CN108760655A - A kind of apple sense of taste profile information method for visualizing - Google Patents

Abstract

The invention discloses a method for visualizing apple taste map information, comprising the following steps: S1, acquisition of taste data of apple samples based on electronic tongue; S2, acquisition and preprocessing of apple sample data based on hyperspectral technology; S3, selection of characteristic bands ; S4, the construction of the apple taste visualization model; S5, after the GS-SVM model is established, the predicted spectral values under the 10 bands are input by points, and the taste predicted output values under the spectral values of each point are calculated as K(i , j), the taste value is converted into the corresponding actual color value, and the apple taste information is visualized. The invention combines the advantages of apple samples in hyperspectral single-point difference with the electronic tongue taste overall information detection technology, thereby realizing the visualization of apple taste map information.

Description

A method for visualization of apple taste map information

technical field

The invention relates to the field of apple taste analysis, in particular to a method for visualizing apple taste map information.

Background technique

China is a big apple producing country, and its output accounts for 65% of the total apple output. Apple fruit is rich in minerals and vitamins. It is highly soluble and easily absorbed by the human body. Therefore, it is called "living water" and is one of the fruits that people often eat. As one of the important factors reflecting its quality, the taste information of apples affects the choice of most consumers when purchasing. Accurate and efficient detection and characterization of taste information has practical guiding significance for apple breeding, planting and storage. Traditional physical and chemical testing methods cannot reflect the taste and sensory information of apples, and the most commonly used artificial sensory evaluation results are not objective enough due to factors such as the assessor's psychology and the surrounding environment. Based on this, the SA-402B electronic tongue, as a single-taste intelligent bionic detection system, gradually replaces the application of traditional detection methods in taste information with its advantages of objectivity and precision. Sampling and batch processing of apple samples to detect the overall taste information of the sample cannot reflect the distribution of various taste information in the sample space.

Contents of the invention

In order to solve the above problems, the present invention provides a method for visualizing apple taste map information.

In order to achieve the above object, the technical scheme that the present invention takes is:

A method for visualizing apple taste map information, comprising the steps of:

S1. Acquisition of taste data of apple samples based on electronic tongue

S11. Select apples picked in the same batch and at the same location with positive shape, uniform size, no defects or pollutants, and select 90 high-quality fruits according to the national standard GB/T10651-2008;

S12. Store the sample at room temperature 20±2°C and relative humidity 55±5% for 24 hours after numbering, keeping the temperature and humidity unchanged;

S13. Wash, peel, squeeze, and filter each apple sample to take 40 mL of the supernatant and place them in two pure measuring cups, and obtain 90 × 3-dimensional sugar content in 30 seconds through the detection of the electronic tongue in sequence according to the numbering order. Sour and salty taste data; before the test, the sensor was cleaned in the positive and negative cleaning solutions for 90s, and then in the reference solution for 120s, and then in another reference solution for 120s, and the sensor balance returned to zero. Zero 30s; after the balance is reached, the test starts, the test time is 30s, and the cleaning step is automatically entered after each measurement;

S2. Acquisition and preprocessing of apple sample data based on hyperspectral technology

S21. Correct the hyperspectral image by using the black-and-white calibration method through the following formula to eliminate the influence of noise:

In the formula, R _d - dark image, R _w - diffuse reflectance image of whiteboard, R _s - original diffuse reflectance spectrum image of apple sample, R - corrected diffuse reflectance spectrum image.

S22. Carry out vectorization processing on the black-and-white calibrated image, obtain an image description curve, delineate the target image, generate a specific mask image of interest, and control the sample image processing area;

S23. Perform image cropping on the masked image, and the spectral values of the masked spectral region are all 0 except for the region of interest.

S24. Obtain 90 apple spectral images after preprocessing, wherein, since apples are spherical fruits, an apple sample is divided into 4 surfaces when a hyperspectral sorter is used to irradiate the sample, and each surface Take 5 regions of interest numbered 1-5 respectively, each region of interest is 300 pixels in size, calculate the average value of the spectra of the 5 surfaces of each sample, and finally obtain 90×256 dimensional apple data;

S3. Selection of characteristic bands

The sensitive bands corresponding to the three basic tastes of sweet, sour and salty were respectively selected, the optimal number of sensitive bands was determined by the coefficient of variation, and then the correlation between the taste information of the apple sample and the spectral image was determined by the gray relational degree (GRA) method degree;

S4. Construction of Apple Taste Visualization Model

Using SVM to predict the concentration value of each discrete point of taste information, the hyperplane function, RBF kernel function and regression function formulas (2), (3) and (4) are as follows:

K(x, x')=exp(g Px-x'P ² ) (2)

f(x)=wφ(x)+b (3)

In the formula, w-the normal vector of the hyperplane, φ(x)-non-linear mapping function, b-bias, g-width coefficient;

S5. Taste visual presentation

After the GS-SVM model is established, the predicted spectral values under the 10 bands are input point by point, and the taste prediction output value under the spectral values of each point is calculated as K(i, j), and the taste value is converted into the corresponding actual value. Color value, which visualizes the taste information of apples.

