MAFNet-DETR: Multi-branch Adaptive Fusion Network for Small Object Detection in UAV Aerial Images

Small object detection in high-altitude unmanned aerial vehicle (UAV) imagery presents formidable challenges, including severely constrained target scale and intricate background interference. To address these issues, this paper proposes MAFNet-DETR, a novel detection framework built upon RT-DETR-R18, designed to augment feature representation capability on densely populated small-object benchmarks such as VisDrone2019. MAFNet-DETR systematically advances detection performance through four core architectural innovations: the Three-branch Gated Enhancement Block (TGEB), which employs large-kernel convolutions to capture expansive global contextual information in conjunction with a gating mechanism for adaptive salient feature modulation, substantially elevating the model’s perceptual sensitivity to sparse and inconspicuous signals; the Global-Local Spatial Attention (GLSA) module, which adopts a dual-branch parallel architecture to independently extract fine-grained local spatial details and long-range global dependencies, thereby achieving an effective equilibrium between local perceptual acuity and holistic contextual comprehension; the Adaptive Bidirectional Feature Pyramid Network (Adaptive-BiFPN), which incorporates a dual-weight fusion mechanism coupling spatially adaptive weights with globally learnable weights to realize fine-grained and computationally efficient multi-scale feature aggregation, significantly mitigating the degradation of small-object detail information during hierarchical feature propagation; and the Lightweight Progressive Upsampling Block (LPUB), which achieves high-fidelity feature map resolution restoration through the synergistic integration of depthwise separable convolutions, channel shuffling, and shift channel mixing operations. Extensive experimental evaluations conducted on three publicly available benchmark datasets—VisDrone2019, SIMD, and AI-TOD—demonstrate that, relative to the baseline RT-DETR-R18, the proposed method achieves improvements in mAP₅₀ of 6.1%, 4.6%, and 4.8%, and in mAP₅₀₋₉₅ of 4.5%, 2.7%, and 3.3%, respectively, while concurrently reducing the parameter count by 12.46%, 8.7%, and 9.45%. The proposed model attains substantial gains in detection accuracy and exhibits strong applicability to small object detection under complex background conditions.