摘要
针对黄板诱捕的害虫体积小、数量多和分布不均匀,难以进行害虫识别的问题,引入当前主流目标检测模型Faster RCNN对黄板上的小菜蛾、黄曲条跳甲和烟粉虱等主要害虫进行识别与计数,提出一种基于改进Faster RCNN的田间黄板害虫检测算法(Mobile terminal pest Faster RCNN,MPF RCNN)。该算法将ResNet101网络与FPN网络相结合作为特征提取网络,并在RPN网络设计多种不同尺寸锚点对特征图像进行前景和背景判断,使用ROIAlign替代ROIPooling进行特征映射,以及使用双损失函数进行算法参数控制。对2440幅样本图像的实验分析表明,在真实复杂的自然环境下,MPF RCNN对烟粉虱、黄曲条跳甲、小菜蛾和其他大型害虫(体长大于5 mm)检测的平均精度分别为87.84%、86.94%、87.42%和86.38%;在35 cm×25 cm黄板上不超过480只的低密度下平均精度均值为93.41%,在480~960只害虫的中等密度下平均精度均值为89.76%。同时实验显示,在中低等密度下晴天和雨天的检测精度无明显差异,本算法计数结果与害虫计数决定系数为0.9255。将该算法置入以“微信小程序+云存储服务器+算法服务器”为架构的小米7手机终端系统中进行应用测试,平均识别时间为1.7 s。研究表明,该算法在精度和速度上均可支持当前便携式应用,为利用手机对蔬菜害虫进行快速监测与识别提供了技术支撑。
Realizing identification and counting of vegetable pests captured by yellow plates under complex conditions in the field is an essential prerequisite for targeted prevention and treatment pests and diseases of crop.Because of the small size,the large number and uneven distribution of pests trapped by yellow plates,it brings a great challenge to both manual and machine identification of pests.The current mainstream machine learning model Faster RCNN was introduced to identify and count the main pests such as diamondback moth,striped flea beetle and bemisia tabaci on the yellow plates.It also proposed a modified Faster RCNN pest detection algorithm(Mobile terminal pest Faster RCNN,MPF RCNN)based on Faster RCNN.This algorithm combined ResNet101 network with FPN network as a feature extraction network and designed a variety of different size anchor pairs in the RPN network to judge the foreground and background of features.This algorithm also adopted ROIAlign instead of ROIPooling for feature mapping and a dual loss function for algorithm parameter control.The experimental analysis of 2440 sample images showed that the average accuracy of MPF RCNN in the detection of bemisia tabaci,striped flea beetle,diamondback moth and other large pests(body length greater than 5 mm)in the realistic and complex natural environment were 87.84%,86.94%,87.42%and 86.38%,respectively.The average accuracy in the low density of 0~480 on 35 cm×25 cm yellow plate was 93.41%,and the mean accuracy in the case of the medium density of 480~960 was 89.76%.There was no significant difference between the detection accuracy in sunny and rainy days in medium and low density and the determination coefficient between the counting result of this algorithm and the insect count was 0.9255.Simultaneously,the average recognition time of the algorithm was 1.7 s when it was put into the Mi 7 mobile terminal system with the architecture of“WeChat applet+cloud storage server+algorithm server”for application test.The results showed that the present algori
作者
肖德琴
黄一桂
张远琴
刘又夫
林思聪
杨文涛
XIAO Deqin;HUANG Yigui;ZHANG Yuanqin;LIU Youfu;LIN Sicong;YANG Wentao(College of Mathematics and Informatics, South China Agricultural University, Guangzhou 510642, China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2021年第6期242-251,共10页
Transactions of the Chinese Society for Agricultural Machinery
基金
广州市科技计划项目(201904010196)
广东省重点领域研发计划项目(2019B020217003、2019B020214002)。
