Development of System for Detecting Railway Surface Defects by Using Deep Learning Technique

Realizing the importance of early prevention of railway deterioration in the railway maintenance, inspection of railway should be carried out regularly. As the railway surface defects can accelerate the deterioration of rail, inspection of railway surface defects becomes a critical task for railway maintenance engineers. Previously, inspection of railway surface defects is conducted by experienced inspectors. However, this subjective visual inspection is labor-intensive, time consuming and costly. So, researchers are finding another alternative to inspect the railway surface defects, and computer vision becomes another alternative to overcome the drawbacks of manual inspection. Previous studies attempted to improve the defect detection of railway surfaces by applying computer vision (CV) based techniques and deep learning; however, there are still limitations. Firstly, the previous research can only conduct the classification of the railway surface defects, but they could not located the defects in the images. Moreover, object detection, which can both classify and locate objects in the image has mainly focused on detecting railway components. Therefore, this paper aims to propose a system that utilizes deep learning techniques to automatically detect railway surface defects within the images. In this research, CNN is trained by using images in which different kinds of railway surface defects include. You Only Look Once (YOLOv4) object detection algorithm is used when the training is conducted. Regarding the collection of image dataset, railway surface defects images are collected from the railway department of Thailand. The system is validated by the testing image dataset, and it can give the result in terms of evaluation matrices. The finding from this research will provide a railway surface defects detection system which can save time and cost and increase the performance.