基于MobileViT模型的小麦收获机喂入密度分类方法

基于喂入量的作业速度智能化控制技术是优化联合收获机作业效率和质量的重要手段。本文针对传统喂入量自动控制技术时滞明显，在喂入量调整时无法及时适应实际情况的问题，采用基于图像的深度学习方法开展了成熟期小麦植株密度等级分类识别方法研究，通过预先感知作物密度，实现联合收获机作业参数的自动调整。首先基于车载相机和无人机图像构建了小麦植株图像数据集，并细分为低密度、中密度、高密度和特高密度4类；其次构建了基于MobileViT-XS轻量化网络的密度等级识别模型，利用建立的数据集进行模型的训练和测试；最后将其与VGG16、GoogLeNet和ResNet进行了比较。结果表明，MobileViT-XS模型的总体识别准确率达到91.03%，且单幅图像推理时间仅为29.5ms。与VGG16、ResNet网络相比，总体识别准确率分别高出3.51、2.34个百分点，MobileViT-XS模型可以较好的完成小麦不同密度等级的分类识别任务，为实时预测小麦喂入密度提供了技术支持。

Classification Method for Wheat Harvester Feeding Density Based on MobileViT Model

Intelligent control technology of operating speed based on feeding rate is an important means to optimize the efficiency and quality of combine harvester operations. Aiming at the obvious time delay of the traditional feeding rate automatic control technology and the inability to adapt to the actual situation in time when the feeding rate is adjusted. By analyzing the influencing factors of the feeding rate, an image based deep learning method was used to carry out a research on the classification and recognition method of wheat plant density in the mature stage. By sensing crop density in advance, the operating parameters of the combine harvester can be automatically adjusted. Firstly, a multi-variety and multi-region mature stage wheat plant image dataset was constructed based on vehicle-mounted cameras and UAV images, and classified it into four categories: low density, medium density, high density, and very high density. Next, a density classification recognition model was built based on the lightweight MobileViT-XS network, and trained and tested the model by using the established dataset. Finally, it was compared with VGG16, GoogLeNet, and ResNet. The results showed that the overall recognition accuracy of the MobileViT-XS model reached 91.03%, and the inference time for a single image was 29.5ms. Compared with VGG16 and ResNet networks, the overall recognition accuracy was 3.51 percentage points and 2.34 percentage points higher respectively. The MobileViT-XS model can effectively accomplish the classification recognition of wheat at different density levels, providing technical support for real-time prediction of wheat feeding rate.