基于改进RT-DETR的轻量级小目标检测方法

Lightweight Small Object Detection Method Based on Improved RT-DETR

摘要:

无人机航拍图像的小目标检测性能易受不规则拍摄角度、视角变化、遮挡及光照变化等因素影响而显著下降。为此，提出轻量级的无人机目标检测模型LDRT-DETR；针对小目标与背景尺度差异显著的问题，引入多尺度边缘特征提取单元，从不同尺度获取特征并增强边缘信息，以提升模型对小目标的感知能力；设计上下文增强特征融合方法，以更有效整合上下文信息并减少融合过程中的细节损失，从而提高检测精度；针对无人机航拍图像中背景复杂、遮挡严重的特点，提出多尺度采样模块，该模块结合上采样与下采样机制，并通过多分支结构构建丰富的特征提取路径，从而增强特征表达能力；实验结果表明，LDRT-DETR在VisDrone2019数据集上的mAP50达到50.05%，较RT-DETR提升2.82%，同时参数量和GFLOPs分别下降24.7%和10%。

Abstract:

The performance of small-object detection in UAV aerial imagery is highly susceptible to degradation caused by irregular shooting angles, viewpoint variations, occlusions, and illumination changes. To address these issues, we propose LDRT-DETR, a lightweight object detection model tailored for UAV scenarios. To mitigate the pronounced scale discrepancy between small objects and their backgrounds, a multi-scale edge feature extraction module is introduced to capture features across different scales and enhance edge information, thereby improving the model’s ability to perceive small objects. A context-enhanced feature fusion strategy is designed to more effectively integrate contextual cues and reduce detail loss during feature fusion, leading to improved detection accuracy. Additionally, to cope with complex backgrounds and severe occlusions in UAV aerial images, a multi-scale sampling module is developed. This module incorporates both upsampling and downsampling mechanisms and constructs diverse feature extraction pathways through a multi-branch structure, thereby enriching feature representation. Experimental results on the VisDrone2019 dataset demonstrate that LDRT-DETR achieves an mAP50 of 50.05%, surpassing RT-DETR by 2.82%, while reducing the number of parameters and GFLOPs by 24.7% and 10%, respectively.