Deep learning-based text detection on architectural floor plan images

Abstract

Abstract Architectural drawings are an important source of information for many construction-related tasks, as they contain geometric and semantic information about building parts. However, the manual extraction of room stamps and the insertion of such gleaned information into facility management systems is quite laborious and, thus, its automation is anticipated. In this paper, a method is proposed to detect and classify obscure or illegible text elements on legacy 2D architectural drawings of possibly poor quality. In contrast to existing approaches, a deep learning model is specifically trained for the task at hand rather than making use of transfer learning approaches. The resulting text snippets can be further processed with natural language processing tools to be fed into a facility management system automatically. Other conceivable applications include the extraction of drawing header information, material type or any additional text given on the drawing, to facilitate the enrichment of digital twins of existing structures with semantic data. To provide training data, two floor plan datasets are annotated in a consistent manner. The influence of different data augmentation techniques is investigated systematically. With regard to performance and efficiency, the presented method is compared to alternative tools for the task at hand.