基于改进YOLOv4算法的番茄叶部病害识别方法  

Identification method of tomato leaf diseases based on improved YOLOv4 algorithm

作　　者：储鑫 李祥[1] 罗斌[2,3] 王晓冬 黄硕[2,3] CHU Xin;LI Xiang;LUO Bin;WANG Xiao-dong;HUANG Shuo(College of Information Engineering,East China University of Technology,Nanchang 330013,China;National Engineering Technology Research Center for Agricultural Informatization,Beijing 100097,China;Research Center of Intelligent Equipment Technology,Beijing Academy of Agriculture and Forestry,Beijing 100097,China)

机构地区：[1]东华理工大学信息工程学院,江西南昌330013 [2]国家农业信息化工程技术研究中心,北京100097 [3]北京市农林科学院智能装备技术研究中心,北京100097

出　　处：《江苏农业学报》2023年第5期1199-1208,共10页Jiangsu Journal of Agricultural Sciences

基　　金：江苏省科技计划重点及面上项目(BE2021379);江西省核地学数据科学与系统工程技术研究中心开放基金项目(JETRCNGDSS201801)。

摘　　要：为快速准确识别自然环境下的番茄叶片病害,提出一种基于改进YOLOv4算法的轻量化番茄叶部病害识别方法。该方法根据番茄病害特征采用K均值聚类算法调整先验框的维度,并使用宽度因子为0.25的MobileNetv1代替YOLOv4原有的主干网络CSPDarknet53进行特征提取,并在特征融合网络PANet中引入深度可分离卷积代替原有的3×3标准卷积,同时在主干网络的2个输出特征层和空间金字塔池化输出层分别嵌入卷积块注意力模块(CBAM),提高模型识别精度。试验结果表明,改进后的模型对8类番茄叶片整体检测精准性(mAP)为98.76%,参数量为12.64 M,传输帧数为1 s 101.76帧,相较于原YOLOv4模型,模型参数量减少80%,每秒传输帧数比原始YOLOv4模型提高了130%。In order to identify tomato leaf diseases in natural environment quickly and accurately,a lightweight tomato leaf disease identification method based on improved YOLOv4 algorithm was proposed.The method used K-means clustering algorithm to adjust the dimensions of the prior box according to the characteristics of tomato disease,and used MobileNetv1 with a width factor of 0.25 instead of the original backbone network CSPDarknet53 of YOLOv4 for feature extraction,and introduced deep separable convolution in place of the original 3×3 standard convolution in the feature fusion network PANet.At the same time,the convolutional block attention module(CBAM)was embedded in the two output feature layers and the spatial pyramid pooling output layer of the backbone network to improve the model recognition accuracy.The results showed that,the overall detection accuracy(mAP)of the improved model for eight types of tomato leaves was 98.76%,the parameter quantity was 12.64 M,and the transmission frame number was 101.76 f/s,which was 80%lower than that of the original YOLOv4 model,and the number of transmitted frames per second was 130%higher than that of the original YOLOv4 model.

关 键 词：YOLOv4 MobileNet 轻量化 注意力机制 病害 

分 类 号：S436.412[农业科学—农业昆虫与害虫防治]