Enhanced RT-DETR: A Lightweight Model for Aerial Infrared Small Target Detection
作者
Litian Ma,Wenjing Xie
出处
期刊:日期:2025-07-28卷期号:: 8245-8250被引量:2
标识
DOI:10.23919/ccc64809.2025.11178700
摘要
This study proposes a lightweight enhanced Real-Time Detection Transformer (RT-DETR) model to detect infrared small targets of unmanned aerial vehicle (UAV). To effectively suppress background interference and resolve multi-scale detection challenges, the proposed model integrates three core modules: (1) CGB-Block for dynamic multi-scale feature fusion, (2) TCFS (Triplet Cross-Dimensional Feature Selection) to enhance the attention on dense small targets, and (3) WT-SFF (Wavelet Transform-Selective Feature Fusion) for computational efficiency optimization. Furthermore, the Inner-Shape-IoU loss function is introduced to refine detection precision. Experimental results demonstrate that compared to the baseline RT-DETR, our improved model achieves a 38.86% reduction in parameters, 30.42% fewer floating point operations (FLOPs), and 7.68% improvements in detection accuracy, respectively. The mAP50 reaches 81.48%, highlighting its strong potential for UAV infrared small target detection scenarios.