A method for rapidly identifying lily powder based on the multimodal technology of electronic tongue and electron microscope

Abstract

The present invention belongs to the technical field of food detection, and discloses a method for rapidly identifying lily powder based on an electronic tongue and electron microscopy multimodal technology. By collecting the array data of the electronic tongue taste sensors and the image data under a scanning electron microscope, and performing preprocessing on the data such as first-order derivative, vector normalization, filtering and smoothing, the dimension is reduced and the characteristic information that can most represent the sample is extracted. A discriminant analysis model for rapidly identifying Lilium brownii var. viridulum, Lilium bulbocapicum, and Lilium speciosum is constructed by a clustering algorithm, and a data model of different types of lily powder is constructed, which is beneficial to realizing the quality monitoring and evaluation of lilies throughout the whole process of lily powder production, processing, circulation, and consumption. This detection method does not require the use of a large amount of reagents, is green and environmentally friendly, has simple operation, good reproducibility, stable and accurate results, and can realize rapid on-line detection and real-time evaluation feedback.

Description

Technical Field

[0001] The present invention relates to the technical field of food detection, and particularly relates to a method for rapidly identifying lily powder based on an electronic tongue and electron microscopy multimodal technology. Background Art

[0002] As a perennial herbaceous plant, lily belongs to the food with both medicinal and edible properties, and has extremely high economic value and broad development prospects. It is rich in starch, protein, various vitamins and minerals, etc., and also contains bioactive ingredients such as saponins, and has the effects of clearing the heart and calming the mind, clearing heat and moistening the lungs. The traits of bulblet lily, bulb lily, and officinal lily are highly similar, especially the first two. Due to the dynamic changes of components during the development cycle of plant bulblets and bulbs, it is extremely difficult to examine various indicators through physical and chemical detection. At present, the standard method for detecting such products is still blank.

[0003] At present, the methods for identifying lily powder types are summarized as follows: using XRD diffraction method and infrared spectroscopy to examine the internal crystal structure and surface layer orderliness; using physical and chemical detection methods to examine the content of amylose and amylopectin, gelatinization temperature, freeze-thaw stability of starch paste, swelling potential and solubility, etc.; using differential scanning calorimetry to examine the characteristic parameters of starch gelatinization, such as the difference in gelatinization peak temperature and endothermic peak, thermodynamic enthalpy value and other thermodynamic characteristics. These reported methods have the following defects:

[0004] (1) Larger error

[0005] For varieties with highly similar traits, especially after being made into powder, it is almost impossible to distinguish them structurally by microscopic identification and morphological identification methods.

[0006] (2) Unable to evaluate its quality as a whole

[0007] Physical and chemical inspection and instrumental inspection, for the characteristic responses of certain specific components, due to the complexity of components and efficacy, it is sometimes difficult to evaluate the integrity of its quality standard. Especially for samples from different regions and different harvesting times in China, due to the differences in plant metabolism and environment, the functional chemical components will be different, which is likely to cause errors.

[0008] (3) Various components will undergo dynamic changes during the plant development process

[0009] As lily plants grow and develop, various components will change dynamically. It is difficult to identify lily species simply from the various indicators of the above-mentioned physical and chemical tests. This is because we have to take into account the characteristics of lily plants and the division of natural cycles. For example, the bulb development process can be used as a "source" and a "sink." The former period provides nutrients for the lily and supplies the growth of the entire plant, while the latter period accumulates components to prepare for the next cycle. The formation of bulbils and the base of the petioles of the aboveground stems and leaves also follow their natural development laws. Therefore, the use of various physical and chemical indicators to identify lily species is limited by the plant development cycle, and physical and chemical testing has limitations.

[0010] (4) The operation is cumbersome and it is difficult to achieve rapid detection

[0011] Physical and chemical testing requires a large amount of reagents, the pre-treatment process is time-consuming and cumbersome, and it is equipped with large analytical instruments and requires operators to have solid analytical experience. Especially when faced with large quantities of samples, it is impossible to quickly test their quality and monitor and identify them online, which makes it difficult to promote to small and medium-sized enterprises. Summary of the invention

[0012] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a method for monitoring and evaluating the quality of lilies in the production, processing, circulation and consumption of lily powder, which is green, environmentally friendly, easy to operate, reproducible, and produces stable and accurate results.

