An efficient deep learning framework for text detection and recognition in engineering drawings

This is the authors' abstract. We don't add key points for this paper.

The efficient digitization of engineering drawings is critical for automating construction workflows, including quantity takeoff, procurement, fabrication, inspection, BIM modeling, digital twin creation, and lifecycle information management. Manual extraction of steel reinforcement annotations is labor-intensive, error-prone, and impedes digital project delivery. To address this, we propose a deep learning-based text detection and recognition framework that accurately extracts rebar annotations from complex construction drawings. Our detection network uses an enhanced Differentiable Binarization Network (DBNet) with reduced computational complexity (3.8M parameters vs. 12.3M for standard DBNet detection pipeline), representing a 69% reduction in detection parameters. Combined with a Transformer-based recognition model (15.56M parameters), the system achieves high detection (94.2%) and recognition (96.3%) accuracy, even under challenging conditions such as noise, occlusion, and varied styles. By enabling automatic, accurate conversion of technical drawings into structured digital information, our method significantly advances construction automation. This contribution supports the broader goals of reducing manual processes, enhancing information reliability, and accelerating digital transformation in construction projects.