A Diffusion-Based Approach to Wafer Defect Image Generation in Semiconductor Manufacturing

Accurate defect detection in semiconductor manufacturing is crucial but often hindered by data scarcity and labeling costs. To address this, we develop a latent diffusion model that generates wafer defect images with specified defect types and locations using mask-based control and textual prompts, ensuring real backgrounds simultaneously. By training defect detection models on this augmented dataset with our generated images, we achieved a 15.4% increase in Precision, 16.5% in Recall, and 16.3% in F1-score over baseline methods, validating the effectiveness of our approach.