
Guanhua Zhu, Pengfei Cao · Engineering Research Express 2026 · 2026
DOI: 10.1088/2631-8695/aeb858
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Complex scene perception and dense small object detection in unmanned aerial vehicle (UAV) imagery frequently encounter severe challenges, including extremely small object scales, strong background interference, and severe occlusion. Conventional vision-only detection frameworks typically perform feature extraction and fusion solely at the pixel level, lacking the synergy between explicit semantic information and fine-grained localization cues. To address these limitations, this paper proposes a novel multimodal object detection framework driven by text prompts, alongside frequency-domain and structural priors. First, a text semantic prompting mechanism is introduced to transform closed-set detection into a cross-modal feature alignment task, utilizing multimodal semantics to explicitly direct visual representations toward target regions. Second, in the deep encoding stage, an Adaptive Differentiable Histogram of Oriented Gradients (HOG)-Guided Self-Attention (ADGA) module is designed to inject multi-scale continuous gradient features as structural priors into the self-attention computation. During the feature fusion stage, a Competitive Interactive Dynamic Alignment Fusion (CAF) module is proposed. By explicitly modeling the discrepancies and consistency among feature branches and introducing a competitive allocation mechanism, it achieves a content-driven reorganization of features at the same scale. Finally, in the pre-decoder stage, a Frequency-Decoupled Dual Attention (FDDA) module is introduced. It utilizes differentiable Gaussian soft masks to decouple high-frequency and low-frequency components and applies customized spatial and channel attention, respectively, thereby providing high signal-to-noise ratio (SNR) features for the decoder. Extensive experiments on the VisDrone2019 dataset demonstrate that the proposed model achieves an overall Average Precision (AP) of 0.246 and a small object average precision (APs) of 0.171, with only 15.80M parameters and 30.92 GFLOPs. Furthermore, additional experiments on the UAVDT dataset yield an AP of 0.225, an AP50 of 0.357, and an APs of 0.154, further demonstrating the robustness and cross-dataset generalization capability of the proposed framework in diverse UAV scenarios.
No comments yet — start the discussion below.