融合注意力及多重知识迁移的茶叶病害轻量化检测方法  

A lightweight detection method for tea disease by merging attention and multiple-knowledge-transfer

作　　者：毛致颖 刘宇航 杨春勇[1,2,3] 田永胜 倪文军 王曦照[2] Mao Zhiying;Liu Yuhang;Yang Chunyong;Tian Yongsheng;Ni Wenjun;Wang Xizhao(College of Electronics and Information Engineering,South-Central Minzu University,Wuhan,430074,China;Hubei Engineering Research Center of Intelligent Internet of Things Technology,Wuhan,430074,China;Hubei Key Laboratory of Intelligent Wireless Communications,Wuhan,430074,China)

机构地区：[1]中南民族大学电子信息工程学院,武汉市430074 [2]智能物联技术湖北省工程研究中心,武汉市430074 [3]智能无线通信湖北省重点实验室,武汉市430074

出　　处：《中国农机化学报》2024年第12期140-147,F0002,共9页Journal of Chinese Agricultural Mechanization

基　　金：湖北省自然科学基金创新群体项目(2024AFA030)。

摘　　要：茶树病虫害是影响茶叶产量及品质的主要原因,精准检测茶叶病虫害种类是当前国内的热点问题之一。针对传统目标检测网络模型参数量大、精确率低导致工业部署困难的问题,建立茶叶病虫害表型图像数据集;对网络模型进行轻量化处理,优裁基于知识蒸馏的多重知识迁移训练模型;构建基于视觉注意力模块(CSA)的YOLOv5目标检测网络模型,优化茶叶病虫害检测方法。结果表明,添加视觉注意力模块(CSA)的YOLOv5目标检测模型与YOLOv5网络模型、添加传统注意力模块SE、CBAM模块的YOLOv5网络模型相比较,其平均准确率分别提高3.1%,1.1%,1%。对比蒸馏前学生模型,构建的模型最佳准确率提升4.1%,对比教师模型,模型容量降低5.4 MB,单帧图片推理时间下降35%。设计的网络模型在不损失准确率的情况下,降低网络计算的开销,可为资源受限的农业信息化领域边缘计算系统提供植入可能。Tea plant diseases and pests are the main factors affecting tea production and quality.The accurate detection of tea diseases and pests is one of the current hot issues in China.Aiming at the problem that the traditional target detection network model is difficult to deploy in industry due to the large number of parameters and low accuracy,the phenotype image dataset of tea diseases and pests is established,the network model is lightened,and the multi-knowledge transfer training model based on knowledge distillation is optimized.A YOLOv5 target detection network model based on the visual attention module(CSA)is constructed,and the tea disease and pest detection method is optimized.The experimental results show that the YOLOv5 target detection model with the visual attention module(CSA)constructed in this paper,compared with the YOLOv5 network model,and the YOLOv5 network model with the traditional attention modules SE and CBAM,respectively,improves the average accuracy by 3.1%,1.1%,and 1%.Compared with the pre-distilled student model,the model constructed in this paper achieves a maximum accuracy improvement of 4.1%.Compared with the teacher model,the model capacity is reduced by 5.4 MB,and the single-frame image inference time is reduced by 35%.The network model designed in this paper reduces the computational overhead without sacrificing accuracy and can provide an implantation possibility for resource-limited edge computing systems in the field of agricultural informatization.

关 键 词：茶叶病虫害 注意力模块 知识迁移 轻量化 农业信息化边缘计算 

分 类 号：S435.711[农业科学—农业昆虫与害虫防治] TP391.41[农业科学—植物保护]