摘要
田间麦穗计数因主要依靠人工而存在耗时长、成本高等问题,为提高麦穗计数的效率和准确性,提出基于人群计数卷积神经网络的麦穗计数方法,在图像基础上进行麦穗数量自动化计数。试验改进了现有人群计数模型中的多列卷积神经网络MCNN和空洞卷积神经网络CSRNet,并对MCNN和CSRNet进行融合,建立了多列卷积神经网络MCSRNet。测试结果表明:MCSRNet网络对麦穗的预测准确率可以达到92.4%,较MCNN和CSRNet分别提高1.0%和4.0%,且训练迭代次数分别减少39次和5次。另基于独立数据集进行了测试,MCSRNet网络平均准确率为81.9%,较MCNN和CSRNet分别提高了0.4%和0.7%。MCSRNet的麦穗计数结果R2为0.80,优于MCNN和CSRNet。以上研究结果表明,MCSRNet网络有较高的麦穗计数准确率,同时有较快的训练速度,可为后续基于图像的小麦产量预测提供技术方法。
Nowadays, wheat spikes counting mainly depends on manual statistics, so there are problems of time consuming and high cost. To improve the efficiency and accuracy of wheat spikes counting, this paper proposes methods of automatic wheat spike counting with crowd counting convolution neural networks based on images. Firstly two original networks CSRNet and MCNN are improved, then MCSRNet is proposed by fusing CSRNet and MCNN. The results show that the prediction accuracy of MCSRNet is 92.4%, 1.0% and 4.0% higher than that of MCNN and CSRNet, and the number of training iterations can be reduced by 39 and 5 times, respectively. The test based on independent data sets shows that MCSRNet has the best average accuracy, which is 81.9%, 0.4% and 0.7% higher than that of the other two networks. R2 of MCSRNet for wheat spike counting is 0.80, which is better than that of MCNN and CSRNet. The above results indicate that MCSRNet has a high accuracy in improving wheat spikes counting and a fast training speed, which can provide a technical method for image-based wheat yield prediction in the future.
作者
刘云玲
张品戈
王千航
周睿琪
赵佳
肖永贵
马韫韬
LIU Yunling;ZHANG Pinge;WANG Qianhang;ZHOU Ruiqi;ZHAO Jia;XIAO Yong-gui;MA Yuntao(College of Information and Electrical Engineering,China Agricultural University,Beijing 100083,China;Institute of Crop Sciences,Chinese Academy of Agricultural Sciences,Beijing 100081,Chi-na;College of Land Science and Technology,China Agricultural University,Beijing 100193,Chi-na)
出处
《吉林农业大学学报》
CAS
CSCD
北大核心
2021年第2期171-180,共10页
Journal of Jilin Agricultural University
基金
国家重点研发计划项目(2016YFD0200700)。
关键词
卷积神经网络
机器视觉
图像识别
麦穗计数
密度图
convolutional neural network
machine vision
image recognition
wheat spikes counting
density map