Prediction Of Ponkan Freshness Based On Machine Learning And Its Application In Sensory Evaluation

Although citrus is one of the most popular healthy fruits in the world,there is a problem in the market that the freshness of citrus cannot be easily and accurately predicted,which leads to consumers who may buy citrus whose taste and nutritional value are lower than expected.In some past studies,electronic noses and Raman spectrometers have been used to detect citrus freshness,but these methods have drawbacks such as complex operation,high cost and low accuracy,resulting in a lack of market application potential.Therefore,there is an urgent need for a new citrus freshness detection method that can solve the above problems.In order to solve this problem,This paper used machine learning algorithms to study the freshness prediction method and related features of ponkan,and explored the sensory evaluation training process based on machine learning.The specific research contents and results are as follows:Based on machine vision and deep learning technology,a high-precision classification method for the freshness of ponkan was constructed.In the research of ponkan freshness detection method based on deep learning,70 ponkan were taken as the research object,and images of these ponkan in different periods were collected through self-built devices and digital microscopes,and the collected images were used to construct datasets respectively.Then,the images in the two datasets were preprocessed to remove the background using the image masking algorithm and the semantic segmentation algorithm respectively.In order to obtain the best freshness classification model for ponkan,five-fold cross-validation was used to compare the classification effects of the models based on different datasets,neural network models,input image pixel values and data enhancement methods.Finally,the data set of images taken by the self-built device,Resnet-18 network model,224×224 image size and geometric change data enhancement method were selected as the optimal combination,and the accuracy rate of the generated ponkan freshness classification model reached 97.1 %.In the feature analysis of ponkan freshness,by observing and analyzing the feature maps of each layer of the above optimal ponkan freshness classification model,hypotheses about the key features of ponkan freshness classification were put forward,and these hypotheses were verified by decision tree algorithm.Based on the feature map of the deep learning model,it was conjectured that the ponkan surface gloss,fruit stem and calyx have a larger classification weight in the early storage period of ponkan,while the peel color has a larger classification weight in the late storage period of ponkan.Semantic segmentation algorithm was used to segment the ponkan peel,calyx and peel regions in the image,and the ponkan surface gloss and the RGB values of the above three regions were extracted.Then,they were input as features into the CART algorithm to build decision trees of different depths,and finally 86.4%of the test set classification accuracy were obtained.Further,the ponkan storage process was divided into four stages from early to late,and the classification feature weights of the decision tree at each stage were output.The results showed that ponkan calyx color,stem color and gloss played an important role in the classification of the first and second stages of ponkan storage,and the sum of the weights was 83% and73%,respectively.The color of ponkan peel is the key classification feature of the third and fourth stages of storage,and the weight value reaches 88% and 89%,respectively.The above results validate the proposed hypothesis and demonstrate the effectiveness of the extracted features in ponkan freshness classification work.A machine learning-based sensory evaluation training method was constructed,and the verification and evaluation of the method were carried out.This study was carried out in the form of an online questionnaire.First,100 groups of ponkan freshness pictures were selected from the data-enhanced data set to generate ranking questions and imported into the created online questionnaire.Next,40 evaluators were randomly invited and divided equally into two groups.Then,one group of evaluators was trained on sensory evaluation,informing them of the features that characterize freshness at each storage stage,and the other group served as a control and did not participate in the training.Finally,the evaluators were invited to answer the questionnaire,and the results of the questionnaire were collected for specific analysis.The results showed that the average accuracy rate of the group after sensory evaluation training was 10.5% higher than that of the control group,confirming that the training process based on machine learning can significantly improve consumers’ ability to evaluate the freshness of tangerines.In conclusion,this paper proposes a ponkan freshness detection method based on machine vision and deep learning,which can complete the ponkan freshness detection work in a non-destructive,convenient,low-cost and high-precision manner.Furthermore,the validity of the proposed ponkan freshness classification feature was verified based on decision tree and sensory evaluation.These features can significantly improve consumers’ ability to predict the freshness of ponkan and its market value,and are critical to increasing their satisfaction.The results can provide reference for other citrus fruit freshness prediction research.