Automatic Semantic Parsing of Structural Drawings and Rebar Schedules Based on a Hybrid Deep Learning Framework

초록

Reinforcement detailing quality in reinforced concrete structures is directly associated with structural safety and project costs. However, manually cross-checking structural construction drawings against rebar schedules remains inefficient and highly susceptible to human error. Although demand for digital transformation in construction has become increasingly urgent, achieving high-precision automated information extraction continues to be challenging owing to the ultra-high-resolution of engineering drawings and the semantic complexity of domain-specific annotations. To address these issues, this paper proposes a staged visual semantic parsing framework tailored to structural construction drawings and rebar schedules. The proposed approach incorporates an overlapping tiling strategy to preserve local contextual information within large-scale drawings and constructs an end-to-end pipeline that integrates YOLOv11-based text detection with TrOCR-based sequence recognition, enabling accurate dense rebar annotation and tabular data extraction. Experimental results from real-world engineering datasets demonstrate that the proposed method exhibits strong robustness under high-density text conditions and complex background interference, achieving a Character Accuracy of 90.93% (CER = 0.0907) on the held-out test set, outperforming the best baseline by approximately 37.2 percentage points. These findings validate the feasibility and accuracy of leveraging deep-learning techniques for automatic structured information reconstruction from two-dimensional engineering documents, establishing a technical foundation for advancing automated quantity take off, intelligent code-compliance assessment, and digital quality management in the construction industry.

키워드