基于改进YOLO v5的夜间温室番茄果实快速识别  

Fast Recognition of Tomato Fruit in Greenhouse at Night Based on Improved YOLO v5

作　　者：何斌[1,2] 张亦博 龚健林 付国 赵昱权[1] 吴若丁 HE Bin;ZHANG Yibo;GONG Jianlin;FU Guo;ZHAO Yuquan;WU Ruoding(College of Water Resources and Architectural Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;Key Laboratory of Agricultural Soil and Water Engineering in Arid and Semiarid Areas,Ministry of Education,Northwest A&F University,Yangling,Shaanxi 712100,China)

机构地区：[1]西北农林科技大学水利与建筑工程学院,陕西杨凌712100 [2]西北农林科技大学旱区农业水土工程教育部重点实验室,陕西杨凌712100

出　　处：《农业机械学报》2022年第5期201-208,共8页Transactions of the Chinese Society for Agricultural Machinery

基　　金：陕西省科技创新引导专项(2021QFY0801)。

摘　　要：为实现日光温室夜间环境下采摘机器人正常工作以及番茄快速识别,提出一种基于改进YOLO v5的夜间番茄果实的识别方法。采集夜间环境下番茄图像2000幅作为训练样本,通过建立一种基于交并比的CIOU目标位置损失函数,对原损失函数进行改进,根据计算函数anchor生成自适应锚定框,确定最佳锚定框尺寸,构建改进型YOLO v5网络模型。试验结果表明,改进YOLO v5网络模型对夜间环境下番茄绿色果实识别精度、红色果实识别精度、综合平均识别精度分别为96.2%、97.6%和96.8%,对比CNN卷积网络模型及YOLO v5模型,提高了被遮挡特征物与暗光下特征物的识别精度,改善了模型鲁棒性。将改进YOLO v5网络模型通过编译将训练结果写入安卓系统制作快速检测应用软件,验证了模型对夜间环境下番茄果实识别的可靠性与准确性,可为番茄实时检测系统的相关研究提供参考。In order to realize the normal operation of the picking robot and the rapid recognition of tomato in the nighttime environment of solar greenhouse,a nighttime tomato fruit detection method based on improved YOLO v5(You only look once)was proposed.Totally 2000 tomato images in the night environment were collected as the initial training samples,and the original loss function was improved by establishing a CIOU target position loss function based on intersection and union ratio,and then an adaptive anchor frame was generated according to the anchor calculation function,the optimal anchor frame size was determined,the network structure was optimized,and an improved YOLO v5 network model was constructed,and the recognition rate of tomato fruit in night environment was improved.The experimental results showed that the average recognition accuracy of improved YOLO v5 network model for tomato green and red fruits and average recognition accuracy in night environment was 96.2%,97.6%and 96.8%.Compared with traditional CNN convolution network model and traditional YOLO v5 model,the recognition accuracy of occluded features and features in dark light was improved and the robustness of the model was improved.The improved YOLO v5 network model compiled and wrote the training results into Android system to make a rapid detection application software,which verified the reliability and accuracy of the model for tomato fruit recognition in night environment,and provided a reference for the relevant research of tomato real-time detection system.

关 键 词：番茄果实 温室 识别 改进YOLO v5 夜间 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]