Development Of Corn Kernel Selection System Based On Convolutional Neural Network

This paper is based on projects funded by the National Natural Science Foundation of China(project number: 51405078),Heilongjiang Returning Overseas Studies Fund(project number: LC2018019),Northeast Agricultural University academic backbone project(project number: 17XG01).The aim of this paper is to design and optimize the corn kernel selection system by using electromagnetic vibration theory,automatic control technol ogy and deep learning technology,so as to achieve the purpose of mass selection and grading of corn kernel quality.Based on the above,the main work of this paper is as follows:(1)The analysis and design of the speed difference integrative separation device,control and vision system of the test prototype.In order to realize the gradual separation of corn kernels,a speed differential electromagnetic vibration transmission and separation device was designed based on the theory of electromagnetic vibration.Orthogonal experiments were carried out with material tank counterweight,installation inclination and system amplitude as factors,adhesion rate and positive rate as indicators.Through orthogonal experiments,the optimum parameters of the system were obtained: counterweight,installation inclination and system amplitude were 0.3 kg,0 degree and 0.36 mm,respectively.The adhesion rate and orientation rate of mechanical system were 9.40% and 92.60% respectively,which could make corn kernels group compete under the drive of electromagnetic vibration system.Gradual transformation and segregation are the transportation of single kernel.Combining with automatic control technology and image acquisition principle,the working principle of the mechanical,control and constant light intensity vision system of the test prototype of corn kernel quality selection and detection is analyzed,and the design of key components and corresponding software of each system is completed,which provides a platform for subsequent sorting work.(2)Overall design of corn kernel data set,image acquisition and preprocessing scheme and data set calibration.Firstly,the overall design of corn kernel data set is completed according to previous research,relevant technical requirements and specifications,which provides design basis for the training and validation of convolutional neural network and the research and evaluation of relevant selected detection data sets.Based on image segmentation and morphological image processing technology,a pre-processing scheme suitable for corn kernel segmentation and denoising in dark box environment is designed.Under the influence of edge and noise,the image of corn kernel data set with the highest clarity can be obtained.The data set calibration of corn kernel image after pre-processing scheme was carried out by programming.Finally,four types of kernels were identified in the data set,626 kernels for each type of excellent corn kernel and 480 kernels for each type of removed kernel,and the image size was 352 × 352 pixels.(3)Analysis and design of convolutional neural network model for corn kernel quality detection.Firstly,the network model,network structure and network training principle of convolutional neural network are analyzed comprehensively;secondly,the framework of convolutional neural network model for corn kernel quality selection is designed;then,on the basis of in-depth study of Faster R-CNN,the size of the maximum convolutional neural network model that can run under limited hardware conditions is determined through experiments,aiming at corn kernel quality.A network model of corn kernel quality selection is designed,which can directly use color image as input: model 1(S),model 2(M),model 3.0(L).Finally,considering the particularity of previous studies and corn kernel quality selection based on convolutional neural network,the evaluation criteria for this paper are formulated.(4)Design and construction of the experimental prototype of corn kernel selection system.In order to realize the automatic sorting of corn kernel quality,the software and hardware system of the prototype was built based on the above research and the functional design requirements of each part of the prototype.The structural para meters of the key components of the prototype were optimized through experiments,so that the prototype could achieve the goal of optimum sorting performance.The results show that the experimental prototype can effectively realize the automatic sorting of corn kernel quality problems.Compared with the traditional recognition method of extracting corn kernel characteristic parameters by artificial feature modeling,this study only needs a pre-labeled data set,which can be used to train the network model with limited corn kernel data set,so as to solve the problems of tedious modeling process,small number of features and insufficient ability of feature expression in the past corn kernel selection artificial feature modeling.This study explored the application of deep learning technology in real-time sorting of corn kernels,and provided theoretical basis for the application of deep learning in corn sorting and other agricultural fields.Comparing and analyzing the classification performance of the four ne twork models in this paper,the results show that the overall classification performance of model 2(M)is the best,with the highest average accuracy rate of 91.21%,average recall rate of 9 1.37% and average F1 value of 91.24%,and model 2(M)has the lowest false positive rate of 2.60% in the four models.Compared with Faster R-CNN,the performance of the network model is improved.According to the classification performance of good corn kernels only,model 1(S)has the highest average recall rate of 98.41% in the four models,model 2(M)has the highest average precision rate and average F1 value of 98.77% and 96.97% in the four models,and model 2(M)has the lowest false positive rate of 0.47% in the four models,which is improved in different aspects compa red with Faster R-CNN two network models.By comparing and analyzing the overall performance of the four network models,the results show that the detection model in this study is superior to Faster R-CNN model.In the four models,model 1(S)has the highest m AP1 value of 97.73%,which is 2.18% higher than Faster R-CNN,and lower requirement for computer hardware;model 2(M)has only 0.42% lower than Faster R-CNN.Compared with Faster R-CNN,model 1(S),model 2(M)and model 3.0(L)are 138.25 MB,136.04 MB and 75.5 MB respectively.Because of the huge size of Faster R-CNN network model,it is too demanding for the existing computers,which is not conducive to the application in the agricultural field.By comparing and analyzing model 1(S),model 2(M)and model 3.0(L),it is found that the recall-accuracy curve of model 2(M)is the most stable of these models,and its overall performance is the most robust and ideal.Finally,through the joint debugging of the mechanical,control and visual system of the experimental prototype and the performance test of the prototype,the results show that the actual detection accuracy of corn kernels is 96.8%,and the actual sorting rate is 98.14%.