Improving the quality of convolutional neural networks learning in the tasks of machine vision of unmanned trains

Unmanned trains are a key component of the next level of railway automation. Launching locomotives in unmanned mode requires the development of reliable computer vision systems using artificial intelligence technologies. The paper presents a method for improving the quality of learning convolutional neural networks for detecting railway infrastructure objects. The reliability of visual object detection by a computer vision system can be achieved through algorithmic expansion of training datasets. The proposed method takes into account the variability of weather conditions in which identical objects must be detected, and it allows generating image modifications with added effects of rain, snow or fog. The original dataset included 21700 annotated images and contained 7 classes of objects. Based on them, an extended set of 65100 images was formed using the developed method. To evaluate the effectiveness of the proposed approach, comparative learning of the advanced YOLOv11 model was carried out on the original and extended datasets. The F1-measure and mean average precision (mAP) metrics were used to compare the learning results. The results of the computational experiments confirm that using the extended dataset improves the quality of learning. In particular, the F1-measure for the YOLO model trained on the original dataset was 0.72, while on the extended dataset this parameter reached an increased value of 0.90. The value of the second used metric mAP (50–95) increased from 0.67 on the original dataset to 0.83 on the extended dataset. Comparative values of the metrics were obtained at the same confidence threshold of 0.8. The developed method has been implemented in a hardware and software system, which is ready for testing as part of an integrated control and safety system for freight trains.

1. Fedorov V.A., Ogorodnikova O.M. Convolutional Neural Networks as a Tool for Detecting Railway Infrastructure Objects. Avtomatizatsiya v promychlennosti. 2024;(12):20–23. (In Russ.).

2. Ozerov A.V., Marshova A.S. On Some Aspects of Train Automation in Railway Transport. Nauka i tekhnologii zheleznykh dorog. 2024;8(3):12–17. (In Russ.).

3. Anagnostopoulos A. High-Speed Railway and Safety: Insights from a Bibliometric Approach. High-Speed Railway. 2024;2(3):187–196. https://doi.org/10.1016/j.hspr.2024.08.004

4. Fedorov V.А. Recognizing Railway Infrastructure Using CNN and Stereoscopic Vision. In: 2024 International Russian Smart Industry Conference, 25–29 March 2024, Sochi, Russia. IEEE; 2024. P. 13–18. https://doi.org/10.1109/SmartIndustryCon61328.2024.10516208

5. Tang Yo., Qian Yu. High-Speed Railway Track Components Inspection Framework Based on Yolov8 with High-Performance Model Deployment. High-Speed Railway. 2024;2(1):42–50. https://doi.org/10.1016/j.hspr.2024.02.001

6. Terven Ju., Córdova-Esparza D.-M., Romero-González Ju.-A. A Comprehensive Review of YOLO Architectures in Computer Vision: From YOLOv1 to YOLOv8 and YOLO-NAS. Machine Learning and Knowledge Extraction. 2023;5(4):1680–1716. https://doi.org/10.3390/make5040083

7. Popov V.A., Zubkov A.V. Application of YOLO Family Machine Learning Models for the Task of Analyzing Tea Raw Materials by Photograph. Modeling, Optimization and Information Technology. 2025;13(2). (In Russ.). https://doi.org/10.26102/2310-6018/2025.49.2.042

8. Gavrilov V.S., Korchagin S.A., Dolgov V.I., Andriyanov N.A. Development of a Lightweight Model for Automatic Classification of Structured and Unstructured Data in Streaming Sources to Optimize Optical Character Recognition. Modeling, Optimization and Information Technology. 2025;13(3). (In Russ.). https://doi.org/10.26102/2310-6018/2025.50.3.006

9. Castillo-Girones S., Munera S., Martínez-Sober M., Blasco Jo., Cubero S., Gómez-Sanchis Ju. Artificial Neural Networks in Agriculture, the Core of Artificial Intelligence: What, When, and Why. Computers and Electronics in Agriculture. 2025;230. https://doi.org/10.1016/j.compag.2025.109938

10. Khanam R., Hussain M. YOLOv11: An Overview of the Key Architectural Enhancements. arXiv. URL: https://doi.org/10.48550/arXiv.2410.17725 [Accessed 31st July 2025].

11. He Z., Wang K., Fang T., Su L., Chen R., Fei X. Comprehensive Performance Evaluation of YOLOv11, YOLOv10, YOLOv9, YOLOv8 and YOLOv5 on Object Detection of Power Equipment. arXiv. URL: https://doi.org/10.48550/arXiv.2411.18871 [Accessed 31st July 2025].

12. Yacouby R., Axman D. Probabilistic Extension of Precision, Recall, and F1 Score for More Thorough Evaluation of Classification Models. In: Proceedings of the First Workshop on Evaluation and Comparison of NLP Systems, 20 November 2020, Virtual Event. Association for Computational Linguistics; 2020. P. 79–91. https://doi.org/10.18653/v1/2020.eval4nlp-1.9

13. Chitraningrum N., Banowati L., Herdiana D., et al. Comparison Study of Corn Leaf Disease Detection Based on Deep Learning YOLO-v5 and YOLO-v8. Journal of Engineering and Technological Sciences. 2024;56(1):61–70. https://doi.org/10.5614/j.eng.technol.sci.2024.56.1.5

14. Ogorodnikova O.M., Ali W. Method of Ripe Tomato Detecting for a Harvesting Robot. In: Physics, Technologies and Innovation (PTI-2019): Proceedings of the VI International Young Researchers' Conference, 20–23 May 2019, Ekaterinburg, Russia. American Institute of Physics; 2019. https://doi.org/10.1063/1.5134297

15. Patruno C., Renò V., Nitti M., Mosca N., Di Summa M., Stella E. Vision-Based Omnidirectional Indoor Robots for Autonomous Navigation and Localization in Manufacturing Industry. Heliyon. 2024;10(4). https://doi.org/10.1016/j.heliyon.2024.e26042