MLRT-DETR: A Multi-Directional and Multi-Level Small Object Detection Algorithm from UAV Perspective

To address the challenges of missed and false detections caused by densely distributed small objects, complex backgrounds, and limited informative cues in UAV-view images, this paper proposes MLRT-DETR (Multi-Layer RT-DETR),a multi-Level modeling small object detection algorithm Based on RT-DETR.First, a feature extraction module SDFE (Selective Dynamic Feature Extraction) is proposed. It leverages dynamic weight attention and a gating mechanism to capture key information, followed by randomized feature enhancement, thereby improving the model's ability to extract edge features of small objects.Second, polarity-aware linear attention and dynamic tanh are introduced to construct the PD-AIFI (Polarity-aware Dynamic AIFI) module, which enhances the model's spatial structure learning capability and nonlinear modeling efficiency.Finally, a feature fusion module MFF (Multi-path Feature Fusion) is designed. This module fuses shallow features and performs multi-level modeling, thereby enhancing the model's ability to perceive image textures, edges, and structural regioans. Experimental results demonstrate that on the Visdrone2019 dataset, compared with RT-DETR, MLRT-DETR achieves a 2.1% improvement in mAP@0.5 and a 1.0% improvement in mAP@0.5:0.95, while reducing the number of parameters by 5.7M and the computational complexity by 3.5Gflops. In comparison with algorithms such as YOLOv11M and DEIM, MLRT-DETR exhibits excellent performance in both accuracy and speed, indicating favorable application value.