Dual-domain attentions for unmanned aerial vehicle small object detection

Abstract

Images captured by unmanned aerial vehicles (UAVs) often suffer from severe degradation in small object quality and resolution due to environmental constraints, posing significant challenges in preserving the dual-domain characteristics of spatial details and frequency components. While large-scale models attempt to address this through complex architectures, aggressive down-sampling and successive convolution operations inevitably erase fine-grained patterns that are essential for detecting small objects. To overcome these challenges, we propose a dual-domain attention mechanism for small object detection, which focuses on both spatial and frequency domains. In the spatial domain, the proposed step-free triple-attention convolution (SFTAConv) reduces information loss during feature propagation by combining spatial–channel interactions and a lossless space-to-depth transform, thereby enhancing subtle object patterns while suppressing background interference. In the frequency domain, the frequency-domain hybrid attention (FD-HAT) jointly recalibrates high- and low-frequency components, moving beyond single-domain recalibration to recover discriminative representations of occluded or blurred small objects. Additionally, a classification-assisted localization (CAL) branch with classification-guided localization further refines detection accuracy. After extensive experiments conducted on the vision meets drone 2019 object detection (VisDrone2019Det), dataset for object detection in aerial (DOTA), and pascal visual object classes (PASCAL VOC) datasets, the results demonstrate that our model achieved the significant gains of 2.2%, 1.7%, 5.3% at $A{P}_s$ metric on three datasets, respectively, and being competitive with the state-of-the-art (SOTA) detectors.