부산대학교 도서관

Computer-aided design (CAD) drawing parsing is a fundamental step to both drawing design expansion and digital management in industrial field. Existing methods can only parse some areas of the complete drawing. In this paper, we propose the task of panoptic CAD drawing parsing, which requires locating all area symbols, composition symbols and their relations in the complete drawing. This task is challenged by locating tiny symbols in high-resolution images and the pixel-wise locating the composition symbols. Therefore, we propose a hierarchical panoptic CAD drawing parsing system based on deep learning. Concretely, we use a hierarchical approach to firstly parse the drawing layout regions and then locate tiny area symbols based on symbol location. Meanwhile, for the composition symbols, we propose the Symbol-as-Points (SaPs) algorithm to obtain pixelwise location. 800 industrial drawings are precisely annotated to verify the effectiveness of our approach. Extensive experimental results prove the effectiveness of the framework.