作者：Shiyan Pang, Zhiheng Yi, Pengcheng Zhou, Yuqin Yang（杨玉芹）, Kejiang Xiao*（肖克江）, Zhiqi Zuo*

出版刊物：Expert Systems with Applications

出版时间：2026

内容摘要：

The detection of individual and organ-level object during the growth period of crops is a core component of precise phenotypic analysis in modern intelligent breeding systems. Utilizing unmanned aerial vehicle (UAV) remote sensing imagery for maize seedling and wheat head detection has become a crucial technique for achieving high-throughput, non-destructive field phenotyping. However, existing methods still face limitations in performance when dealing with dense and overlapping agricultural scenes in the field. Major challenges include small object scales, high-density distributions, complex backgrounds, and variable imaging perspectives. To address these issues, this paper proposes a dense and overlapping object detection method based on wavelet-enhanced sparse Mixture-of-Experts (S-MoE). First, multi-level features are extracted through a backbone network to capture multi-scale spatial representations. Next, discrete wavelet transform (DWT) is introduced at each feature level to decompose spatial features into the frequency domain, enabling collaborative modeling of frequency and spatial domain features, thereby enhancing the preservation of fine details and texture information. Additionally, a dynamic hypergraph aggregation module is incorporated into the deepest feature layer to adaptively learn hyperedge structures, explicitly modeling high-order relationships within local regions. Furthermore, during the decoding phase, we implement an IoU-aware query selection mechanism to filter and re-weight candidate queries. Specifically, a sparse Mixture-of-Experts (S-MoE) dynamic routing mechanism is introduced to enable specialized, query-adaptive representation learning. These combined strategies effectively improve detection accuracy and reliability in complex scenarios. Finally, to validate the generalization capability and effectiveness of the proposed method, experiments are conducted on a self-constructed maize seedling dataset comprising nearly thirty thousand images, as well as on a public wheat head detection dataset. The proposed method is also compared against state-of-the-art DETR-based models (DINO, DAC-DETR, RT-DETR, DN-DETR and Relation-DETR) and YOLO-based models (YOLOv8-L, YOLOv10-L, YOLOv11-L and YOLOv13-L). Extensive experiments show that our method, MoE-DETR, effectively balances accuracy and efficiency through its dynamic routing mechanism, achieving AP50: 95 scores of 93.2% on the maize seedling dataset and 33.5% on the wheat head detection dataset–outperforming existing state-of-the-art approaches.

作物生长期个体与器官级目标检测是现代智能育种系统中精准表型分析的核心组成部分。利用无人机（UAV）遥感影像进行玉米幼苗与麦穗检测，已成为实现高通量、无损田间表型分析的关键技术。然而，现有方法在应对田间密集且重叠的农业场景时仍存在性能局限，主要挑战包括目标尺度小、分布密度高、背景复杂以及成像视角多变。为解决上述问题，本文提出一种基于小波增强稀疏专家混合（S-MoE）的密集重叠目标检测方法。首先，通过骨干网络提取多层级特征，以捕捉多尺度空间表征。其次，在各特征层级引入离散小波变换（DWT），将空间特征分解至频域，实现频域与空间域特征的协同建模，从而增强细节与纹理信息的保留能力。此外，在最深层特征层中引入动态超图聚合模块，以自适应学习超边结构，显式建模局部区域内的高阶关系。进一步地，在解码阶段，我们实现了一种IoU感知的查询选择机制，对候选查询进行筛选与重加权。具体而言，引入稀疏专家混合（S-MoE）动态路由机制，以实现专门的、查询自适应的表征学习。上述策略相结合，有效提升了复杂场景下的检测精度与可靠性。最后，为验证所提方法的泛化能力与有效性，在自建的包含近三万张图像的玉米幼苗数据集以及公开麦穗检测数据集上开展了实验。所提方法还与当前最先进的基于DETR的模型（DINO、DAC-DETR、RT-DETR、DN-DETR和Relation-DETR）以及基于YOLO的模型（YOLOv8-L、YOLOv10-L、YOLOv11-L和YOLOv13-L）进行了对比。大量实验表明，我们的方法MoE-DETR通过其动态路由机制有效平衡了精度与效率，在玉米幼苗数据集上取得AP50:95为93.2%、在麦穗检测数据集上取得33.5%的成绩，优于现有最先进方法。