Fine-Tuning Convolutional Neural Network Based Railway Damage Detection

Abstract

Due to the rapid development of the railway industry, the task of checking the fit and defects of rails has become of high importance. The train tracks, which are kilometers long, are obtained with hours of video recording. It is almost impossible to examine the images obtained by one or more human eyes. Even if factors that may affect people (such as discomfort, fatigue) are ignored, we can easily state that the time required for the completion of damage assessment will take weeks or months. During the period of investigation, the condition of serious damage may worsen and undesirable results may occur. Therefore, it will save time and cost if the flaws on the rails are made by a deep learning model instead of being made by humans. At the same time, safety in rail transport will be ensured. We propose a high-performance fine-tuning convolutional neural network model that can be improved with negligible losses by using image data to detect defects that occur depending on time or impact on the rail surfaces they use for the transportation of trains. In our study, a two-step approach is followed. In the first stage, we get cropped images focused on the train tracks instead of the rail image captured with a large area. In the second stage, various convolutional neural network models were applied using the cropped images and the classification was provided. While our model continues to work with high success, it works with increasing parameters that accelerate training, such as batch size, and it works very little or even without any loss of success. Experimental results show that our model is better than previous studies.