Research On Non-destructive Determination Technology Of Fruit Sugar Degree By Visible/Near-Infrared Spectroscopy Based On Deep Learning

With the development of the big data on the Internet,deep learning is widely used in image recognition,text processing and speech processing due to its strong computing power and generalization ability.Therefore,in this thesis,a new algorithm of regression model of fruit sugar in visible/near infrared spectroscopy based on deep learning is proposed.The front part of the algorithm is based on Multilayer Percep,and the back is based on Convolution Neural Network,fusing principal component analysis and two-dimensional correlation.The algorithm of regression model uses Adam optimizer to optimizing the network.The experiment takes the oranges of Southern Jiangxi and the fragrant pears of Xinjiang as the research object,and the main detection index with their sugar.In this thesis,the experiment make compared with the new algorithm of regression model and the traditional Partial Least Square Regression,Principal Component Regression,Multilayer Perceptron,Convolution Neural Network regression model.It mainly focuses on the new algorithm of regression model of fruit sugar in visible/near infrared spectroscopy with nondestructive testing?The main works include:(1)A new algorithm of regression model of fruit sugar based on depth learning is proposed in this thesis.The experiment shows that the model(MLP-CNN)is better than the traditional Partial Least Squares Regression model and the Principal Component Regression model after the pre processing and the feature band screening.In the process of predicting the sugar degree of oranges,compared with the PLSR model,this new algorithm reduces the root mean square root error from 0.61 ° Brix to 0.46° Brix,and the determination coefficient is increased from 0.79 to 0.85.In the process of predicting the sugar degree of pears,compared with the PLSR model,this new algorithm reduces the root mean square root error from 0.34° Brix to 0.29° Brix,and the determination coefficient is increased from 0.88 to 0.93.(2)A static fruit spectrum collection experiment device is setted up.It include the selection of the light source,the spectrometer,the link of the circuit and the work of the fruit calibration.In the same time,the experiment complete the measurement of the fruit sugar degree.(3)The fruit sugar measurement software based on deep learning is developed,which realized the basic functions of fruit spectrum input,output,display,storage,preprocessing and regression.Compared with the traditional regression algorithm,MLP-CNN regression model has simple operation and better prediction ability.At the same time,compared with the simple Convolution Neural Network and Mmulti-layer Perceptron model,the MLP-CNN regression model's convergence speed is faster,the self-learning level is higher and the prediction ability of the model is stronger.The MLP-CNN not only ensures the integrity of the samples,but also expands the area of the application with deep learning.It promotes the development of nondestructive testing technology,and make the nondestructive,accurate and rapid measurement of the fruit sugar degree come true.