Identifying wagon numbers using transformers

作     者：Gladys, Bartlomiej Wojakowicz, Bartlomiej Bober, Michal Nalepa, Jakub 

作者机构：Euvic SA Ul Przewozowa 32 PL-44100 GLIWICE Silesia Poland Silesian Tech Univ Akad 16 PL-44100 GLIWICE Silesia Poland 

出 版 物：《ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE》 (Eng Appl Artif Intell)

年 卷 期：2025年第143卷

核心收录：

学科分类：0808[工学-电气工程] 08[工学] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：Silesian University of Technology  Poland 

主　　题：Railway wagon identification Transformer Sequence-to-sequence model Automated identification systems Object detection 

摘      要：Identifying wagons in the railway industry from image data is essential in streamlining the process of wagon inspection and fleet monitoring. However, this is a laborious and time-consuming task, hence automating this process is pivotal in accelerating the identification process and removing human bias. This challenge is addressed in this work, and a multi-stage approach for identifying wagons by detecting and assembling their unique International Union of Railways (UIC) numbers from image data is introduced. Employing a Transformer architecture, the methodology surpasses other methods in accuracy obtained over real-world images and offers fast operation. The model s effectiveness is demonstrated through extensive experiments performed on a unique dataset of more than 3,000 images captured in various weather conditions. The top-performing machine learning model exhibits exceptional speed and real-time performance, processing an average of 100 UIC-containing frames in less than 0.1 son an NVIDIA A10G Tensor Core Graphics Processing Unit, and offering the accuracy of assembling UIC numbers exceeding 95% in scenarios where 1 to 4 digits are missing. The findings presented in this article not only open avenues to apply sequence-to-sequence models to various industrial object detection tasks, but can also significantly improve railway operations.