A Deep Learning-Based Hybrid Approach to Detect Fastener Defects in Real-Time

Abstract

A fastener is an important component used to fix the rail in railways. Defects in this component cause the rail and ballast to remain unstable. If the defective fasteners are not replaced in time, it is inevitable that the train will derail, and serious accidents will occur. Therefore, this component should be inspected periodically. Conventional image processing-based control systems are affected by noise and different lighting conditions in the real environment. In this study, it is aimed to determine the defects of fasteners with a deep learning-based hybrid approach. The YOLOv4-Tiny method is used for fastener detection and localization. This method gives accurate results, especially for the detection of small objects. After the fastener position is determined, a new lightweight convolutional neural network model is used for defect classification. The proposed convolutional neural network for classification has a small network structure because it uses depth-wise and pointwise convolution layers. When the experimental results are compared with other known transfer learning methods, better results were obtained in terms of training/test time and accuracy.

Keywords

• defect detection
• deep learning
• fastener
• object detection
• railway system