Weakly Supervised Image Segmentation for Detecting Defects From Scanning Electron Microscopy Images in Semiconductor

The continuous miniaturization of semiconductor patterns improves performance but also leads to frequent pattern disruptions that significantly reduce manufacturing yields. In the field, scanning electron microscope (SEM) images of defects are analyzed to gain insight for improvement. However, manually reviewing a large volume of images to gather additional information is inefficient. Semantic segmentation methods based on artificial intelligence prove effective for defect area extraction, but existing studies largely rely on supervised learning methods that require high-cost ground truth labels. The purpose of this study is to secure accurate segmentation performance without relying on costly labels. To this end, we propose weakly supervised image segmentation for semiconductor defect SEM images, a distinctive approach that uses weak pseudo-labels instead of detailed labels and enhances segmentation performance for defect areas through recursive learning. Pseudo-labels capture the characteristics of SEM images and single pattern bridges to closely approximate the quality of ground truth. These pseudo-labels are used for training the semantic segmentation model and are updated through a post-processing process to be reused as ground truth in the next round, achieving continuous accuracy improvements. The efficacy of the proposed method is demonstrated through single pattern bridge defect images obtained from the latest semiconductor mass production processes. Experimental results show our method significantly reduces the labeling costs, achieving a mean intersection over union (mIOU) of 0.8723. This represents 96% of performance of traditional supervised learning methods. This demonstrates the efficacy of our approach in maintaining high accuracy in defect segmentation with considerably lower resource expenditure.