基于深度卷积神经网络的田间麦穗密度估计及计数

Estimation and counting of wheat ears density in field based on deep convolutional neural network

小麦的最终产量可由单位面积的小麦麦穗数侧面反映,为了快速准确统计小麦麦穗数,该研究给出一种在单幅图像上利用深度卷积神经网络估计田间麦穗密度图并进行麦穗计数的方法。首先对采集的田间小麦图像进行直方图均衡化及阈值分割预处理,以减少图像中光照及一些复杂背景对计数的影响;然后根据灌浆期田间小麦图像麦穗密集的特点,引入拥挤场景识别网络(Congested Scene Recognition Network,CSRNet)构建麦穗密度图估计模型,并采用迁移学习方法,利用小麦图像公开数据集对模型进行预训练,再用所采集的小麦图像数据集进行模型参数调整和优化;利用得到的模型生成单幅小麦图像的麦穗密度图,根据密度图中所有密度值的总和对图像进行麦穗计数。最后根据对单幅麦穗图像的试验数据,构建田间麦穗计数函数模型,实现田间小麦麦穗数估计。通过对所采集的安农170、苏麦188、乐麦608和宁麦24这4个品种共296幅小麦图像进行试验,平均绝对误差(Mean Absolute Error,MAE)和均方根误差(Root Mean Squared Error,RMSE)分别为16.44和17.89,4个品种小麦的麦穗计数值与真实值的决定系数R^2均在0.9左右,表明该方法对单幅图像小麦麦穗计数精度较高。此外,通过对田间小麦麦穗数进行估计试验,结果表明,随面积的增大麦穗估计的误差越小,研究结果可以为小麦的产量自动估计提供参考。

Wheat is one of the most important grain crops in the world.The stability of wheat yield is crucial to global food security.The final yield of wheat can be calculated by the number of ears per unit area.Using deep learning technology to accurately and automatically count the number of wheat ears can save a lot of manpower and material resources.The images of wheat ears from four different wheat varieties during grain filling period were collected,and 296 images with a total of 14964 ears of wheat were selected to construct the WEE data set.The method of point labeling was used to label the wheat ears in the images.In order to quickly and accurately count the number of wheat ears in complex crowded scenes,a method of estimating the density map and counting the number of wheat ears in field wheat images was proposed.Firstly,histogram equalization and threshold segmentation were used to preprocess the collected field wheat images to reduce the influence of light and some complex backgrounds on counting.Then,according to the characteristics of dense wheat growth at grain filling period,the Congested Scene Recognition Network(CSRNet)was introduced to construct the wheat ear density map estimation model.The CSRNet network was composed of front-end and a back-end networks.The front-end network uses the pre-training model VGG16 to extract the features of the wheat images,and the back-end network uses the dilation convolution to generate the distribution density map while expanding the receptive field.In order to improve the accuracy of the model in the training process,the transfer learning method was used to pre-train the model by using the public wheat image data set,and the collected wheat image data set was used to adjust and optimize the model parameters.The trained model was used to generate the wheat ear density map of a single wheat image,and the wheat ears were counted according to the sum of all density values in the density map.Finally,according to the test data of a single wheat ear image,a wheat ear counting function was constructed to estimate the number of wheat ears in the field.Experiments were conducted on a total of 296 wheat images collected from four varieties(Annong170,Sumai 188,Lemai 608 and Ningmai 24).The Mean Absolute Error(MAE)and Root Mean Square Error(RMSE)were 16.44 and 17.89,respectively.The correlation coefficient R^2 between the ear count values and the true values of the four varieties of wheat was about 0.9,indicating that the method in this paper has a higher accuracy for counting wheat ears in a single image.In addition,the experiment of estimating the number of wheat ears in the field showed that the error of estimation of wheat ears was smaller with the increase of area.The results of this study can provide the possibility of automatic estimation of the number of wheat ears in practical application process,and also provide a reference for wheat yield estimation.