Research On Quality Detection Method Of Light Post-harvest Mango Based On 3D Structured System

As a popular fruit,mango is loved by consumers.In recent years,mango industry has developed rapidly,however,the post-harvest sorting stage is mainly done manually,which is not only inefficient but also susceptible to subjective and environmental factors.Detection of machine vision is not subjective and has been widely used in agriculture.However,most methods of mango detection using machine vision are two-dimensional,which have many limitations.Therefore,this paper uses mango as experimental material,building a mango visual detection system based on 3D structured light system.Detecting property of post-harvest mango by extracting two-dimensional(2D)image information and three-dimensional(3D)points cloud information of its surface.Main research contents are as follows:(1)A mango visual detection system based on 3D structured light system is built,which can collect 2D images and 3D points cloud information online.After tilt correction of points cloud,the maximum error of distance measurement is within 1%,and the R~2 of real length calibration on two coordinate axes using vision system reaches 0.998.At the same time,coordinates of height matrix and image coordinates are matched,the R~2 of coordinate matching on x-axis and y-axis both reach more than 0.996.(2)The problem that two different postures of mango on assembly line is studied,2D and3D features are extracted to detect the posture.When 3D information has not been introduced,using first three principal components obtain after dimensionality reduction by principal component analysis to detect the posture.The identification accuracy of the"lying"posture is95%,and the identification accuracy of the"upright"posture is 90%,the recognition error is caused by the shape difference between the samples.Then,3D information is introduced.At this point,only two principal components can explain close to 90%of the variance,the accuracy of posture identification using first two principal components reaches 100%.(3)A non-contact detection method of shape grade,volume,weight and external grade in two postures’mango is studied.For identification of mango shape grade,part of 2D and 3D features which are common and exclusive to the two postures are extracted.Mango shape grade recognition models of corresponding postures are established,then,using prediction set to identify theirs accuracy.The accuracy rate of shape grade identification in"lying"posture is95%,and that in"upright"posture is 90%.For volume prediction of mango,pixel projection areas of mango and height information of its surface are extracted,different volume prediction models of mango are established under two postures respectively.Using 2D and 3D information to build models and predict,the average error rate under two postures is about 4%,which is much smaller than that of 2D modeling method,and there will be no posture recognition errors which leading to wrong model use.For weight prediction of mango,the method of fixing density and modeling analysis are adopted.The average error rate of prediction sets using two methods are less than 5%.Finally,using shape grade,volume and weight which predicted and identified above,combine with surface defects and color information,a mango external grade recognition model based on random forest algorithm is established and prediction set is used to identify its accuracy.Mango external grade identification accuracy rate reaches 95%.(4)The relationship between external color and internal property of mango is studied.Color information of different parts of mango is extracted to establish a prediction model of mango’s brix,acidity and fruit firmness,then using prediction set to predict.Established BP-neural network model achieves good results,predicted Rp~2 for these three target values are all above 0.83.A stepwise multiple regression model is established on postharvest storage time using mango’s brix,acidity and fruit firmness.After selecting significant variables and removing a insignificant variable,prediction set is predicted according to the built model,the predicted Rp~2 is 0.858,RMSEP is 1.205,and the accuracy of classifying prediction storage time reaches 85%.The taste of mango is divided into three grades,a mango taste grade recognition model based on random forest algorithm is established,and prediction set is used to identify its accuracy.Identification accuracy reaches 85%.