Dual-domain perception network for small object detection in UAV aerial images

Aerial object detection based on Unmanned Aerial Vehicles (UAVs) is crucial for applications such as mapping, logistics, and agriculture. However, UAV aerial scenarios contain numerous small objects with limited visual information, which are easily overwhelmed by complex background, posing significant challenges for detectors. To address this problem, we propose a Dual-Domain Perception Network (DDPNet) that comprehensively enhances small object representation through joint spatial- and frequency-domain perception, comprising three core modules: Global-Local Context Modeling Module (GLCMM), Frequency-Domain Understanding Module (FDUM), and Selective Feature Fusion Module (SFFM). GLCMM models global object-scene relationships to adaptively capture foreground-relevant context. FDUM employs Discrete Wavelet Transform (DWT) with selective feature extraction and adaptive weighting to enhance small object discrimination against complex background. SFFM adaptively fuses spatial- and frequency-domain features, reinforcing small object cues while suppressing irrelevant background interference. DDPNet achieves mAPs of 58.1%, 40.7%, and 62.2% on AI-TOD, VisDrone-DET, and TinyPerson datasets, respectively, surpassing state-of-the-art detectors and demonstrating its effectiveness and robustness in small object detection.