Automated measurement method with human-pose estimation model for cross-sectional SEM images of semiconductor devices

Extracting feature lengths, such as width, depth, and so on, from cross-sectional scanning electron microscopy (SEM) images is an inevitable task in the process development of semiconductor devices. If this extraction task is done manually, the precision of the result depends on the operator's skill, and this task will be time consuming. We previously proposed a deep-learning-based automated measurement method that combines two image-recognition tasks: (1) semantic segmentation for obtaining the boundaries of each area (mask, substrate, and background) and (2) object detection for determining the coordinates of each unit of a line/space (L/S) pattern. However, it required annotation data consisting of segmented images and bounding boxes, which are not easily made by operators. In this study, we propose a novel measurement method based on a human-pose estimation (HPE) model, which is easier to use.