【摘要】 针对密集粘连小麦籽粒高通量检测存在籽粒计数精度与效率难以兼顾、嵌入式部署困难的问题，本文提出一种基于LMA-DEIM的轻量化密集粘连小麦籽粒目标检测模型。构建了包含5种小麦品种共2.3??Wingdings 2MC@10~5粒籽粒目标框的WSD数据集，为模型训练与评估提供数据基础。在DEIM框架基础上，设计轻量化全局建模主干网络Lite-Mamba,以线性计算复杂度增强对粘连区域的区分能力；提出基于选择性状态空间的MIFI模块替换原AIFI模块，实现高效特征交互的同时降低计算复杂度；引入轻量化下采样模块ADown,缓解下采样过程细节丢失，提升细节特征保留能力。试验结果表明，LMA-DEIM模型在WSD测试集上精确率、召回率与mAP@50分别达到92.7%、91.5%和92.0%,检测速度达143.1 f/s,相比原DEIM框架，在精度显著提升的同时，参数量减少66.9%,速度提高240%,在高密度测试子集上计数平均相对误差仅为3.08%,满足嵌入式设备对高通量籽粒实时、高精度检测的部署需求。更多还原

【Abstract】 Aiming to address the challenges of balancing accuracy and efficiency, and the difficulty of embedded deployment, in high-throughput detection of densely clustered wheat grains, a lightweight LMA-DEIM-based model for detecting densely clustered wheat grains was proposed. Firstly, a WSD dataset containing 230 000 grain target boxes from five wheat varieties was constructed to provide a data foundation for model training and evaluation. Secondly, based on the DEIM framework, a lightweight global modeling backbone network, Lite-Mamba, was designed to enhance the ability to distinguish clustered regions with linear computational complexity. A MIFI module based on selective state space was proposed to replace the original AIFI module, achieving efficient feature interaction and reduced computational complexity. A lightweight downsampling module, ADown, was introduced to mitigate detail loss during downsampling and improve the ability to preserve detailed features. Experimental results showed that the LMA-DEIM model achieved precision, recall, and mAP@50 of 92.7%, 91.5%, and 92.0% on the WSD test set, respectively, with an inference speed of 143.1 f/s and an average relative counting error of only 3.08% on the high-density test subset. Compared with the original DEIM framework, the proposed method significantly improved accuracy while reducing the number of parameters by 66.9% and increasing speed by 240%, meeting the deployment requirements of embedded devices for high-throughput, real-time, and high-precision grain detection.更多还原