[0013] In order to achieve the above-mentioned purpose of the present invention, the following technical scheme is adopted:

[0014] A method for quickly identifying lily powder based on electronic tongue and electron microscope multimodal technology, comprising the following steps:

[0015] S1. Preparation of electronic tongue samples: remove impurities from different varieties of lily powder samples, crush them into powder, pass through a No. 3 sieve, dissolve and let stand, filter, and put the filtrate into a special test cup for the electronic tongue;

[0016] S2. Electronic tongue operation: Activate the electronic tongue with activation solution, calibrate and pre-balance the taste sensor, prepare reference solution, balance the activated sensor in the reference solution, and then immerse it in the reference solution and the sample solution to be tested in step S1 in turn. According to the membrane potential value, combined with the signal acquisition software provided by the electronic tongue, convert the first taste signal and the aftertaste signal into taste information acquisition data;

[0017] S3. Scanning electron microscopy: The lily powder samples were analyzed by scanning electron microscopy at magnifications of 1500 times, 1800 times, and 2000 times, respectively, to obtain electron microscopy images of the sample surface morphology and structure;

[0018] S4. Electron Microscope Image Dataset Processing: Collect images of lily powder under a scanning electron microscope, and divide them into a training set and a test set according to a ratio of 7:3. Perform processing such as rotation, flipping, and filtering on the training set images. The rotation angle of the images is randomly selected within [0°, 135°], and the blank areas at the edges after rotation are filled with the adjacent background. The images are flipped horizontally. Gaussian filtering is used for image filtering. After enhancement processing by these three methods, the training set image dataset is enlarged.

[0019] S5. Electronic Tongue Data Preprocessing and Principal Component PCA Extraction: Use the taste values of the samples measured by the electronic tongue as analysis variables. After mathematical preprocessing methods such as vector normalization and ND filter smoothing, the PCA method is adopted. According to the eigenvalue size, the principal component factors are extracted for data dimensionality reduction. The SIMCA software is used to perform PCA analysis on the samples. The scaling method is selected as centering processing. According to the sample information and the cumulative contribution rate obtained by the PCA method, select the information representing the principal components, which can reflect the basic characteristics and main information of different types of lily powder. Using the extracted different principal components as coordinate axes, construct a principal component plane graph, and project the multivariate variables of the samples onto a two-dimensional plane through dimensionality reduction, which is convenient for observing the overall distribution of the samples and the contribution of each variable to the sample distribution.

[0020] S6. Establish a Rapid Identification Model for Different Types of Lily Powder: According to the picture feature information and electronic tongue taste feature information of different types of lily powder, use a deep learning convolutional neural network and build a prediction model for rapidly identifying lily powder with the ResNet algorithm.

[0021] Furthermore, the activation solution described in step S2 is prepared by mixing 3.33 mol / L KCl and 0.07 mmol / L AgCl in a ratio of 1:1, and the reference solution is prepared by mixing 0.3 mmol / L tartaric acid and 30 mmol / L KCl in a ratio of 1:1.

[0022] Furthermore, the electron microscope image dataset processing described in step S4 also includes preprocessing the electron microscope images: Set the color space, set a threshold on the saturation channel to achieve the segmentation of the target object and the background, and gray-scale and fuse the images under the electron microscope into a complete image. Use the gray-level co-occurrence matrix method, define the image as a gray-scale image, and use the graycomatrix function and graycoprops function to calculate the gray-level co-occurrence matrix of the image and the corresponding parameters such as contrast, correlation, and homogeneity.

[0023] Furthermore, the electron microscope image data set processing described in step S4 also includes: cutting out the image into a rectangle with the larger length and width as the edge, and then proportionally enlarging or reducing it to the same resolution (224×224), and then mapping the image pixel value from [0, 255] to [0, 1] or [-1, 1] to complete the normalization operation.

