Exploring Diffusion Model for Semiconductor Defect Detection

The development of X-ray microscopy (XRM) technology has enabled non-destructive inspection of buried semiconductor structures. Combined with deep learning and computer vision techniques, it is possible to automatically identify defects in XRM scans. As samples containing defects are scarce in quantity, the resulting class-imbalance can harm the performance of the deep models. In this work, we propose to employ the recently emerging diffusion model to address this problem. More specifically, we learn the data distribution of XRM scans by a classifier-free diffusion model. The learnt model is then used to generate more defect data to train the deep model. Our experimental results show that the diffusion model is able to generate novel and realistic defect data and is a promising technique to address the data scarcity problem in deep learning-based defect detection for XRM scans.