Abstract

This paper focuses on developing and evaluating machine learning methods to automate the extraction and structuring of information from railway signaling technical drawings. The goal is to address the challenge of digitizing paper-based documentation in the railway infrastructure sector. The study addresses this challenge by developing a pipeline that integrates object detection, classification, and optical character recognition (OCR) to identify and extract symbols and text from these technical drawings. The research focuses on two primary areas: symbols and text. For symbols, the project explores how machine learning-based object detection and classification can be designed to identify relay components. This involves a two-step approach: first, detecting the location of symbols using a YOLO (You Only Look Once) model, and second,
classifying the detected symbols using deep learning architectures. The results show a very high performance in symbol detection, with near-perfect recall, and almost perfect accuracy in symbol classification, indicating that these symbols are easily learnable due to their consistent appearance. For text, the study tunes machine learning-based OCR methods to domain-specific data. This includes text detection using a YOLO model to locate text regions, and text
recognition using a fine-tuned transformer-based OCR model to read the text content. While text detection performs strongly, text recognition, particularly the reading of specific character
formats like Roman numerals and subscripts, presents the most errors in the overall pipeline. The results demonstrate that machine learning-based methods can effectively automate parts of
the digitization process for railway signaling drawings. Individual components of the pipeline already perform at a level sufficient for practical use, and the work provides a strong foundation for further development toward a more complete automated system.