SMART RAIL VEHICLE MONITORING: A COMPARATIVE STUDY OF CNN ARCHITECTURES

Abstract

This paper presents performance analysis of two deep learning architectures for object detection YOLOv2 and YOLOv4-Tiny aimed at identifying the model state for inspecting the braking systems of Series 441 and 444 locomotives. This models were trained and evaluated using a unified dataset comprising images of 3220 Braking system, as well as instances where these components were absent. Model evaluation was conducted using standard metrics, with a particular focus on the Precision-Recall curve and Average Precision (AP). Experimental results demonstrate that the YOLOv2 architecture achieving a high AP value of 0.97199 after 700 training iterations, accompanied by rapid loss function convergence. In contrast, YOLOv4-Tiny achieved stable convergence after 2500 iterations but yielded a lower AP value of 0.92546. These findings indicate that YOLOv2 can deliver superior performance and higher precision in object detection tasks where fine details are not the primary focus, while simultaneously significantly reducing the demand for complex computational resources.