Preferably, the gray relational degree method specifically includes the following steps: first, convert the multidimensional data into a unified dimensionless form, define a reference sequence as a taste information sequence, and compare the sequence as a spectrum information sequence under each band of 380-1038nm to perform dimensionless processing, so as to eliminate the difference in dimension between taste and spectral information; then, calculate the gray correlation coefficient between each single taste information sequence and the spectral information sequence under the full band; finally, calculate the gray correlation degree of each band .

Preferably, said step S4 optimizes the parameters c and g based on genetic algorithm (GA) and grid search (GS). In genetic algorithm, the maximum genetic algebra is 100, and the initial population number is 20. The search range of the parameter c is from 0 to 100, and the value of g is from 0 to 100; in the grid search method, the parameters are optimized at an interval of 0.5, and the search range of the parameters c and g is from 2 ^-10 to 2 ¹⁰ .

Preferably, in the process of establishing the model in the step S4, when performing spectral data and taste information mapping, the spectral values under each band are averaged before establishing the model; wherein the Kennard-Stone method is used Select 2/3 spectral data samples as the training set and 1/3 spectral data as the prediction set, and realize the visual analysis of spectral-taste information based on GS-SVM and GA-SVM.

The present invention combines the advantages of apple sample hyperspectral single-point difference with the electronic tongue taste overall information detection technology, thereby realizing the visualization of apple taste map information, thereby providing a relatively accurate analysis method for apple taste analysis.

Description of drawings

Fig. 1 is a spectrum image description curve in an embodiment of the present invention.

Fig. 2 is a vector mask image in an embodiment of the present invention.

Fig. 3 is a spectral curve of a region of interest in an embodiment of the present invention.

Fig. 4 is the variation curve diagram of coefficient of variation in the embodiment of the present invention

FIG. 5 is a schematic diagram of a parameter optimization process in an embodiment of the present invention;

Among them, (a) is the genetic algorithm parameter optimization process; (b) is the grid search parameter optimization process.

Fig. 6 is a graph showing the visualization results of taste sensation based on GS-SVM in the embodiment of the present invention.

Detailed ways

In order to make the objects and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the examples. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

Example

In this example, Aksu candied apples picked in the same batch and at the same location are selected as the research object, with fruit shape, uniform size, and no defects or pollutants, and 90 high-quality fruits are selected according to the national standard GB/T10651-2008; including the following steps :

S1. Acquisition of taste data of apple samples based on electronic tongue

After numbering the sample, store it at room temperature at 20°C (±2°C) and relative humidity at 55% (±5%) for 24 hours, keeping the temperature and humidity basically unchanged. Measure 90 samples sequentially according to the numbering order, wash, peel, squeeze, and filter each apple sample to get 40mL of supernatant and place them in 2 pure measuring cups for testing. Before the test starts, the sensor is cleaned in the positive and negative electrode cleaning solution for 90s, and then in the reference solution for 120s, then in another reference solution for 120s, and the sensor balance is reset to zero for 30s. After the balance is reached, the detection is started, the test time is 30s, and the cleaning step is automatically entered after each measurement. The 90×3-dimensional three-dimensional taste (sugar, sour, salty) data at 30 seconds is obtained through the detection of the electronic tongue.

S2. Acquisition and preprocessing of apple sample data based on hyperspectral technology

During the hyperspectral image acquisition process, due to the difference in the intensity distribution of the light source in each band and the influence of the dark current noise of the camera, some noise information will be included. These noise information will affect the quality of hyperspectral images, and then affect the accuracy and stability of qualitative or quantitative analysis models of hyperspectral images. Therefore, the black and white calibration method is used to correct the hyperspectral image to eliminate the influence of noise, as shown in formula (1).

In the formula, R _d - dark image, R _w - diffuse reflectance image of whiteboard, R _s - original diffuse reflectance spectrum image of apple sample, R - corrected diffuse reflectance spectrum image.

Vectorize the black and white calibrated image to obtain the image description curve as shown in Figure 1. After delineating the target image, generate a specific mask image of interest as shown in Figure 2 to control the sample image processing area. Since the edge of the acquired image contains a lot of spectral noise, which will increase the difficulty of data processing in the later stage, the image after mask processing is cropped, and the spectral values of the masked spectral region are all 0 except for the region of interest.

