MicroSight-DETR: spatial-preserving real-time transformer with multi-domain fusion for UAV micro-object detection

Addressing the technical challenges of low detection accuracy for small targets in UAV monitoring scenarios, this paper proposes MicroSight-DETR, an enhanced real-time detection model based on RT-DETR-r18. Through systematic analysis of RT-DETR’s behavior on UAV aerial imagery, we identify three stage-specific degradation mechanisms operating at successive stages of the detection pipeline: a receptive field bottleneck at the feature extraction stage, a single-domain representation limitation at the feature enhancement stage, and a spatial information collapse at the feature fusion stage. Guided by this analysis, the model introduces three complementary modules each targeting a specific identified weakness: the Global Efficient Modeling (GEM) backbone utilizes EfficientViM’s linear-complexity Mamba architecture combined with CGLU gating mechanisms to achieve global feature modeling with O(L) complexity; the Multi-domain Adaptive Fusion Dynamics (MAFD) encoder integrates polarized attention, spectral frequency enhancement (SEFM), efficient feature mixing (MonarchMixer), and dynamic multi-scale aggregation (DyHead) to achieve adaptive spatial-frequency dual-domain fusion; and the Spatial Preserving Aggregation with Multi-scale (SPAM) neck employs SPDConv spatial-preserving transformation and CSPOmniKernel multi-scale convolutions to preserve shallow detail features critical for small target detection. Extensive experiments on the VisDrone2019 dataset demonstrate that MicroSight-DETR achieves 51.3% mAP@0.5, representing a 9.9% relative improvement over the baseline, while mAP@0.5:0.95 increases by 12% from 28.4% to 31.8%. Progressive ablation experiments show that the three modules contribute largely additive improvements, providing empirical evidence for their complementary design. The proposed MicroSight-DETR achieves real-time inference (78 FPS) on server-grade GPUs with only 16.1M parameters and 64.5 GFLOPs, providing a robust solution for ground station-based processing of UAV-captured imagery.