Abstract: Object detection from an Unmanned Aerial Vehicle (UAV) perspective faces severe challenges, including highly complex backgrounds, drastic scale variations, and a lack of discriminative features for small objects. To overcome the performance bottlenecks of existing models in complex airspace environments, a UAV-oriented object detection model, LHM-DEIM, based on the DEIM (DEtection TRansformer with Improved Matching)-DFINE architecture, was proposed. To address the issue of texture loss in small objects, a Laplacian of Gaussian (LoG) Stem was introduced for initial feature extraction, utilizing the Laplacian operator to strengthen the representation of high-frequency spatial features. To capture cross-scale global correlations, the model's neck was reconstructed into a Hypergraph Hybrid Encoding Neck (H-HEN). In this neck, a Spatial Scale Alignment (SSA) mechanism was implemented, using the P4 layer as the base scale and other layers as enhancement features within a hierarchical fusion strategy. Furthermore, a hypergraph modeling strategy was incorporated to significantly increase the correlation density of high-order features. Additionally, a MutualGuideFusion module was designed to effectively mitigate feature attenuation in deep networks through a bidirectional information excitation mechanism, achieving efficient integration of multi-scale features. Experimental results on the VisDrone2019 dataset demonstrate that the LHM-DEIM model increases the average precision at an IoU threshold of 0.5 (AP₀.₅) from the baseline value of 34.3% to 38.2%, representing an increase of 3.9 percentage points. Specifically, the average precision for small objects (APs) increases by 2.5 percentage points. Ablation and generalization experimental results further confirm that the model has strong robustness and practicality in drone scenarios. 

Key words: Unmanned Aerial Vehicle (UAV), Laplacian of Gaussian (LoG) operator, hypergraph feature processing, feature fusion, small object detection, DETR with Improved Matching (DEIM) 

摘要： 无人机(UAV)视角下的目标检测面临背景高度复杂、尺度剧烈变化以及小目标特征匮乏等严峻挑战。为突破现有模型在复杂空域环境下的性能瓶颈，以DEIM(DETR with Improved Matching)-DFINE(Redefine Regression Task in DETRs as Fine-grained Distribution Refinement)为基准架构，提出一种UAV视角下的目标检测模型——LHM-DEIM。针对小目标纹理丢失问题，引入高斯-拉普拉斯(LoG)初始特征提取模块(LoGStem)，利用拉普拉斯算子强化对高频空间特征的表征能力；为了捕获跨尺度的全局关联，重新构建模型颈部,提出一种超图混合编码颈部(H-HEN)，在颈部使用空间尺度对齐(SSA)机制，使用P4层作为基准尺度，其余各层作为增强特征的层级融合策略，并且引入超图（Hypergraph）建模策略，可以显著提升高阶特征的关联密度；设计互引导融合（MutualGuideFusion）模块，通过双向信息激励机制有效缓解深层网络中的特征衰减问题，以实现多尺度特征的高效集成。在VisDrone2019数据集上的实验结果表明，LHM-DEIM模型的平均精度AP0.5从基线模型DEIM-DFINE的34.3%提升至38.2%，AP0.5相较于基线提升了3.9个百分点，小目标的检测精度APs提升了2.5个百分点。消融实验和泛化实验结果进一步验证了，该模型在UAV场景下具有较强的鲁棒性与实用性。

关键词: 无人机, 高斯-拉普拉斯算子, 超图特征处理, 特征融合, 小目标检测, DEIM