Research On Cucumber Target Detection And Picking Location Based On Deep Learning

The planting area of cucumber in China ranks first in the world and occupies an important po sition in the development of agricultural economy.At present,the cucumber picking operation in China is basically completed by manual picking,which has high labor intensity,low production e fficiency and high production cost.Therefore,the research and development of picking robot with recognition and positioning function instead of manual picking has important practical significanc e.In view of the application requirements of cucumber picking robot in natural environment,this p roject proposed a method of cucumber target detection and picking location based on deep learning.The main content on this paper is as follows:1.Research on image processing methods based on cucumber target recognition.Through the field investigation of the layout and planting situation of the cucumber plantation,the collection ti me of the cucumber images and the shooting Angle and distance of the camera were determined.A total of 3000 cucumber images were collected.In the process of image preprocessing,RGB colo r space is selected as the color mode of the image after comparison.The Gaussian filter can not onl y effectively remove the noise,but also well preserve the characteristic details of the fruit,which is ideal for the study of this project.2.Considering that cucumber fruits and plants have similar color characteristics,it is diffic ult to recognize cucumber fruits in natural environment.Therefore,a deep learning based method is proposed.This paper briefly introduces the development process of deep learning target detectio n algorithm,analyzes the algorithm principle and framework of RCNN,Fast RCNN,Fathers RC NN and SSD,and summarizes their advantages and disadvantages.It is concluded that SSD has g ood comprehensive performance,and the detection accuracy and speed meet the requirements of t his topic.Finally,the single-order multi-layer detector SSD was selected as the recognition model of cucumber image.3.The network structure and working principle of SSD algorithm are introduced in detail fr om the algorithm level.Based on Windows platform,this paper builds a deep learning target Detec tion test environment of Tensor Flow and Tensor Flow Object Detection API framework.Labelimg tool is used to label the preprocessed images,and the training set and test set are established.Throu gh experiments,the learning rate,batch size,training steps and other super parameters were optim ized,and the learning rate was determined to be 0.001,the batch size was 12,and the training step s were 10000.Using the determined super parameters to train the model had the best effect.The e xperimental results showed that the average accuracy of the final model on the training set was 85.49%.The average detection time of each image was 0.96 s.4.Study on three-dimensional spatial localization method of cucumber picking point.The Bo uguet algorithm was used to calibrate the camera,and the internal and external parameters of the c amera were obtained.Based on the method of feature information matching,the SSD detection alg orithm was used to extract the feature points from the left and right images,and the matching accu racy was 76% in the cucumber matching experiment.Experimental results show that the relative e rror of localization is within 19 mm,which indicates that the binocular vision system can relativel y accurately locate the spatial position of cucumber.