Benchmarking Deep Learning Approaches for AEC Engineering Drawing Layout Detection and Information Extraction

Abstract: Information Extraction (IE) from Architecture, Engineering, and Construction (AEC) drawings is hindered by inefficient manual processes. Although layout detection is a key prerequisite for downstream semantic interpretation—organizing graphical and textual hierarchies—it remains under-explored in this domain. Most general document layout analysis models are optimized for text-centric documents. Consequently, their performance on complex engineering drawings remains unvalidated. To address this gap, we construct a custom dataset and benchmark five diverse deep learning architectures on AEC-specific layouts. Results demonstrate that RF-DETR achieves state-of-the-art performance, with an mAP@50 of 0.949, proving the DETR architecture’s superiority in capturing high-density, sparse graphical layouts. Additionally, Vision-Language Models like Qwen3-VL demonstrate strong interpretive ability, achieving an F1-score of 0.911. In contrast, models pre-trained on general document datasets exhibit domain interference, leading to degraded performance. This research confirms that distinct data features prevent effective model transfer from general documents, establishing a robust technical foundation for automated IE.