SDFA-Net: Synergistic Dynamic Fusion Architecture With Deformable Attention for UAV Small Target Detection

This study proposes an enhanced detection algorithm based on a real-time detection transformer (RT-DETR) framework. To Address the problem of leakage and misjudgment in small-target detection in aerial images. At the network architecture level, first, a lightweight multi-scale attention feature extraction architecture is designed to improve the base module by introducing a heavily parameterized convolution and incorporating a multi-scale attention mechanism. Second, the dynamic attention interaction mechanism (attention) is used to replace the traditional Attention-based Intrascale Feature Interaction (AIFI) module, which effectively alleviates the computational complexity problem and strengthens the feature interaction efficiency. In feature processing, a Dynamic Sampling Feature Enhancement Module (Dysample-SDI) is innovatively constructed to realize the refinement of multi-scale features; whereas the Shallow Detail Fusion Module (SDFM) is designed to enhance the characterization ability of shallow features. In terms of the optimization strategy, the composite bounding box regression loss function (Focaler-MPD-IOU) was proposed to integrate the advantages of the multi-dimensional positioning evaluation index. The experimental results show that the improved algorithm improves the mAP<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0.5</sub> and mAP<sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0.5:0.95</sub> metrics by 2.3% and 1.1%, respectively, on the Visdrone2019 benchmark dataset. Through cross-domain validation of the ALU and HIT-UAV datasets, the ablation experiments and comparative analysis fully validated the effectiveness of the algorithmic improvement, and its detection accuracy and robustness were significantly improved over the baseline model. This study provides a new technical solution for small target detection in complex scenes.