Dynamic multi-scale mixture-of-experts with cross-scale feature enhancement for UAV small object detection

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).

Detecting small objects in unmanned aerial vehicle (UAV) imagery remains a persistent challenge due to scale variations of small objects in UAV scenarios and their low pixel occupancy. Existing approaches suffer from two issues in feature extraction and fusion. First, backbone networks employ fixed receptive fields, precluding input-adaptive adjustment and thereby failing to capture adequate contextual information for small objects. Second, cross-scale feature fusion predominantly relies on layer-by-layer chain propagation, which induces progressive degradation of high-resolution shallow details during transmission. Moreover, existing adaptive weighting mechanisms are restricted to channel-wise interactions between adjacent layers, lacking cross-layer semantic guidance and deep-level noise suppression; consequently, small object features are easily overwhelmed by large objects or background clutter. To address these issues, we propose DCE-DETR, a dedicated framework for UAV small object detection. For feature extraction, the Dynamic Multi-scale Mixture-of-Experts (DMS-MoE) adaptively selects optimal receptive field combinations through a dynamic routing mechanism that perceives global context, substantially enhancing scale perception. For feature fusion, the Gather-Scatter Pyramid Network (GSPN) adopts a middle-layer-centric nonlinear interaction architecture to shorten the transmission path between shallow details and deep semantics. Moreover, it integrates the Multi-scale Semantic Gated Attention (MSGA) module to enable cross-layer semantic guidance and suppress irrelevant noise. Extensive experiments on four public benchmarks, VisDrone2019, UAVDT, HazyDet, and SIMD, demonstrate consistent improvements of 3.0%, 2.4%, 2.6%, and 2.6% in AP over the baseline, alongside a 15.8% reduction in parameters.Under matched computational budgets, DCE-DETR significantly outperforms current state-of-the-art(SOTA) methods. Code and models are available at https://github.com/LXYSOD/DCE-DETR .