A Multi-branch Fusion Hierarchical Representative Network for Foreign Object Detection in High-speed Railway Catenary Systems

The catenary system serves as a power transmission line for train operation. Foreign objects such as plastic bags attached on the catenary lines pose significant safety risks to train operation. Therefore, timely and accurate detection of foreign objects in catenary systems is of great significance. However, existing detection methods suffer from insufficient accuracy and high computational complexity when dealing with small object detection in complex environments. To address these challenges, this paper proposes two key improvements: (1) a Multi-branch Fusion Hierarchical Representative network (Multi-branch Fusion-HR backbone) for efficient multi-scale feature extraction through five progressive stages, (2) a Small Object-aware Enhancement Module (Small Object-aware Enhancement) that combines multi-scale channel attention with spatial and gradient-aware mechanisms. Tests on the catenary dataset shows significant improvements over You Only Look Once version 11, that is accuracy increased from 91.3% to 93.5%, and mean Average Precision at 0.5 Intersection over Union from 93.3% to 96.9%. Furthermore increasing parameters from 2.58 million to 4.25 million, inference speed is improved from 163.93 to 200 Frames Per Second, which indicates an effective balance between accuracy and efficiency for catenary system monitoring.