Answers are grounded in this paper and cite their sources.
Small-object detection in UAV remote sensing imagery is challenging because target boundaries and spatial cues are easily weakened by successive downsampling operations, while substantial scale variation and complex backgrounds further increase classification and localization difficulties. Although existing multi-scale detectors aggregate hierarchical features, fusion often occurs after part of the shallow detail has been lost, whereas repeated high-resolution processing increases computational cost. To improve the representation of weak and densely distributed objects, this paper proposes AeroPDT, an efficient end-to-end detector based on progressive detail transmission. AeroPDT preserves, enriches, and transfers small-object information throughout feature extraction. First, an EdgeAwareStem is designed to introduce a fixed Laplacian-based high-frequency residual pathway before the second downsampling operation, thereby preserving object contours and local contrast. Subsequently, HG_Stage_PKI_Lite employs lightweight multi-receptive-field feature extraction on compressed channels to capture local structures and contextual information with limited computational overhead. Finally, a PDH-FPN is developed to inject directional P2 details into the P3 feature pathway and enhance detail-semantic interaction through high-frequency perception and local cross-scale dependency modeling. Experimental results on the VisDrone2019 dataset show that AeroPDT achieves an mAP₅₀ of 44.4%, an mAP₅₀₋₉₅ of 26.2%, and scale-specific mAP_S , mAP_M , and mAP_L values of 22.7, 36.1, and 42.5%, respectively. The model contains 9.63 M parameters and reaches 126.9 FPS. Further ablation studies, visual analyses, and experiments on the HIT-UAV infrared dataset provide preliminary evidence of the framework’s cross-modal adaptability. AeroPDT provides an effective balance among small-object detection accuracy, model complexity, and real-time inference for UAV remote sensing applications.
No takes yet. Share an insight, caveat, or question.
Sun et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: