| Jujube,Chinese characteristic agricultural product,is popular with consumers for its health benefits and sweet flavor.China has always been a great producer of jujube,but the improvement of additional value of jujube and the increase of export are not so good.An important reason is that the classification and grading of jujube are limited by current techniques.At present,the classification and grading methods of jujube in the market mainly include artificial method,mechanical method and method based on machine vision.There are weaknesses such as high cost,low efficiency,easy to get broken by the machine and grading only by single characteristic,which affect the further development of jujube industry to some extent.The research in this paper further explores the external quality classification and grading methods for jujube based on machine vision.The main research work is as follows:(1)Analyze GB/T 5835-2009 and standards for dried red jujubes in related literature,and take the external quality characteristics of the experimental objects selected into consideration to develop a set of grading standards which is feasible in actual application.Build a machine vision jujube image acquisition system,and collect multi-view of images for each jujube.In addition,manually label the jujube images collected,and prepare data sets for jujube classification research based on convolutional neural network(CNN).(2)Combine the characteristics of deep learning algorithms and classification requirements,the convolutional neural network algorithm is chosen as the model for jujube classification.Through classification experiment in four different networks,including VGGNet-19,ResNet-18,ResNet-50,DenseNet-121,the experiment results show that the classification accuracy and speed of the four networks are satisfactory,and the classification result of ResNet-18 is the best.Thus,ResNet-18 is selected as the backbone network and it is connected with identification sub-network and verification sub-network.(3)Explore the method of three-dimensional fitting based on multi-view images of jujubes to realize jujube grading by size.After a series of image preprocessing and edge extraction on the jujube images collected from four different perspectives,the least squares ellipse fitting method is used to fitting the jujube.With the complementary advantages of the four perspective images,the accuracy of the extracted semi-axises of the fitting ellipse is optimized.The volume of the fitting ellipse is adopted as the grading basis.Compared with the jujube grading by least squares ellipse fitting of single two-dimensional jujube images,the experimental results show that the size grading method based on three-dimensional fitting of multi-view red jujube images has obvious advantages in grading accuracy.Based on the deep learning theory framework,this paper constructs jujube classification algorithm,further explores the machine vision classification method,solves the technical problem of low classification and recognition accuracy of the defective jujube,improves the accuracy of jujube size classification,and meets the requirements of agricultural modernization development.At the same time,it will also help increase the additional value and market competitiveness of jujubes as well as promote the export and sales of jujubes. |