[0024] Furthermore, the conversion of the first taste signal and the aftertaste signal into taste information acquisition data described in step S2 includes: the activated sensor is balanced in the reference solution and then immersed in the reference solution and the sample solution to be tested in turn, and the membrane potential values Vr and Vs are measured respectively; after being briefly washed with the reference solution, the sensor is immersed in a new reference solution again to measure Vr'; and according to the measured potential values, the first taste signal value (R) and the aftertaste signal value (CPA) of the sample are obtained, and the calculation formula is as follows:

[0025] R=Vs-Vr

[0026] CPA=Vr'-Vr

[0027] Based on the Weber-Fechner law, the analysis software of the electronic tongue converts the measured first taste signal value and aftertaste signal value into certain taste characteristic information.

[0028] Furthermore, the step S6 of establishing a prediction model for rapid identification of lily powder includes:

[0029] (1) Convolutional neural network structure: Convolutional neural networks need to complete the function of feature extraction through convolution operations. Network structures include VGGNet, ResNet, DenseNet, etc. When modeling convolutional neural networks, in order to reduce the dimension of each feature map, reduce the number of convolution operations and retain important feature information, pooling layer processing is introduced; in the connection layer, a large number of feature maps are obtained, and the normalization function softmax is used to classify with a classifier;

[0030] (2) ResNet residual learning unit: The network model introduces the ResNet residual unit. The information of the previous residual block is directly transmitted to the next residual block, which is conducive to deepening the number of network layers and extracting more high-level image features to complete the classification of different types of lily powder of the same variety;

[0031] (3) Deep learning convolutional neural network modeling based on the ResNet algorithm: The ResNet deep convolutional network structure has a total of 18 layers, including five parts. The output sizes of each layer are different. The structures are [7*7, 64] (conv1); [3*3] (max pooling layer); [3*3, 64]×2, [3*3, 64]×2 (conv2); [3*3, 128]×2, [3*3, 128]×2 (conv3); [3*3, 256]×2, [3*3, 256]×2 (conv4); [3*3, 512]×2, [3*3, 512]×2 (conv5), followed by a fully connected layer and a softmax classifier;

[0032] (4) Model verification: Adopt the method of five-fold cross-validation. Each type of dataset is randomly and evenly divided into a training set and a test set for cross-validation; The initial value of the learning rate for network training is set to 0.001. Use the Adam optimization algorithm to dynamically adjust the learning rate of each parameter. Use the SoftMax classifier for classification. Adopt the cross-entropy loss function to evaluate the gap between the true value and the predicted value, and determine the number of training iteration steps;

[0033] After the pooling layer, use L2 norm regularization and the dropout function to optimize the network model. The value of the L2 regularization parameter is 0.0008, and the dropout function parameter is set to 0.05; Input the feature information of the prediction set into the model to obtain the classification result of lily powder. The recognition accuracy of the model is 95.24%. The prediction is stable and the precision is good, both of which can remain above 95%.

[0034] An electronic tongue is a detection instrument developed on the basis of bionics that covers the analysis and identification of flavor components. Its composition includes three systems: sampling, sensors, and data processing. It can objectively analyze the overall flavor of the substance to be measured and analyze the differences in the overall flavor between samples from the sensor response values. By analyzing the signal response values of the sensors, the relative intensity of the flavor can be quantitatively predicted within a certain range, which helps to examine the differences between components and compare with the data obtained by the sensory evaluation panel.

[0035] The principle of the scanning electron microscope is to use a very narrow and focused high-energy electron beam to scan the sample. Through the interaction between the electron beam and the substance, various physical information is excited, and these information are collected, amplified, and re-imaged to achieve the purpose of characterizing the microscopic morphology of the substance.

[0036] Perform multi-fusion data modeling on the electronic tongue sensor response data and image data. The detection precision of the model depends on (1) data preprocessing method (2) chemometric method for extracting sample information (3) determining the clustering algorithm and optimizing the modeling parameters for modeling.

[0037] The pattern recognition method is a commonly used qualitative method in chemometrics, which is to identify characteristic components, exclude invalid components and unify the same kind of substances in the complex components of substances, and finally achieve the purpose of prediction classification and qualitative analysis.

