WCSE

Intelligent Reading for Multi-scale Engineering Drawings Based on Adaptive Object Detection and Optical Character Recognition

Abstract— Aiming at the problem that a single OCR model cannot extract signs from engineering drawings, and that the multi-size and super-size of input drawings seriously affect the performance of deep learning models, a new intelligent engineering drawing reading scheme is proposed in this paper. Our scheme contains several deep learning detectors, such as YOLO object detector, text angle predictor, DB text detector and CRNN line-level text recognizer. In addition, some important rules and methods are included in the pipeline to unify the input dimensions. We first set up the data set provided by the grid company. For the collected drawings, we carried out format conversion, angle adjustment and size adjustment, and constructed the annotated sample data set of secondary wiring drawings. Then, YOLO V5 model is used to detect the object in the drawing, DB+CRNN is used to detect and recognize the text in the drawing, and knowledge distillation strategy is used to train an ultra-lightweight small model for mobile inferring. The experimental results show that the YOLO V5 object detector achieves 0.98 score under map@0.5 index in this task, and the lightweight OCR model achieves 0.648 score on F1 index when its volume is compressed to 4.7% of the large model volume, which is only 0.031 lower than the large model. In addition, the size transformation rules made by us ensure that the model can keep relatively uniform performance for different sizes of drawings, and prevent the memory explosion caused by large size drawings in industrial applications.

Index Terms— Digitization of Engineering Drawings, Optical Character Recognition, Object Detection, Secure Redundancy Segmentation Strategy.

Rui Yao
School of Computer science, Nanjing University of Posts and Telecommunications, China
Xin Cheng
School of Computer science, Nanjing University of Posts and Telecommunications, China
Zhilei Hui
School of Computer science, Nanjing University of Posts and Telecommunications, China
Zexin Li
School of Computer science, Nanjing University of Posts and Telecommunications, China
Zhaoe Min
School of Computer science, Nanjing University of Posts and Telecommunications, China

[Download]

Cite: Rui Yao, Xin Cheng, Zhilei Hui, Zexin Li, Zhaoe Min, " Intelligent Reading for Multi-scale Engineering Drawings Based on Adaptive Object Detection and Optical Character Recognition, " WCSE 2022 Spring Event: 2022 9th International Conference on Industrial Engineering and Applications, pp. 1484-1492, Sanya, China, April 15-18, 2022.