, surface defect inspection remains challenging due to the scarcity and irregularity of defective samples. While generative models have been utilized for data augmentation, generating realistic and seamless defect samples has been a challenging task. In this paper, we propose a context-aware defect sample generation method that produces high-quality defect samples while maintaining structural consistency with the background surface. Our approach leverages surrounding contextual information by employing binary masks and applying a conditional reverse diffusion process to iteratively embed realistic defect patterns. Experiments were conducted on five benchmark datasets including GC10-DET, NEU-DET, MT-dataset, TILDA, and MVTec AD-fabric. Experimental results demonstrate that the proposed method effectively improves the performance of surface defect inspection, outperforming existing data augmentation techniques.