Get 90 preprocessed apple spectral images. Among them, since apples are spherical fruits, an apple sample is divided into 4 surfaces when a hyperspectral sorter is used to irradiate the sample, and 5 regions of interest (numbered 1-5) are taken on each surface ( Each region of interest is approximately 300 pixels in size). The average spectral values of the 5 regions of interest on each section are shown in Figure 3, and the average spectral values of the 5 surfaces of each sample are calculated to obtain the 90×256-dimensional apple data. S3. Selection of characteristic bands

Spectral images reflect the differences in the internal quality of apples, and the electronic tongue obtains taste information through the free functional groups in apple samples. Therefore, the method of gray relational degree (GRA) is used to determine the degree of correlation between the two, and the sensitive bands corresponding to the three basic tastes of sweet, sour and salty are selected respectively. Before gray correlation analysis, the coefficient of variation was used to determine the optimal number of sensitive bands. The larger the comprehensive coefficient of variation, the lower the correlation between variables. To study the influence of the number of characteristic bands with an interval of 2 on the coefficient of variation, it can be seen from Figure 4 that the number of characteristic bands achieved the highest value at 10. coefficient of variation. Among them, in the gray relational degree method, firstly, the multi-dimensional data is converted into a unified dimensionless form, the reference sequence is defined as a taste information sequence, and the comparison sequence is the spectral information sequence under each band of 380-1038nm for dimensionless processing. This eliminates the difference in the dimensions of gustatory and spectral information. Then, calculate the gray correlation coefficient between each single taste information series and the spectral information series under the full band. Finally, the gray correlation degree of each band is calculated. The results are shown in Table 1, the top 10 characteristic bands with higher correlation value for each single taste information.

Table 1 The top 10 characteristic bands of each taste

S4. Construction of Apple Taste Visualization Model

Spectral and taste information data are discrete, and there is a local linear and nonlinear relationship between them, and the main idea of SVM is to establish a regression hyperplane as a decision surface, and use the kernel function to map multidimensional data to high-dimensional space, making it Be as linear as possible to solve the local nonlinearity of the original data, and finally make the distance between all the data in the set and the hyperplane the shortest. Based on this, this paper uses SVM to predict the concentration value of each discrete point of taste information, in which the hyperplane function, RBF kernel function and regression function formulas (2), (3), and (4) are as follows. The result of the SVR modeling process depends on the parameters c and g, and the correct and effective parameter selection has good regression performance for the support vector machine. Therefore, this paper optimizes the parameters c and g based on genetic algorithm (GA) and grid search (GS), as shown in Figure 5. In the genetic algorithm, the maximum genetic algebra is 100, the initial population number is 20, the search range of parameter c is 0 to 100, and g is 0 to 100. In the grid search method, parameter optimization is carried out at an interval of 0.5, and the search range of parameters c and g is 2 ^-10 to 2 ¹⁰ . In the process of building the model, because the taste value detected by the experiment is the taste value of the whole sample, when mapping the spectral data and taste information, the average value of the spectral value under each band is calculated before the model is established. The Kennard-Stone method is used to select 2/3 of the spectral data samples as the training set, and 1/3 of the spectral data as the prediction set. Based on GS-SVM and GA-SVM, the visual analysis of spectral-taste information is realized. The results are shown in Table 2, which shows the parameter optimization process of GA-SVM. The genetic algorithm quickly screened out the optimal modeling parameters c as 6.3859 and g as 83.0159 under the condition that the optimal fitness Mse was 0.01667. The GS-SVM In the parameter optimization process, the grid search screened out the optimal modeling parameters c as 22.6274 and g as 0.0019531 under the condition that the optimal fitness Mse was 0.0037346, so GS-SVM was selected as the model for calculating the apple taste value.

K(x, x')=exp(g Px-x'P ² ) (2)

f(x)=wφ(x)+b (3)

In the formula, w-the normal vector of the hyperplane, φ(x)-non-linear mapping function, b-bias, g-width coefficient

Table 2 Result values of c and g optimization process

S5. Taste visual presentation

After the GS-SVM model is established, the predicted spectral values under the 10 bands are input point by point, and the taste prediction output value under the spectral values of each point is calculated as K(i, j), and the taste value is converted into the corresponding actual value. Color value, which visualizes the taste information of apples. Figure 6 is a single-point taste distribution map of sweetness, sourness, and saltiness. The taste of the legend gradually becomes stronger from dark blue to brick red. It can be seen from the picture that Aksu candied apples are mainly sweet and sour, with a very light salty taste, and the sweetest and most widely distributed. The sweet taste is mainly distributed on both sides of the equator line, and the central axis is sweeter; the distribution position of the sour taste is similar to the sweet taste; the salty taste is more uniform and widely distributed in the equator section.