[0038] The principal component analysis (PCA) and clustering algorithm used in the present invention belong to one of the unsupervised pattern recognition methods. The PCA method can analyze the main influencing factors from complex phenomena, reduce evaluation indicators, simplify the evaluation process, and is suitable for the comprehensive analysis of multiple indicators.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] 1. The present invention can be used for the qualitative discrimination of lily powder types. Based on the electronic tongue and scanning electron microscopy technology, the present invention uses the principal component analysis method of chemometrics to mine sample information and extract principal components, and establishes a prediction model of lily powder with a neural network. This method can quickly predict the adulterated types with powerful machine learning ability. The model is stable, the prediction is accurate, and the method has good reproducibility. Especially for lily powder samples with extremely high similarity, without the need for biological and chemical experimental design, rapid analysis and real-time monitoring of lily powder quality evaluation can be achieved through scientific modeling.

[0041] 2. The present invention does not require physical and chemical pretreatment of samples. The present invention does not need to use physical and chemical detection methods, avoiding the disadvantages of a large amount of chemical reagent waste and long time consumption. Through the data processing of electronic tongue sensors and the scanning of electron microscopy images, and scientific modeling with methods such as chemometrics and neural networks, the advantages of simple operation, good reproducibility, and environmental protection are achieved.

[0042] 2. The present invention has good detection results and low detection costs. This method is based on chemical sensor and electron microscopy image data. By modeling and optimizing the chemical measurement process with algorithms, the data used in constructing the model must be provided by standard methods and the mathematical models established through chemometrics. The data needs to be preprocessed by certain methods to optimize the algorithm modeling process. Finally, the model established with the modeling parameters can achieve stable learning and training, accurate prediction, and reliable precision. The samples in the prediction set can be quickly subjected to adaptive learning through model calibration verification to complete the species pattern recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Electron microscopy image of Lilium brownii var. viridulum Baker powder in an embodiment of the present invention

[0044] Figure 2 Electron microscopy image of Lilium bulbocapense powder in an embodiment of the present invention

[0045] Figure 3 Electron microscopy image of Lilium speciosum powder

[0046] Figure 4The principal component analysis results based on the electronic tongue feature data

[0047] Figure 5 The training process of deep learning convolutional neural network model based on ResNet algorithm

[0048] Figure 6 The accuracy of the prediction set samples based on the deep learning convolutional neural network model of the ResNet algorithm DETAILED DESCRIPTION

[0049] A method for quickly identifying lily powder based on electronic tongue and electron microscope multimodal technology comprises the following steps:

[0050] S1. Preparation of electronic tongue samples: remove impurities from lily powder samples of lily bulb powder, lily bulb powder, and lily medicinal powder, crush them into powder, pass through a No. 3 sieve, dissolve and let stand, filter, and put the filtrate into a special test cup for electronic tongue;

[0051] S2. Electronic tongue operation: Use the SA402 electronic tongue device from Japan Insent Company, which consists of AAE, CA0, CT0, C00, AE1, GL1 and Ag / AgCl reference electrodes, a 16-bit automatic sampler, and a data acquisition system. Use activation liquid to activate the electronic tongue, calibrate and pre-balance the taste sensor, and prepare the reference solution. After the activated sensor is balanced in the reference solution, it is immersed in the reference solution and the sample solution to be tested in step S1 in turn. According to the membrane potential value, combined with the signal acquisition software provided by the electronic tongue, the first taste signal and the aftertaste signal are converted into taste information acquisition data; when testing samples, each sample is tested for 120s, and each sample is tested 8 times, with 1 data tested per second. Each time a sample is tested, the cleaning time is 30s, and the average of the last 6 data is taken as the test result. The sweetness value of the GL1 sensor is collected as the basic data for data modeling.

[0052] S3. Scanning electron microscopy: The lily powder samples were analyzed by scanning electron microscopy at magnifications of 1500 times, 1800 times, and 2000 times, respectively, to obtain electron microscopy images of the sample surface morphology and structure;

[0053] S4. Processing of electron microscope image data set: The images of lily powder under scanning electron microscope were collected and divided into training set and test set in a ratio of 7:3. The training set images were rotated, flipped, filtered, etc. The image rotation was randomly selected between [0°, 135°], and the blank space on the edge after rotation was filled with the adjacent background; the image flipping was done by horizontal flipping; the image filtering was done by Gaussian filtering; after the three enhancement methods, the training set image data set was expanded;

