Rail Track Surface Defects Detection Using CNN

Abstract

Railway transportation plays an important role in global transportation systems, providing efficient and rapid movement of passengers and goods. The safety and reliability of train travel heavily depend on the railway tracks' quality. rail surface defects present a substantial risk, which can result in accidents and service interruptions. Identifying rail surface defects presents several challenges, particularly in extracting discriminant features for effective defect detection. This task is complex and non- trivial due to the diverse nature of defects and their appearances. Timely identification is crucial to ensure railway operations' safety and continuous functioning. This paper proposes performing image classification using Convolutional Neural Networks (CNNs) to detect defects on rail surfaces. A publicly available dataset of 1838 images is split 70-20-10 for training, testing, and validation. Rail surface images are preprocessed through resizing, noise reduction, and pixel normalization for CNN compatibility. Data augmentation, including rotation, zooming, brightness adjustment, channel shifting, and horizontal flipping, generates diverse samples with varied perspectives. Convolutional Neural Networks are employed, utilizing transfer learning techniques such as Mobilenet-V2, VGG-16, SqueezeNet, and Inception-V3 to train the classification model with the addition of channel attention. Several pre-trained models are evaluated, and the fine-tuned Inception-V3 model demonstrates a classification accuracy of 95.76%. This research contributes to the development of the railway industry, offering cost-effective solutions for detecting defects in railway tracks in earlier stages