Semantic Scholar Open Access 2025

2D Engineering Drawing Information Detection Based on Improved RT-DETR

Zhongpeng Li Xinyan Lu He Liu Guanghui Liu

Abstrak

D engineering drawings play an irreplaceable role in mechanical manufacturing area. Engineers need to read such drawings and extract information consistently for cost analysis, quality control, design optimization and some other work. To automate this extracting process while ensuring both efficiency and accuracy, this paper proposes an improved model based on RT-DETR for information detection in 2D drawings. First, a backbone network combining partial convolution (PConv) with efficient multi-scale attention (EMA) is used to reduce model parameters and obtain more detailed image features. Second, deformable attention is used to take the place of multi-head attention in the AIFI of original model to improve recognition accuracy for objects of different sizes. The experiment results show that improved RT-DETR increases precision by $6.43 \%$, increases recall by $14.26 \%$, increase mAP @ 50 by $13.35 \%$ compared to the base model, while requiring less inference time compared to other models. The algorithm demonstrates faster arithmetic speed and higher precision in the real-time detection of 2D engineering drawings.

Penulis (4)

Z

Zhongpeng Li

X

Xinyan Lu

H

He Liu

G

Guanghui Liu

Format Sitasi

Li, Z., Lu, X., Liu, H., Liu, G. (2025). 2D Engineering Drawing Information Detection Based on Improved RT-DETR. https://doi.org/10.1109/IC2ECT66838.2025.11290968

Akses Cepat

Informasi Jurnal
Tahun Terbit
2025
Bahasa
en
Sumber Database
Semantic Scholar
DOI
10.1109/IC2ECT66838.2025.11290968
Akses
Open Access ✓