beam-internal-structural-details-extraction-dataset

This record provides the dataset associated with a multimodal drawing information parsing framework for scanned beam structural construction drawings. The dataset supports three tasks:(1) object detection for locating structural drawing elements and annotation-related regions,(2) optical character recognition (OCR) for recognizing beam annotation texts in scanned drawings, and(3) named entity recognition (NER) for extracting structured engineering information from recognized texts. The dataset was developed for research on automated extraction of refined structural modeling parameters from existing-building drawings in the context of urban renewal. It supports a workflow that integrates object detection, OCR, and NER to parse scanned beam structural construction drawings and facilitate the automatic generation of refined beam structural models. According to the associated study, the proposed two-stage object detection scheme effectively addresses small-object detection in scanned drawings, with both stage-1 and stage-2 models achieving mAP50 above 0.95. The OCR model trained on the augmented dataset achieved 99% recognition accuracy on beam annotation texts in real scanned drawings. In addition, the deep learning-based NER model achieved over 90% recognition accuracy across all defined entities in beam annotation text parsing. This record includes datasets and related files for the YOLO-based object detection task, the OCR task, and the NER task. These materials are intended to support reproducibility and further research on structural drawing digitization, information extraction, and automated structural modeling. If some original engineering drawings are subject to copyright, institutional, or project-related restrictions, this record provides the shareable processed data, annotations, and supporting files that can be made publicly available.