UAV Small Object Detection Algorithm Based on Dynamic Feature Aggregation and Hierarchical Attention Mechanism

Abstract

Uncrewed Aerial Vehicle (UAV)-based small object detection is crucial for traffic monitoring, precision agriculture, and power system inspection. However, UAV images face challenges including small objects with limited pixels, inadequate features, and poor target-background discrimination. To address these challenges, we propose an enhanced small object detection algorithm called DFA-DETR. First, we design the Hierarchical Attention Feature Extraction with Adaptive Gating (HAFE-AG) module, which enhances detail feature representation for small objects through cross-stage local fusion and multi-scale multi-head self-attention mechanisms. Second, we propose the Dynamic Multi-Path Feature Pyramid Network (DMP-FPN), which employs bidirectional multi-path information flow and Shift-Enhanced Channel Reorganization (SECR) operations to achieve deep feature interaction and fusion at different scales. Finally, we design the Wavelet-Driven Contrast Enhancement Aggregation (WCEA) module, which utilizes Haar wavelet decomposition to separate low-frequency approximation and high-frequency detail information, enhancing the contrast between objects and background. Compared to the baseline RT-DETR-R18 model, DFA-DETR achieves performance improvements of 3.3% and 1.9% in AP50 and APs metrics on the VisDrone2019 dataset. Additionally, the model shows strong robustness, achieving 2% and 4.3% improvements in AP50 and APs metrics on the HIT-UAV dataset, providing an effective solution for small object detection in UAV monitoring applications.