Improving Important Object Detection Accuracy Based on a Two-Stage YOLOX and EfficientNetB2 Approaches

Abstract

In application fields requiring high safety standards, such as autonomous driving and railway systems, the accurate detection of critical objects is essential for system reliability. This study focuses on improving the detection accuracy of railway switches, key components that control train direction, using the RailSem19 dataset. Switches are visually similar to other structures and backgrounds in railway environments, making their detection inherently challenging. We propose a two-stage method that combines YOLOX, a state-of-the-art object detection model, with Efficient-NetB2 for fine-grained classification. YOLOX detects candidate regions for switches, which are refined by Efficient-NetB2 to classify them into three sub-categories (switchright, switch-left, and switch-unknown) or background. Experiments on the RailSem19 dataset showed that the proposed methods improved detection accuracy (AP0.5:0.95) by approximately 5 points for switch-right and switch-left, and by 0.6 points for switch-unknown, compared to YOLOX alone. Additionally, applying the same mechanism to YOLOv10, the latest model of YOLO series, demonstrated similar improvements.