Research On Citrus Quality Identification Based On Machine Vision

Rapid detection and classification of citrus quality is the key technology in circulation and processing of citrus.Accuracy and speed of quality detection directly affect the economic value of citrus.Traditional quality detection method relies heavily on manual work and is easy to cause damages to citrus.With the maturity of machine vision technology and the reduction of cost,it provides a new method for fast,accurate and non-destructive citrus quality detection.In this study,machine vision technology,machine learning and deep learning methods are used to realize rapid and non-destructive detection and classification of citrus.The characteristics of a variety of classical models are comprehensively compared and analyzed,and a convolution structure combining depth separable,multi-scale and attention mechanism is proposed as feature extractor to fuse the shallow,deep and multi-scale features of citrus.SVM is used as a lightweight hybrid model for classification.The detailed research content is as follows:Fruit quality grading model based on traditional machine learning.In this paper,913 citrus images of four different qualities are taken as the research object,which are randomly divided into training set(70%)and test set(30%).Thirty-one features of citrus color,shape and texture are extracted by image processing methods,and PCA is used for feature dimension reduction.The classification results of SVM are the best,and the accuracy rate is 93.43%.Citrus quality classification model based on convolutional neural network.The data was expanded by rotation and adding noise,and a data set of 6391 citrus images was obtained.Four classical convolutional neural network models including VGG16,Res Net50,Inception V3 and Mobile Net V3 were used to identify the quality of citrus.The accuracy of Res Net50 model is the highest 97.57%,and the accuracy of VGG16,Inception V3 and Mobile Net V3 models are 95.54%,96.63% and 96.79%,respectively.Compared with traditional machine learning models,the classification accuracy of convolutional neural network model is better than that of traditional models.Optimization and improvement of convolutional neural network models.In this study,Mobile Net V3 is used as the backbone network,combined with the convolution structure of Inception V3 to strengthen the multi-scale feature extraction of citrus,and the CBAM module is used to improve the attention of effective features of citrus,and a lightweight multi-scale convolutional neural network,namely Mobile-Inception Net model,is constructed.While effectively controlling the number of parameters and reducing the computational complexity,the model can extract features of citrus more comprehensively,focus on salient features,fuse features of different scales for collaborative reasoning,and improve the recognition accuracy and generalization ability.The accuracy of the model reaches 97.73%.In addition,on the basis of the improved convolutional neural network model,combined with traditional machine learning classification models,which is used for classification of citrus.The results of CNN_KNN,CNN_SVM,CNN_MLP and CNN_Adaboost are compared and analyzed,and the CNN_SVM combination model has the highest accuracy.Compared with classical machine learning models and convolutional neural network models,the proposed method achieved the best accuracy and provides technical support for citrus quality classification.