Research On Detection Method Of Colored Potato Internal Components Based On Spectral And Image Information Fusion

Colored potato has the function of both food and vegetable.It contains not only various amino acids and trace elements needed by human body,but also anthocyanins.The anthocyanin can resist oxidation,inhibit inflammation,improve vision,prevent cancer and other effects,so the color potato is deeply loved by consumers and has a good market prospect.However,at present,the detection of potato internal components mainly depends on chemical methods,these methods are affected by the operator’s operating experience and fatigue degree,and the detection accuracy is unstable,and the detection process is time-consuming and laborious.this thesis,hyperspectral and image technology will be used to integrate spectral and image technology,and deep learning algorithm will be introduced to explore the detection methods of anthocyanin,protein and starch content in potato.The main research contents are as follows:(1)The rapid detection methods of potato anthocyanins,starch and protein were studied.In this study,hyperspectral imaging technology was used to obtain the spectral and image information of the top,navel and middle part of two colored potatoes,Hongmei and Vajra,from 382nm-1004 nm,and SG,SNV and DET spectral preprocessing methods were used to eliminate the interference information.At the same time,CARS,VCPA,IRIV,GA,SPA and other methods were used to screen the characteristic spectral variables,and the prediction model of anthocyanin,starch and protein,which were the quality indicators,was established by linear partial least squares(PLS).The results showed that the model established by potato top spectrum was the best,and the best model for predicting starch was SPA-PLS,Rp=0.8454,RMSEP=0.786%,RPD=1.868;The best model to predict protein is IRIV-PLS,Rp=0.8832,RMSEP=0.145%,RPD=2.100;he best model for predicting anthocyanins is CARS-PLS,Rp=0.8918,RMSEP=0.232%,RPD=2.219;Using MCO function to constrain PLS algorithm,MCO-PLS prediction model of starch,protein and anthocyanin content was constructed,which improved the prediction accuracy of the model.The starch Rp=0.8633,RMSEP=0.625%,RPD=2.347;The protein Rp=0.9011,RMSEP=0.113%,RPD=2.696;The anthocyanin Rp=0.9116,RMSEP=0.184%,RPD=2.795;The visual distribution of starch,protein and anthocyanin content was established by using the best model.(2)Based on information fusion,the detection methods of anthocyanin,starch and protein in potato were optimized by using the image and spectral information of hyperspectral data.Gray level co-occurrence matrix GLCM and gray level statistical matrix GLGCM are used to extract the texture information of potato hyperspectral images,and a prediction model of starch,protein and anthocyanin based on PLS algorithm is established under the multi-level fusion of full texture,texture information characteristic variables and information.The final result shows that the middle-level fusion model based on GLCM is the best,In which starch predicts Rc=0.9083,RMSEC=0.613%,Rp=0.8988,RMSEP=0.647%,RPD=2.242;Protein predicted Rc=0.8711,RMSEC=0.151%,Rp=0.8833,RMSEP=0.138%,RPD=2.206;Predictive anthocyanin Rc=0.8981,RMSEC=0.224%,Rp=0.8976,RMSEP=0.222%,RPD=2.311.Aiming at the problem of insufficient directionality of GLCM algorithm,DTCWT algorithm is proposed to optimize texture extraction information.Constructed starch Rp=0.9139,RMSEP=0.578%,RPD=2.539;protein Rp=0.9011,RMSEP=0.110%,RPD=2.753;anthocyanin Rp=0.9112,RMSEP=0.184%,RPD=2.805.(3)Convolutional neural network model was established to predict the anthocyanin content under the microstructure of potato.The micro-hyperspectral technology was applied to make slices for the outer skin and inner skin of the potato internal structure and extract the spectral information at the micro level.SG,SNV,DET and their combination are used to denoise spectral data,At the same time,a variety of feature variable screening algorithms were used in combination with PLS and convolutional neural network CNN model to establish a regression prediction model for the anthocyanin content of potatoes.The results showed that the CNN model had the best prediction performance,and the optimal model was SG+DET+CNN model in the endothelial region,with Rc=0.9499,RMSEC=0.0359%,Rp=0.9439,RMSEP=0.2384%,and RPD=4.6516.(4)In order to further improve the prediction accuracy of CNN model,we proposed to use a variety of ways to improve and optimize the model,and the way of data fusion to quantitatively detect the potato anthocyanin under the microstructure.Based on the micro-spectral image information of the potato endothelial part,a novel method using Gaussian error linear element(GELU),Long and short-term memory(LSTM),Self-attention mechanism(SELF-ATTENTION),The results show that,the best model is CNN+GELU+LSTM+Self-Attention,Rc=0.9556,RMSEC=0.0385%,Rp=0.9514,RMSEP=0.174%,RPD=5.2436.A anthocyanin prediction model of CNN+GELU+LSTM+Self-Attention was established by color and spectral information fusion.Rc=0.9550,RMSEC=0.0345,Rp=0.9588,RMSEP=0.1725%,RPD=5.4085,And the de tection precision of the potato anthocyanin is further improved.