面向无人机边缘计算的小麦麦穗计数轻量化模型研究

Lightweight Model of Wheat Ear Counting for UAV Edge Calculation

小麦麦穗的高效计数对快速、准确掌握小麦产量具有重要意义。无人机由于具有效率高、成本低等特点被广泛应用于大田小麦信息的采集。但已有的用于小麦麦穗计数的深度学习模型结构复杂、参数量大,不能直接部署在存储空间有限的无人机的边缘设备上。针对这一问题,提出了一种融合剪枝策略和知识蒸馏的模型压缩方法,基于YOLOv5s模型构建了一种轻量化模型,并设计了面向无人机边缘计算的小麦麦穗计数轻量化方案。试验结果表明,经过模型剪枝和知识蒸馏轻量化处理的YOLOv5s模型,在小麦计数任务上的计数准确率为93.3%,模型的mAP(mean Average Precision,平均精度均值)达到94.4%,模型大小缩小了约76%,模型参数量减少了79.61%。因此,模型在保持较高的计数准确率的同时将会占用更少的计算资源和存储空间,显著的压缩效果使模型可以部署在无人机的边缘设备上,为小麦麦穗的实时计数提供了可能。

The efficient counting of wheat ears is important for the rapid and accurate control of wheat yield.UAV is widely used in collecting field wheat information due to its high efficiency and low cost.However,the existing deep learning models for wheat ear counting have complex structure and large number of parameters,and cannot be directly deployed on the edge devices of UAVs with limited storage space.To address this issue,we proposed a model compression method combining pruning strategy and knowledge distillation,constructed a lightweight model based on YOLOv5s model,and designed a wheat ear count lightweight scheme for UAV edge computing.Experimental results show that the pruned and distilled YOLOv5s model achieves a counting accuracy of 93.3%and m AP of 94.4%in wheat counting tasks.The model size is reduced by approximately 76%,and the parameter count is decreased by 79.61%.Consequently,the model will take up less computing resources and storage space with high counting accuracy.The remarkable compression effect makes the model enables the model to be deployed on the edge equipment of UAV,providing the possibility for real-time counting of wheat ears.