[0054] S5. Preprocessing of electronic tongue data and extraction of principal component PCA: Using the taste values of the samples measured by the electronic tongue as analysis variables, after mathematical preprocessing methods such as vector normalization and ND filter smoothing, the PCA method is adopted. According to the eigenvalue size, the principal component factors are extracted for data dimensionality reduction; the SIMCA software is used to perform PCA analysis on the samples. The scaling method is selected as centering processing. According to the sample information and the cumulative contribution rate obtained by the PCA method, the information representing the principal components is selected, which can reflect the basic characteristics and main information of different types of lily powder; using the extracted different principal components as coordinate axes, a principal component plane graph is constructed, and the multivariate variables of the samples are projected onto a two-dimensional plane through dimensionality reduction, which is convenient for observing the overall distribution of the samples and the contribution of each variable to the sample distribution.

[0055] S6. Establish a rapid recognition model for different types of lily powder: According to the picture feature information and electronic tongue taste feature information of different types of lily powder, using the deep learning convolutional neural network and modeling with the ResNet algorithm, a prediction model for rapid recognition of lily powder is established.

[0056] Among them, the activation solution described in step S2 is prepared by mixing 3.33 mol / L KCl and 0.07 mmol / L AgCl in a ratio of 1:1, and the reference solution is prepared by mixing 0.3 mmol / L tartaric acid and 30 mmol / L KCl in a ratio of 1:1.

[0057] The processing of the electron microscope image dataset described in step S4 also includes preprocessing the electron microscope images: setting the color space, setting a threshold on the saturation channel to achieve the segmentation of the target object and the background, graying and fusing the images under the electron microscope into a complete image; using the gray-level co-occurrence matrix method, defining the image as a gray-scale image, and using the graycomatrix function and graycoprops function to calculate the gray-level co-occurrence matrix of the image and the corresponding parameters such as contrast, correlation, and homogeneity.

[0058] The processing of the electron microscope image dataset described in step S4 also includes: cutting out a rectangle along the center point of the obtained picture with the larger value of the length and width as the side, and then scaling it up or down proportionally to the same resolution (224×224). After that, the pixel values of the picture are mapped from [0, 255] to the interval [0, 1] or [-1, 1] to complete the normalization operation.

[0059] The conversion of the initial taste signal and the aftertaste signal into taste information acquisition data described in step S2 includes: after the activated sensor is balanced in the reference solution, it is successively immersed in the reference solution and the sample solution to be measured, and the membrane potential values Vr and Vs are measured respectively. After a short cleaning with the reference solution, it is immersed in a new reference solution again, and Vr’ is measured. According to the measured potential values, the initial taste signal value (R) and the aftertaste signal value (CPA) of the sample are obtained. The calculation formulas are as follows:

[0060] R = Vs - Vr

[0061] CPA = Vr’ - Vr

[0062] Based on the Weber-Fechner law, the analysis software built into the electronic tongue converts the measured initial taste signal value and aftertaste signal value into certain taste characteristic information.

[0063] The establishment of the prediction model for rapid identification of lily powder described in step S6 includes:

[0064] (1) Convolutional neural network structure: The convolutional neural network needs to complete the function of feature extraction through convolutional operations. The network structure includes VGGNet, ResNet, DenseNet, etc.; when building a model with a convolutional neural network, in order to reduce the dimension of each feature map, reduce the number of convolutional operations and retain important feature information, a pooling layer is introduced for processing; in the connection layer, a large number of obtained feature maps are used, the normalization function softmax is used, and a classifier is used for classification;

[0065] (2) ResNet residual learning unit: The ResNet residual unit is introduced into the network model, and the information of the previous residual block is directly transmitted to the next residual block, which is conducive to deepening the number of network layers and extracting more high-level features of the image to complete the classification of different categories of lily powder of the same variety;

[0066] (3) Convolutional neural network modeling based on deep learning of the ResNet algorithm: The ResNet deep convolutional network structure has a total of 18 layers, including five parts, and the output sizes of each layer are different. The structures are [7*7, 64] (conv1); [3*3] (max pooling layer); [3*3, 64]×2, [3*3, 64]×2 (conv2); [3*3, 128]×2, [3*3, 128]×2 (conv3); [3*3, 256]×2, [3*3, 256]×2 (conv4); [3*3, 512]×2, [3*3, 512]×2 (conv5), followed by a fully connected layer and a softmax classifier;

[0067] (4) Model verification: The five-fold cross-validation method is adopted. Each type of dataset is randomly and evenly divided into a training set and a test set for cross-validation. The initial value of the learning rate for network training is set to 0.001. The Adam optimization algorithm is used to dynamically adjust the learning rate of each parameter. The SoftMax classifier is used for classification. The cross-entropy loss function is adopted to evaluate the gap between the true value and the predicted value, and the number of training iterations is determined.

