FR-DETR: a streamlined DETR for remote sensing object detection

Abstract Nowadays, remote sensing object detection has demonstrated broad application prospects in numerous fields. However, the variety of target types, large scale variations, and complex backgrounds pose significant challenges to accurate detection. To address these issues, this paper proposes a novel algorithm, FR-DETR, which builds upon the RT-DETR framework, with its core comprising the innovative lightweight backbone network FLNet. FLNet achieves efficient feature extraction and performance enhancement through the streamline fusion block module. Additionally, the integration of the Conv3XC module to form the Dual Residual Bottleneck module enhances feature fusion capability, while the adoption of the hierarchical weighted downsampling module preserves image details and improves detection accuracy. Experimental validation demonstrates that FR-DETR significantly boosts detection metrics across multiple authoritative remote sensing datasets, showcasing its robustness in complex scenarios. Moreover, as an RT-DETR-based method, FR-DETR excels in processing speed and parameter count, particularly in lightweight design, making it suitable for real-time applications and deployment in resource-constrained environments.