| 基于YOLO v5和多源数据集的水稻主要害虫识别方法 |
| |
| 引用本文： | 梁勇,邱荣洲,李志鹏,陈世雄,张钟,赵健.基于YOLO v5和多源数据集的水稻主要害虫识别方法[J].农业机械学报,2022,53(7):250-258. |
| |
| 作者姓名： | 梁勇 邱荣洲 李志鹏 陈世雄 张钟 赵健 |
| |
| 作者单位： | 福建省农业科学院 |
| |
| 基金项目： | 福建省自然科学基金项目（2020J011376）、福建省农业高质量发展协同创新工程项目（XTCXGC2021015）、福建省智慧农业科技创新团队项目（CXTD2021013-1）和福建省农科院特色现代农业产业发展研究项目(AA2018-9) |
| |
| 摘 要： | 针对水稻稻纵卷叶螟和二化螟成虫图像识别中自动化程度较低的问题，引入目标检测算法YOLO v5对监测设备和诱捕器上的稻纵卷叶螟和二化螟成虫进行识别与计数。依据稻纵卷叶螟和二化螟的生物习性，采用自主研发的水稻害虫诱集与拍摄监测装置，自动获取稻纵卷叶螟和二化螟成虫图像，并与三角形诱捕器和虫情测报灯诱捕拍摄的稻纵卷叶螟和二化螟成虫图像共同构建水稻害虫图像数据集；采用左右翻转、增加对比度、上下翻转的方式增强图像数据集；对比了不同训练模型对三角形诱捕器和监测设备诱捕拍摄的水稻害虫图像的检测性能，并对比稻纵卷叶螟成虫不同训练样本量对识别结果的影响，用精确率、召回率、F1值、平均精度评估各模型的差异。测试结果表明，测试集图像为三角形诱捕器和监测设备诱捕拍摄虫害图像时，稻纵卷叶螟识别的精确率和召回率分别达到91.67%和98.30%,F1值达到94.87%,二化螟识别的精确率和召回率分别达到93.39%和98.48%,F1值达到95.87%。不同采样背景、设备构建的多源水稻害虫图像数据集可以提高模型对水稻害虫识别的准确性。基于YOLO v5算法设计的水稻害虫识别计数模型能够达到较高的识别准确率，可以用于...
|
| 关 键 词： | 水稻 稻纵卷叶螟 二化螟 深度学习 YOLO v5 自动识别 |
| 收稿时间： | 2022/2/11 0:00:00 |
Identification Method of Major Rice Pests Based on YOLO v5 and Multi-source Datasets |
| |
| LIANG Yong,QIU Rongzhou,LI Zhipeng,CHEN Shixiong,ZHANG Zhong,ZHAO Jian.Identification Method of Major Rice Pests Based on YOLO v5 and Multi-source Datasets[J].Transactions of the Chinese Society of Agricultural Machinery,2022,53(7):250-258. |
| |
| Authors: | LIANG Yong QIU Rongzhou LI Zhipeng CHEN Shixiong ZHANG Zhong ZHAO Jian |
| |
| Institution: | Fujian Academy of Agricultural Sciences |
| |
| Abstract: | Accurate prediction of rice pest plays a very important role in ensuring high yield of rice and reducing economic losses. Manual rice pest survey methods in paddy fields are time-consuming. With the developments in computer vision technology and theory, machine learning and deep learning have been applied in automatic identification of agricultural pests, which have greatly improved image classification accuracy of rice pest. Target detection algorithm YOLO v5 was introduced to identify and count Cnaphalocrocis medinalis and Chilo suppressalis on monitoring equipment and traps. Based on the biological habits of C.medinalis and C.suppressalis, a multi-source rice pest images dataset was constructed from the images of adults of C.medinalis and C.suppressalis , which were captured by a self-developed rice pest trap and photo device, triangular traps and light traps. The number of images in the dataset were increased by flipping the image left to right, increasing the image contrast, and flipping the image up and down. The detection performance of different training models on rice pest images captured by triangle traps and monitoring device was compared, and the effects of different training sample sizes on the identification results were compared. Precision, recall, F1 score and average precision were used to evaluate the difference of each model. The models were tested on the rice pest images captured by triangle traps and monitoring device. The results showed that the precision and recall of C.medinalis were 91.67% and 98.30%, respectively, and the F1 score was 94.87%. The precision and recall of C.suppressalis were 93.39% and 98.48%, respectively, and the F1 score was 95.87%. Multi-source rice pest images dataset constructed by different sampling background and equipment can improve the accuracy of rice pest by identification model. The rice pest identification and counting model developed based on YOLO v5 algorithm can achieve high identification accuracy and it can be used to monitor population of C.medinalis and C.suppressalis in the field. |
| |
| Keywords: | rice Cnaphalocrocis medinalis Chilo suppressalis deep learning YOLO v5 automatic identification |
|
| 点击此处可从《农业机械学报》浏览原始摘要信息 |
| 点击此处可从《农业机械学报》下载免费的PDF全文 |
|
