OC-DETR: DETR with Orthogonal Channel Attention and CSPO-Fusion for component detection of TEDS

In recent years, the evolution of automated maintenance technology for high-speed railways has positioned the Train of EMU failures Detection System (TEDS) as a critical technology for ensuring the safe operation of high-speed trains. TEDS employs line-scan cameras along railway tracks to capture undercarriage and side-view images of Electric Multiple Units (EMUs). However, the diverse shapes of EMU components, complex component backgrounds, and a large number of small components have intensified the challenges of designing component detection algorithms. To address these limitations, this paper introduces a novel real-time component detection framework known as OC-DETR (DETR with Orthogonal Channel Attention and CSPO-Fusion). The framework incorporates a specialized feature fusion module named CSPO-Fusion, within the Cross-Scale Feature Fusion Module (CCFM) to adeptly handle the small target features of EMU components. Additionally, the framework integrates the orthogonal channel attention into the block of the backbone, effectively reducing the redundant information caused by traditional convolution and improving the accuracy of component detection. To further enhance performance across diverse scales, we propose an IoU-aware size-dependent weighted loss function which increases detection precision for small, medium, and large components. In this paper, the TEDS installed at the throat of the train is used to collect the data of EMUs, and the TEDS dataset including CRH380 and CR400AF is constructed. Extensive experimental validation using this dataset demonstrates that the accuracy and speed of the proposed method have improved compared with the current mainstream target detector, and the detection mAP50 of this method is increased by 2.2% compared with the RT-DETR benchmark target detection algorithm, which effectively verifies the effectiveness of this algorithm.

Keywords: TEDS, Orthogonal channel attention, CSPO-fusion, Component detection, Lightweight network.