Information Extraction from Railroad Signal Layout Drawings Based on Deep Learning

With the acceleration of intelligent transformation of railroad signaling system, the traditional way of relying on manual interpretation of signaling engineering drawings in the railroad field has been difficult to meet the demand for high efficiency and accuracy in modern engineering. In this paper, for the problems of low efficiency of character information extraction and high false detection rate of station signal layout drawings in engineering drawings, we propose an intelligent parsing framework for station signal layout drawings that integrates target detection algorithms and text recognition technology. First, the data region to be collected is labeled, and then the target detection model is trained. Then, the recognized regions are segmented to prepare for the next text recognition dataset. The framework utilizes YOLOv8 to quickly locate text regions in drawings and combines with PaddleOCR for adaptive enhancement recognition to construct a complete text detection-recognition-structured output solution. The experimental results show that the accuracy of this method in detecting the text target area on the homemade railroad signal drawing dataset reaches 0.998, and the accuracy of PaddleOCR in recognizing the text in the table of turnout types in the target area reaches 0.961, which significantly improves the automation level of the signaling engineering drawings and reduces the cost of manual recognition and the risk of errors.