Data-Driven Anomaly Control Detection for Railroad Lines Using Sobel Filter and VGG-16 Model, Res-Net50, InceptionV3

Abstract

Rail inspection is essential to ensure the safety and performance of the rail system. In rail inspection, object detection is an important task to locate and identify damage or obstructions on the rail. This research discusses the use of Sobel features on three convolutional neural network (CNN) architectures, namely VGG-16, ResNet50, and InceptionV3 for object detection in rail inspection. The purpose of this research is to improve the accuracy of object detection in rail inspection by utilizing the edge information obtained from the Sobel filter. This research involves several stages, namely collecting rail image data, image processing with the Sobel filter, feature extraction using three CNN architectures, namely VGG-16, ResNet50, and InceptionV3, and evaluating object detection performance using accuracy metrics. The results show that the use of Sobel features in the three CNN architectures can improve the accuracy of object detection in rail inspection. The evaluation results show that the ResNet50 model provides the best performance with detection accuracy reaching 96%, followed by the InceptionV3 model with 90% accuracy, and the VGG-16 model with 90% accuracy. Based on the results of this study, it can be concluded that the use of Sobel features in CNN architecture can improve object detection accuracy in rail inspection. In addition, the ResNet50 model has the best performance compared to the VGG-16 and InceptionV3 models in object detection in rail inspection. This can be a reference in the development of future rail inspection object detection systems.