LDATA-Net: Dynamic feature adaptation for efficient feature learning in resource-limited UAV detection

Unmanned Aerial Vehicle (UAV) image analysis faces the dual challenges of complex background interference and limited onboard computational resources, particularly when processing extreme scale variations across multiple viewpoints. Existing approaches typically enhance detection accuracy by increasing model complexity, but this often leads to parameter proliferation that exceeds the deployment limits of airborne platforms. To address this fundamental contradiction, we propose LDATA-Net (Lightweight Dynamic Aggregation Task-Aligned Network), which pioneers a “Dynamic Feature Adaptation” design paradigm aimed at achieving synergistic optimization between parameter efficiency and detection accuracy. This framework systematically realizes end-to-end dynamic adaptive capabilities through three core components that operate collaboratively across feature extraction, fusion, and detection stages: (1) Dynamic Multi-Branch Depthwise Block (DMBD-Block), whose core innovation is our proposed novel operator DIDWConv, which adaptively adjusts receptive fields according to input features to capture targets of extreme scales and orientations; (2) Lightweight Dynamic Aggregation Network (LDANet), which effectively preserves critical spatial contextual information through hierarchical fusion architecture and dynamic weighting mechanisms; (3) Dynamic Adaptive Head (DA-Head), which effectively mitigates task conflicts through geometric and semantic dynamic feature alignment. LDATA-Net achieves 35.4 %, 77.9 %, and 51.2 % AP${}_{50}$ on VisDrone2019, DOTA1.0, and AI-TODv2 datasets respectively with only 2.8M parameters, establishing a new paradigm for designing memory-efficient yet high-performance detection systems, particularly for resource-constrained heterogeneous computing aviation platforms.