The above is only a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, and these improvements and modifications should also be It is regarded as the protection scope of the present invention.

Claims (4)

1. an apple taste map information visualization method, is characterized in that, comprises the steps: S1. Acquisition of taste data of apple samples based on electronic tongue S11. Select apples picked in the same batch and at the same location with positive shape, uniform size, no defects or pollutants, and select 90 high-quality fruits according to the national standard GB/T10651-2008; S12. Store the sample at room temperature 20±2°C and relative humidity 55±5% for 24 hours after numbering, keeping the temperature and humidity unchanged; S13. Wash, peel, squeeze, and filter each apple sample to take 40 mL of the supernatant and place them in two pure measuring cups, and obtain 90 × 3-dimensional sugar content in 30 seconds through the detection of the electronic tongue in sequence according to the numbering order. Three kinds of taste data: sour taste and mature taste; before the test starts, the sensor is cleaned in the positive and negative electrode cleaning solution for 90s, after the end of the test, it is washed in the reference solution for 120s, and in another reference solution for 120s, and the sensor is balanced. Zero 30s; after the balance is reached, the test starts, the test time is 30s, and the cleaning step is automatically entered after each measurement; S2. Acquisition and preprocessing of apple sample data based on hyperspectral technology S21. Correct the hyperspectral image by using the black-and-white calibration method through the following formula to eliminate the influence of noise: In the formula, R _d - dark image, R _w - diffuse reflectance image of whiteboard, R _S - original diffuse reflectance spectrum image of apple sample, R - corrected diffuse reflectance spectrum image. S22. Carry out vectorization processing on the black-and-white calibrated image, obtain an image description curve, delineate the target image, generate a specific mask image of interest, and control the sample image processing area; S23. Perform image cropping on the masked image, and the spectral values of the masked spectral region are all 0 except for the region of interest. S24. Obtain 90 apple spectral images after preprocessing, wherein, since apples are spherical fruits, an apple sample is divided into 4 surfaces when a hyperspectral sorter is used to irradiate the sample, and each surface Take 5 regions of interest numbered 1-5 respectively, each region of interest is 300 pixels in size, calculate the average value of the spectra of the 5 surfaces of each sample, and finally obtain 90×256 dimensional apple data; S3. Selection of characteristic bands The sensitive bands corresponding to the three basic tastes of sweet, sour and salty were respectively selected, the optimal number of sensitive bands was determined by the coefficient of variation, and then the correlation between the taste information of the apple sample and the spectral image was determined by the gray relational degree (GRA) method degree; S4. Construction of Apple Taste Visualization Model Using SVM to predict the concentration value of each discrete point of taste information, the hyperplane function, RBF kernel function and regression function formulas (2), (3) and (4) are as follows: K(x, x')=exp(g Px-x'P ² ) (2) f(x)=ωφ(x)+b (3) In the formula, w-the normal vector of the hyperplane, φ(x)-non-linear mapping function, b-bias, g-width coefficient; S5. Taste visual presentation After the GS-SVM model is established, the predicted spectral values under the 10 bands are input point by point, and the taste prediction output value under the spectral values of each point is calculated as K(i, j), and the taste value is converted into the corresponding actual value. Color value, which visualizes the taste information of apples. 2. a kind of apple taste map information visualization method as claimed in claim 1, is characterized in that, The gray relational degree method specifically includes the following steps: first, convert the multidimensional data into a unified dimensionless form, define a reference sequence as a taste information sequence, and compare the sequence as a spectral information sequence under each band of 380-1038nm to carry out infinite Then, the gray correlation coefficients between each single taste information sequence and the spectral information sequence under the full band are calculated; finally, the gray correlation degree of each band is calculated. 3. a kind of apple taste map information visualization method as claimed in claim 1, is characterized in that, described step S4 carries out optimization to parameter c and g based on the method of genetic algorithm (GA) and grid search (GS) , in the genetic algorithm, the maximum genetic algebra is 100, the initial population number is 20, the search range of the parameter c is 0 to 100, and the g is 0 to 100; in the grid search method, the parameters are optimized at an interval of 0.5 , the search range of parameters c and g is 2 ^-10 to 2 ¹⁰ . 4. a kind of apple taste map information visualization method as claimed in claim 1, is characterized in that, in the process of building a model in described step S4, when carrying out spectral data and taste information mapping, the The average value of the spectral value is calculated before the model is established; 2/3 of the spectral data samples are selected as the training set by the Kennard-Stone method, and 1/3 of the spectral data are selected as the prediction set, based on GS-SVM and GA- SVM realizes visual analysis of spectral-taste information.