[0068] After the pooling layer, L2-norm regularization and the dropout function are used to optimize the network model. The value of the L2 regularization parameter is 0.0008, and the parameter of the dropout function is set to 0.05.

[0069] The training process and results of the network are shown in the appendix Figure 5 , and it is found that in the first 15 epochs of training (one epoch means training once with all samples in the training set), the loss function of the model (the loss function, where loss represents the gap between the predicted value and the target value) drops rapidly, and the classification accuracy rises rapidly, indicating that the model has good adaptability to the data. As the number of training times continues to increase, the recognition accuracy grows slowly, the loss function gradually decreases and finally converges to a stable state.

[0070] The feature information of the prediction set is input into the model to obtain the classification result of lily powder. The recognition accuracy of the model is 95.24%, and the prediction is stable and has good precision, all of which can be maintained above 95%.

Claims

1. A method for rapidly identifying lily powder based on electronic tongue and electron microscopy multimodal technology, characterized in that, The steps include: S1. Preparation of electronic tongue samples: remove impurities from different varieties of lily powder samples, crush them into powder, pass through a No. 3 sieve, dissolve and let stand, filter, and put the filtrate into a special test cup for the electronic tongue; S2. Electronic tongue operation: Activate the electronic tongue with activation solution, calibrate and pre-balance the taste sensor, prepare reference solution, balance the activated sensor in the reference solution, and then immerse it in the reference solution and the sample solution to be tested in step S1 in turn. According to the membrane potential value, combined with the signal acquisition software provided by the electronic tongue, convert the first taste signal and the aftertaste signal into taste information acquisition data; S3. Scanning electron microscopy: The lily powder samples were analyzed by scanning electron microscopy at magnifications of 1500 times, 1800 times, and 2000 times, respectively, to obtain electron microscopy images of the sample surface morphology and structure; S4. Electron microscope image dataset processing: Collect pictures of lily powder under a scanning electron microscope, and divide them into a training set and a test set according to a ratio of 7:

3. Perform operations such as rotation, flipping, and filtering on the training set images. The rotation of the images is randomly selected between [0 o , 135 o . After rotation, the blank areas at the edges are filled with the adjacent background. The image flipping is performed in a horizontal flipping manner. The image filtering is performed using Gaussian filtering; After three enhancement processes, the training set image dataset is expanded; S5. Electronic tongue data preprocessing and principal component PCA extraction: The taste value of the sample measured by the electronic tongue is used as the analysis variable. After mathematical preprocessing methods such as vector normalization and ND filtering smoothing, the PCA method is used to extract the principal component factors according to the size of the eigenvalues to reduce the data dimension; SIMCA software was used to perform PCA analysis on the samples, and the scaling method was selected as central processing. According to the sample information and the size of the cumulative contribution rate obtained by the PCA method, the information representing the principal component was selected, which can reflect the basic characteristics and main information of different types of lily powder; the principal component plane diagram was constructed with the extracted different principal components as the coordinate axes, and the multivariate variables of the samples were projected on a two-dimensional plane by dimensionality reduction, which facilitated the observation of the overall distribution of the samples and the contribution of each variable to the sample distribution; S6. Establish a fast identification model for different types of lily powder: Based on the image feature information of different types of lily powder and the taste feature information of the electronic tongue, a deep learning convolutional neural network was used to build a model with the ResNet algorithm to establish a prediction model for fast identification of lily powder; Among them, the conversion of the pre-taste signal and the aftertaste signal into taste information acquisition data described in step S2 includes: after the activated sensor is balanced in the reference solution, it is successively immersed in the reference solution and the sample solution to be measured, and the membrane potential values Vr and Vs are measured respectively. After being briefly washed with the reference solution, it is immersed in a new reference solution again, and Vr is measured. ’ , according to the measured potential values, the pre-taste signal value (R) and the aftertaste signal value (CPA) of the sample are obtained, and the calculation formulas are as follows: R=Vs-Vr CPA = Vr ’ -Vr Based on the Weber-Fechner law, the analysis software of the electronic tongue converts the measured first taste signal value and aftertaste signal value into certain taste characteristic information.

