BF-DETR: A Lightweight and Adjustable Model Based on RT-DETR

As the first real-time detection model in the DETR (DEtection TRansformer) series, RT-DETR (Real-Time DEtection TRansformer) surpasses previous models in both speed and accuracy, while demonstrating strong competitiveness against the YOLO (You Only Look Once) series. However, it faces the problems of excessive parameters and limited capability in detecting small targets. To mitigate these problems, this paper proposes a lightweight and adjustable model: BF-DETR.We introduce a lightweight Branch-based Feature Extraction Network (BFENet) as the backbone, the Branch-based Feature Extraction Module (BFEM) within BFENet extracts multi-scale features and achieves cross-level feature fusion. Additionally, by employing a differentiated down-sampling strategy, BFEM effectively preserves spatial detail features. BFENet resolves the issues of singular feature extraction and excessive parameters in the original backbone. Furthermore, we propose the Frequency-Divided Attention-based Intra-scale Feature Interaction (FDAIFI) module. Its core component, Frequency-Divided Self-Attention, performs targeted attention computations on features of different frequencies, reducing computational complexity while effectively utilizing multi-frequency features. Experiments conducted on the VisDrone2019 dataset validate the effectiveness of BF-DETR. Compared to RT-DETR-R18, BF-DETR reduced the number of parameters by 22% and GFLOPs by 24%, while improving themAP50 andmAP50:95 by 0.9% and 0.7%, respectively. Additionally, experiments on the Pascal VOC and Urinary Sediment 11(Urised11) datasets further confirm the generalization capability of BF-DETR.