Study On Wheat Yield Estimation Based On Deep Learning Ear Recognition

Wheat ear counting is an important part of wheat yield estimation in field wheat.Accurate and rapid ear detection and identification can greatly improve the efficiency of wheat count and yield estimation.The traditional counting methods of wheat ears are mostly manual counting in the field,which is time-consuming and labor-intensive,poor in accuracy and low in efficiency.With the rapid development of computer technology and image technology,it is possible to count wheat head intelligently based on deep learning.Under this background,the field image count and wheat yield estimation problem,this paper carried on the thorough research,put forward a set of wheat grain count and yield estimation based on the technology of deep learning system,improved the wheat grain yield estimation link count and the accuracy and efficiency,reduces the labor cost,provide reference for the automated management of wheat field.The main work of this paper is as follows:（1）A wheat ear detection network PPYOLO-SE was proposed,which paid attention to the balance between detection accuracy and detection speed,and added channel attention module on the basis of the original PPYOLO network to improve the learning ability of wheat ear characteristics.In order to verify the effectiveness of the PPYOLO-SE model,the original network and several common detection network training models were used for comparative analysis on the same data set.The detection accuracy of the PPYOLO-SE model reached95.75%,which was better than the original PPYOLO network,SSD,YOLOv3 and FasterRCNN.The detection speed of each image is 0.6s,and the speed and accuracy of the network are high,which meets the requirements of real-time use.（2）In this paper,a wheat head counting model based on deep regression was constructed,and wheat head counting in field image was realized by using Tassel Net V2+ network training on wheat field data set.The experiment verified that the average detection accuracy of the model on the field wheat ear data set reached 95.13%,and the relative error between the counting result and the actual number of wheat ear was about 7.17%,indicating that the counting accuracy of the model was high.The relative error of PPYOLO-SE model is about6.31%,and the recognition and counting effect of PPYOLO-SE model is close to Tassel Net V2+model.（3）The estimation model of wheat ear yield in field was established.The PPYOLO-SE network was used to calculate the mean density per unit area of each field.The regression analysis was made between the data and the actual wheat yield of the field,and the regression model between the wheat yield and the number of ears per unit area of the field was established.Finally,the relative error between the actual wheat yield and the model is 2.30%,which is relatively small.（4）A wheat ear counting and yield estimation system for field wheat was developed.In order to facilitate the counting and yield estimation of wheat ears in field wheat,a mobile application of Wheat assistant in Android system was designed and developed by using Java language,which realized the function of real-time counting results and yield estimation by uploading images of wheat ears in field and added the common functions of wheat encyclopedia.