Research On Deep Learning Of Edible Oil Interference And Spectral Information

Along with the rapid development of China's edible oil market and industry,the detection of edible oil,especially the identification of "gutter” became an imminent food safety issue.However,the existing traditional detection technologies,namely conventional physical and chemical detection technology,nuclear magnetic resonance identification method,Raman spectroscopy detection,near-infrared spectroscopy are tanglesome and poorly systematic.It is obvious that these oil identification and detection technologies cannot form an effective testing system,and there is a lack of high-quality national testing standard.This paper focuses on exploring the feature extraction ability of deep learning methods for edible oil visible spectral images,and studies the recognition and classification efficiency of four kinds of edible oils and four kinds of adulterated oil spectral information in different convolutional neural network(CNN)models.The detection optimization model for a variety of different edible oil products is obtained through the generalization analysis of the model,which provides a theoretical and practical basis for the establishment of a general detection system and even standards for edible oil.This paper takes Hujihuapeanut oil,Arowana brand rapeseed oil,Arowana brand soybean oil,kitchen waste oil(Luhua brand peanut oil that has been fried many times),and four kinds of adulterated oils as the research objects,the spectral information of oil at ??(400,800)nm is obtained by adopting two different types of fourier transform spectrometer.Subsequently,three convolutional neural networks(CNNs)are built by using the stochastic gradient descent algorithm to achieve feature extraction and identification classification of spectral information.Finally,this paper analyzes the generalization of the model comprehensively on basis of the four model evaluation indicators.Main contents and results of the above research are as follows:(1)The influence of the self-made double-layer slide holding device containing oil products on the oil identification results is explored.The spectra obtained by the LSPF-1experimental equipment of Hujihuapeanut oil,Arowana brand rapeseed oil,double-layer slide holding device,and kitchen waste cooking oil are imported into three CNN models for training.Under the condition that the training samples are consistent,the training accuracy of the three models varies within 2%,which shows that the influence of the double-layer slide holding device on the experimental results can be ignored.(2)Effectively utilizing three new CNN models and implementing deep learning methods to achieve feature extraction,recognition and classification of edible oil visible light band spectral information.A total of 600 spectrograms of three kinds of edible oils are selected and imported into three new CNN models: Resnet,Dense Net,and Mobile Net.The recognition accuracy rates obtained after training are 88.5%,93.67%,and 89.5% respectively.(3)Complete the identification of the above three new CNNs in more eight oil products effectually.When the input data of the model are four kinds of edible oil spectra obtained by LSPF-1 equipment,the accuracy rates of Resnet,Dense Net,and Mobile Net are 90.71%,94.57%,and 92.16% respectively;According to the oil doping scheme of 2:8 and 5:5,another four kinds of mixed oil products are obtained,and the LFY-11 spectrometer is used to measure the visible light band spectra of the 4 kinds of mixed oil products.The model training is obtained.Resnet,Dense Net,and Mobile Net of the accuracy rates are 91.3%,91.3%,and 87.5%separately.Thus,the identification and classification of eight kinds of oil products can be realized,and the regular identification changes caused by the difference of equipment and oil products can be disclosed.(4)The generalization and robustness of the three models are verified by different evaluation indicators,and the recognition accuracy is significantly higher than the existing research.The training results of the three models are compared and analyzed.The evaluation indicators of each model are mainly the Accuracy rate,Precision,Recall and F1 Score.By comparing the evaluation indicators,it is found that for oil spectrum training,the three CNN models built in this paper all present good generalization and robustness.Compared with existing oil identification research,its identification accuracy is generally 9% higher,which provides a new digital method for the detection,identification and analysis of edible oil.This paper contains four innovation points as follows.(1)It is discovered that the self-made double-layer slide holding device used in obtaining the edible oil spectrum has little interference with the results of the deep learning of oil spectrum.The results of this research not only provide theoretical premise and experimental process for spectral technology similar to oil,but also break through the idea of "must remove the interference of the holding device on the spectral information of the sample",that is,the optical mixing information can be investigated holistically with the new method.(2)Compared with the spectral information in the near-infrared band or ultraviolet band of oil products studied by most institutions,this study adopts a high-precision and high-resolution fourier transform spectrometer,and obtains eight kinds of oil products in the 400-800 nm band.The spectral information is found to have reliable spectral characteristic information,which breaks through the limitation of using near-infrared band and ultraviolet band to detect and identify oil products.Moreover,the protocol for selecting two healthy oils and one non-healthy oil for identification studies was validated for generalizability.(3)Discarding most of the data smoothing preprocessing methods used in deep learning of edible oil spectrum,in this study,only rotation,translation and scaling of the spectrograms are carried out to maintain excellent accuracy rate,more original data features are retained to facilitate the identification of more complex edible oils,which will open up the research of multi-component analysis of oil products.(4)The generalization and robustness of the model are compared and analyzed using four evaluation indicators.It innovates the accuracy,the multi-index comprehensive evaluation method of accuracy,F1 score and standard deviation,and effectively establishes an oil identification method based on deep learning technology,reflecting ideological innovation.In conclusion,three new convolutional neural networks based on stochastic gradient descent algorithm can efficiently identify eight kinds of oil spectral images,which verify the feasibility of deep learning method for spectral classification and recognition of edible oil.The oil identification and classification model established in the ??(400,800)nm band provides a reliable basis for the study of edible oils in the near-infrared and ultraviolet bands.Futhermore,the multi-index comprehensive evaluation method provides valuable analytical ideas for the testing and identification of edible oil.Therefore,this paper has significant scientific value and application prospect,and is expected to establish a general testing system for edible oil.