Research On Application Of Multispectral Imaging Technology In Moisture Detection Of Dried Fruits And Vegetables

Fruits and vegetables are favored by people who are pursuing a green and healthy life because of their rich nutrients and low fat content.However,fruits and vegetables are perishable.This will lead to huge losses in the storage process and food safety problems.Drying technology has become the focus of modern fruit and vegetable processing because it can effectively inhibit decay.Moisture content is an important index to measure the drying effect,which can reflect the changes of food storage and flavor quality.Using traditional physical methods to detect water content will cause destructive effects on samples,which will increase the detection cost and require longer processing time.Multispectral image detection technology is a new rapid nondestructive testing technology in recent years.It can obtain multi-dimensional information such as spectral characteristics and spatial characteristics of samples by taking multi-spectral images,and can be used to construct the relationship with water content in the drying process,and is widely used in water content detection of agricultural products.In this study,the representative carrots in fruits and vegetables are selected as the main research object.Based on the background of multispectral image detection technology,this paper studies a set of fast,nondestructive and accurate water content prediction model suitable for Assembly line production.Based on the characteristics of band correlation,high dimensional characteristics and vulnerability to multiple factors in the application of multispectral images,the model will explore the introduction,improvement and migration of dynamic network model,and finally realize the detection of moisture content of carrot slices in the drying process.The specific research contents are as follows:1.A long-short term memory(LSTM)using multispectral band correlation is studied.The model uses multispectral image information in the range of 675-975 nm to predict the water content of carrot slices during drying.Twelve features including spectral features and spatial features in each multi-spectral band were selected and input into the LSTM model according to the band sequence to achieve the prediction of water content.The control model group includes partial least squares regression(PLSR),and least squares support vector machine(LS-SVM).The results showed that LSTM model had the best effect,and the highest indexes were Rp=0.956,RMSEP=8.4(%),RPD=3.5.It is also proved that the dynamic network model can effectively learn the band correlation information of multispectral images and realize the moisture content detection of dried carrot slices with higher accuracy.2.From the perspective of model improvement,the model can automatically extract and process the high-dimensional information of multispectral images,and a convolutional-long short term memory(C-LSTM)is used to predict the moisture content of dried carrot slices with the original multispectral images as input.C-LSTM deeply integrates convolution kernel structure into LSTM unit structure,so that the model can automatically extract image information.In addition,the depth of the model is increased to ensure that the original high-dimensional information can be fully utilized.The results show that the prediction accuracy of C-LSTM is higher than that of LSTM and PLSR and LS-SVM model after feature selection by continuous projection algorithm(SPA)without using prior knowledge,Besides,The highest indexes were Rp=0.964,RMSEP=8.0(%),RPD=3.7.and the model simplifies the experimental process and increases the real-time detection.3.Model updating is studied from the perspective of model migration.In this paper,a model migration method based on self-coding network was used to predict the water content of carrot slices in the target set by using the C-LSTM trained by a large number of samples.Based on the freezing fine tuning principle,this transfer learning method realizes the information compression of the high-dimensional features of the frozen layer by introducing the self coding network and constructs the relationship between the features after taking off and landing dimensions and the water content through supervised learning in the terminal full connection layer.The results show that when the proportion of labeled samples is 10%,the highest index can reach: RP = 0.928,RMSEP =10.7(%),RPD =2.6.And the fluctuation range of multiple detection accuracy is reduced.In the experiment of migrating from the dried carrot slice data set prepared based on multispectral technology to the dried soya bean data set prepared based on hyperspectral technology,the method achieved the following performance indicators at the label ratio of 30%: RP = 0.903,RMSEP =8.5(%),RPD =2.3.This method is a effective way to measure the water content of carrot slices with small sample.It also provides a new idea for the migration of data sets of different kinds of dried fruits and vegetables prepared by different spectral techniques.