Automated measurement method based on deep learning for cross-sectional SEM images of semiconductor devices

Abstract

Feature length measurement in cross-sectional scanning electron microscopy images of modern semiconductor devices is time-consuming and laborious. We propose an automated measurement method based on deep learning technology and applied it to trench pattern images. The method combines two image-recognition tasks: (1) object detection for determining the coordinates of each unit of a pattern and (2) semantic segmentation for obtaining the boundaries of each area (mask, substrate, and background). By combining the results of these two tasks, typical feature lengths, such as width and depth, are precisely and immediately measured. The extraction speed of the proposed method is 240 times faster than manual measurement and provides measurement results independent of the engineer’s skills.