Semi-supervised semantic segmentation of SEM images considering multi-scale structural consistency loss in semiconductor pattern layouts

In the fabrication of semiconductor devices, increased yield is achieved using Scanning Electron Microscopes (SEM) to measure and inspect circuit patterns. With recent decreasing scale and increasing complexity of semiconductor circuit patterns, it has become increasingly difficult to recognize patterns accurately using rule-based image processing methods. As such, we propose a method that uses semi-supervised learning for segmentation processing, to recognize which pattern level each pixel represents. With existing methods, the pseudo-labels used for training were not accurate enough, and there were issues such as inconsistent recognition of repeated-pattern layouts and mixed-up results in large unmarked areas distant from the pattern contour. Accordingly, the proposed method is able to perform highly accurate segmentation with the design of new types of loss for evaluating consistency in pattern structure at various scales. When compared with Unimatch and CAC, which are well-known high-performance segmentation methods, the accuracy relative to visual identification increased dramatically, from 10-12% to 100%. In quantitative evaluation using mean Intersection-over-Union (mIoU) at the pixel level, mean values also increased from a range between 0.45 and 0.65 to over 0.94, confirming that the proposed method is effective.