FDM-RTDETR: A Multi-Scale Small Target Detection Algorithm

This is the authors' abstract. We don't add key points for this paper.

To address challenges in UAV object detection including complex backgrounds, severe occlusion, dense small objects, and varying lighting conditions, we propose FDM-DETR, a novel detection algorithm specifically designed for small objects in UAV imagery. This method effectively captures global image information by fusing multi-scale spatial features and performing feature extraction in the frequency domain within the backbone network. We design a Dynamic Feature Interaction (DIFI) module with position-based biases in the encoder, enhancing the model’s perception of local features for small objects. In the neck network, we introduce a Multi-Scale Feature Enhancement Pyramid (MSFEP) module to improve feature extraction capabilities for small object detection. Compared to RT-DETR, our improved model achieves performance gains of 2.5% and 2.7% in AP on the Vis-Drone2019 validation and test sets, respectively. While maintaining low computational complexity and parameter count, the method demonstrates significant improvements in detection performance. FDM-DETR exhibits robust practicality and reliability in UAV-based small object detection tasks.