Paper Title: DefectDiffusion: A Generative Diffusion Model for Robust Data Augmentation in Industrial Defect Detection
Conference Name: International Conference on Artificial Intelligence in Information and Communication (ICAIIC)
Abstract: he accurate detection of industrial defects is critical for ensuring product quality and minimizing operational inefficiencies. However, deep learning models for defect detection often require large, balanced datasets, which are challenging to obtain in industrial settings due to the rarity and variability of defects. In this study, we propose DefectDiffusion, a novel generative diffusion model designed for robust data augmentation in industrial defect detection tasks. By leveraging the progressive noise reduction process inherent to diffusion models, DefectDiffusion synthesizes high-quality, diverse defect images that closely mimic real-world conditions. Unlike traditional augmentation techniques, our approach selectively augments defective regions while preserving the structural integrity of defect-free areas, ensuring realistic and meaningful data augmentation. Experimental results demonstrate that integrating DefectDiffusion-generated images significantly enhances the performance of state-of-the-art defect detection models, improving both precision and recall.