2. The method for rapidly identifying lily powder based on the multi-modal technology of electronic tongue and electron microscope according to claim 1, characterized in that, In step S2, the activation solution is prepared into 3.33 mol / L KCl+0.07 mmol / L AgCl in a ratio of 1:1, and the reference solution is 0.3 mmol / L tartaric acid+30 mmol / L KCl.

3. The method for rapidly identifying lily powder based on the multi-modal technology of electronic tongue and electron microscope according to claim 1, wherein, The electron microscope image data set processing described in step S4 also includes preprocessing the electron microscope image: setting the color space, setting the threshold on the saturation channel, realizing the segmentation of the target object and the background, graying the image under the electron microscope, and fusing it into a complete image; using the grayscale co-occurrence matrix method, defining the image as a grayscale image, and using the graycomatrix function and the graycoprops function to calculate the grayscale co-occurrence matrix of the image, as well as the corresponding contrast, correlation, homogeneity and other parameters.

4. A method for rapidly identifying lily powder based on electronic tongue and electron microscope multimodal technology according to claim 1 or 3, characterized in that The processing of the electron microscope image dataset described in step S4 further includes: cutting out a rectangle with the larger value of the length and width as the side along the center point of the picture, and then scaling it up or down proportionally to the same resolution, which is 224×224. After that, map the pixel values of the picture from [0, 255] to the interval of [0, 1] or [-1, 1] to complete the normalization operation.

5. A method for rapidly identifying lily powder based on an electronic tongue and electron microscopy multimodal technology according to claim 1, characterized in that, The establishment of a prediction model for quickly identifying lily powder described in step S6 includes: (1) Convolutional neural network structure: The convolutional neural network needs to complete the function of feature extraction through convolutional operations. The network structure includes VGGNet, ResNet, DenseNet, etc.; when building a convolutional neural network model, in order to reduce the dimension of each feature map, reduce the number of convolutional operations and retain important feature information, a pooling layer is introduced for processing; in the connection layer, a large number of obtained feature maps are used, the normalization function softmax is used, and a classifier is used for classification. (2) ResNet residual learning unit: The network model introduces the ResNet residual unit, and the information of the previous residual block is directly transmitted to the next residual block, which is beneficial to deepening the number of network layers and extracting more high-level features of the image to complete the classification of different categories of lily powder of the same variety. (3) Convolutional neural network modeling based on ResNet algorithm for deep learning: The ResNet deep convolutional network structure has a total of 18 layers, including five parts, and the output sizes of each layer are different. The structures are conv1: [7*7, 64]; max pooling layer: [3*3]; conv2: [3*3, 64]×2, [3*3, 64]×2; conv3: [3*3, 128]×2, [3*3, 128]×2; conv4: [3*3, 256]×2, [3*3, 256]×2; conv5: [3*3, 512]×2, [3*3, 512]×2, followed by a fully connected layer and a softmax classifier. (4) Model verification: Adopt the method of five-fold cross-validation, randomly and evenly divide each type of dataset into a training set and a test set for cross-validation; the initial value of the learning rate for network training is set to 0.001, use the Adam optimization algorithm to dynamically adjust the learning rate of each parameter, use the SoftMax classifier for classification, adopt the cross-entropy loss function to evaluate the gap between the true value and the predicted value, and determine the number of training iteration steps. After the pooling layer, L2 norm regularization and a dropout function are used to optimize the network model. The value of the L2 regularization parameter is 0.0008, and the parameter of the dropout function is set to 0.05; input the feature information of the prediction set into the model to obtain the classification result of lily powder. The recognition accuracy of the model is 95.24%, the prediction is stable and the precision is good, all of which can be maintained above 95%.

Images