| As the demand for tomatoes increases in both domestic and international markets,the quality of tomatoes has become a growing concern.This requires the grading of picked tomatoes to meet the needs of different consumers and to meet the development trend of consumer diversification.Therefore,it is of great practical importance to develop a non-destructive inspection and grading system for the internal and external quality of tomatoes.In this study,based on the demand of tomato nondestructive inspection and grading,we use machine vision technology to detect the external quality of tomatoes and combine with NIR spectroscopy technology to detect the internal quality,and develop an online tomato inspection and grading system to realize comprehensive inspection and grading of internal and external quality of tomatoes,so as to improve the efficiency and accuracy of tomato grading.The main research contents and conclusions of this thesis are as follows:(1)Overall design of online tomato internal and external quality inspection system.Based on the research and relevant standards,the internal and external quality grading indexes and standards of tomatoes were determined.The online inspection and grading system is defined to consist of five parts:information collection mechanism,conveying mechanism,upper computer grading system,control system,and sorting mechanism.The scheme design,hardware selection and installation of the information collection mechanism are carried out.According to the tomato online inspection process,the control system is designed and the models of hardware such as programmable logic controller,server and servo motor are determined,and the control program is written in ladder diagram language on this basis.Develop the spectral online acquisition program to realize continuous acquisition of tomato equatorial surface spectral data for averaging and saving.(2)Research on tomato external quality inspection model based on machine vision technology.The tomato images are collected,the tomato images are labeled using Label Img open source software,and the training set samples are expanded,the number of training set,validation set,and test set are 5550,618,and 618,based on YOLOv7 network model,the model is improved by adding decoupling head and introducing C2f module,and the training set images are used in the same batch of seven commonly used deep learning models The improved YOLOv7 model was trained to detect tomato external defects at a frame rate of 40frames with an average accuracy of 94.7%.In the detection results of YOLOv7 adaptive threshold segmentation was performed on the color channel based A-channel of tomatoes,and the fruit shape index and the pixel size of tomato fruit diameter were calculated on the segmented mask image,and the actual fruit diameter of tomatoes was calculated by the small-aperture imaging model,and the average absolute error of fruit diameter calculation was 1.9mm by experiment.(3)Study of tomato internal quality detection model based on visible/near-infrared spectroscopy.Firstly,160 tomatoes of Golden Roc varieties were used as experimental samples to compare the effects of two feature extraction methods,Competitive adaptive reweighted sampling(CARS)and Successive projections algorithm(SPA),on The effects of the two modeling methods,Partial least squares regression(PLSR)and Support vector regression(SVR),on the modeling accuracy.It was concluded that the model built using a combination of SNV preprocessing algorithm plus CARS feature extraction algorithm plus PLSR modeling method was the best for predicting tomato soluble solids and hardness,and the correlation coefficient R_p and root mean squared error of prediciton(RMSEP)for the prediction set of tomato soluble solids and hardness,respectively,were In order to improve the applicability of the detection model,two additional tomatoes,Provence and Toyota,were added to participate in the validation and optimization of the model,and due to the limited applicability of the model established by a single variety,a global model was established by mixing three varieties of tomatoes,and the global model was found to have the best prediction set of soluble solids The set correlation coefficient R_p was 0.859 and the root mean square error RMSEP was 0.322°Brix;the set correlation coefficient R_p was 0.877 and the root mean square error RMSEP was 0.643 kg/cm2 for the prediction of tomato hardness.(4)Grading system construction and system accuracy verification test.The established tomato internal quality detection model and external quality detection model were ported to run in the same environment of the upper computer to develop a human-computer interaction interface based on PyQt5.The system accuracy validation test was conducted on 100 tomatoes under the optimal speed conditions,and the accuracy rates were 94.9%,87.3%,and 88.5%for external quality classification,internal quality classification,and fusion quality classification,respectively,with an efficiency of 19 tomatoes per minute.The system can be used for online inspection and grading of tomatoes. |
