A Hierarchical Context-Fusion and Synergistic Alignment Network for small object detection

Small-object detection in UAV imagery remains challenging because fine-grained target features progressively degrade, complex backgrounds cause substantial interference, and classification confidence is often misaligned with localization quality. To address these challenges, we propose a single-stage detector based on a Hierarchical Context-Fusion Network (HCFN). For feature preservation, Shallow-to-Head Feature Reorganization and Dynamic Soft-Weighted Interpolation (DSWI) retain high-resolution target cues and support stable cross-scale semantic interaction. For feature propagation, Expanded Receptive-Field Lightweight Convolution (ERL-Conv) transfers the preserved fine-grained details to deeper semantic representations while enlarging the receptive field. For feature refinement, the Adaptive Channel-Context Prioritization Module (ACCPM) combines asymmetric channel allocation, large-kernel context modeling, and frequency-domain gating to suppress background interference and enhance object-sensitive responses. Finally, the Synergistic Task-Alignment Head (STAH) improves consistency between classification and localization by combining shared task-interactive features, classification-aware modulation, and deformable regression alignment. Experiments on VisDrone2019, AI-TOD, AI-TODv2, SODA-D, and HIT-UAV demonstrate consistent improvements over the corresponding baselines.