Method for evaluating internal and external physicochemical characteristics of dwarf banana based on SVM (Support Vector Machine) multispectral model

Abstract

The invention relates to the technical field of dwarf banana quality detection, and discloses a method for evaluating internal and external physicochemical characteristics of dwarf bananas based on an SVM (Support Vector Machine) multispectral model. S2, measuring external physical and chemical characteristics; s3, measuring internal physical and chemical characteristics; s4, image acquisition and processing: connecting a camera with a computer, adjusting parameters, placing the dwarf banana sample under a lens, and laying black flannelette at the bottom of a camera obscura to avoid light reflection; the image processing comprises image conversion and storage, image visual optimization and image understanding; and S5, establishing a partial least square discrimination method. The SVM-based multi-spectral model can effectively evaluate the external physicochemical characteristics of the dwarf bananas, and has certain potential feasibility of evaluating the internal physicochemical characteristics of the dwarf bananas.

Description

Technical Field

[0001] The present invention belongs to the technical field of plantain quality detection, and specifically relates to a method for evaluating the internal and external physical and chemical characteristics of plantains based on an SVM multi-spectral model. Background Art

[0002] As an important tropical fruit, the quality of plantains directly affects the consumer experience and industrial benefits. Traditional plantain quality detection methods mainly rely on manual experience judgment and destructive physical and chemical analysis, which have disadvantages such as low efficiency, strong subjectivity, and inability to achieve non-destructive detection, and are difficult to meet the requirements of modern agricultural production and market demands.

[0003] In recent years, optical non-destructive detection technology has been widely used in the field of fruit quality detection due to its advantages of rapidity, non-destructiveness, and environmental friendliness. Among them, visible / near-infrared spectroscopy (Vis / NIRS) can reflect the internal composition and structural information of fruits, providing a new technical means for non-destructive detection of fruit quality; at present, Vis / NIRS technology has achieved certain results in the quality detection of fruits such as apples, pears, and peaches, but there is relatively little research on plantain quality detection, and there are the following problems:

[0004] Lack of systematic research on the optical properties of plantains: Existing research mainly focuses on the prediction of single quality indicators, lacking systematic measurement and analysis of the optical property parameters of plantains (such as absorption coefficient μa, reduced scattering coefficient μs′), and it is difficult to comprehensively reflect the quality information of plantains.

[0005] The prediction accuracy and stability of the model need to be improved: Existing prediction models are mostly constructed based on single algorithms, with limited prediction accuracy and stability, and it is difficult to meet the actual application requirements.

[0006] Lack of comprehensive prediction of internal and external quality indicators of plantains: Existing research mainly focuses on the prediction of single quality indicators, lacking comprehensive prediction of the internal and external quality indicators of plantains (such as single fruit weight, soluble solids, acidity, etc.), and it is difficult to comprehensively evaluate the quality of plantains.

[0007] In view of this, the present invention is specifically proposed. Summary of the Invention

[0008] To solve the above technical problems, the basic concept of the technical solution adopted by the present invention is:

[0009] A method for evaluating the internal and external physical and chemical characteristics of plantains based on an SVM multi-spectral model, comprising the following steps:

[0010] Step S1: Sample preparation, select plantains with a maturity of 6-7; after placing the plantains in a constant temperature incubator at 30 °C for 3-4 days, select plantains with a maturity of 8-9 and no diseases, pests, scars, deformities, or mechanical damage as samples;

[0011] Step S2: External physical and chemical characteristics measurement. Use an electronic balance to measure the fresh weight of the fruit, a flexible tape measure to measure the outer diameter, inner diameter, and girth of the fruit; use a 3nh precision colorimeter NR20XE to measure the chromaticity at three points (upper, middle, and lower) of the selected sample, and take the average value.

[0012] Step S3: Measurement of internal physical and chemical characteristics. Cut the banana into small pieces and homogenize it, then wrap it with sterile gauze and squeeze out the juice onto the induction mirror surface of the Atago PAL-BX / ACID6 special sugar-acid analyzer for bananas to measure the soluble solids content of the plantain.

[0013] Step S4: Image acquisition and processing. Connect the camera to the computer and adjust the parameters. Place the plantain sample directly below the lens, and lay a black flannel cloth at the bottom of the dark box to avoid reflection. Image processing includes: image conversion and storage, image visual optimization, and image understanding.

[0014] Step S5: After establishing the partial least squares discriminant method and extracting features from the training samples, establish a regression model between the independent variables and dependent variables of the training samples, and then find the most effective feature information for classification.

[0015] Step S6: Characterize the deviation between the observed value and the true value. The algorithm is as follows

[0016]

[0017] y i is the actual value, is the predicted value, and n is the number of samples.

[0018] As a preferred embodiment of the present invention, in step S3, weigh about 1 g of banana sample with an electronic scale, put it into a 100 mL beaker, then add 50 mL of distilled water with a measuring spoon, stir evenly with a glass straw to fully dilute the sample, and suck an appropriate amount of the diluted acidity measurement sample and drop it onto the instrument induction mirror surface to obtain the acidity of the plantain; repeat each measurement 3 times and take the average value.

[0019] The present invention has the following advantages compared with the prior art:

[0020] 1. Multi-index fusion and model optimization

[0021] By integrating the detection data of external physical and chemical characteristics (single fruit weight, outer diameter, inner diameter, girth, chromaticity) and internal physical and chemical characteristics (soluble solids, acidity), combined with models such as partial least squares regression, support vector machine (SVM), and random forest for comprehensive modeling. Compared with traditional single-index detection or single-model application, this method improves the generalization ability of the model through multi-dimensional data fusion. For example, in the outer diameter prediction of the SVM model, R 2 = 0.81, RPD = 2.30, which is significantly better than some existing chemical determination methods.

[0022] 2. High efficiency of non-destructive testing

[0023] By adopting multi-spectral technology combined with machine learning, non-destructive evaluation of the quality of powder bananas is realized. Compared with the existing method that relies on manual sampling and chemical determination, this method does not need to damage the fruits, has a faster detection speed, can simultaneously obtain external morphology and internal component information, and reduces the detection cost.

[0024] 3. Model selection and improvement of grading accuracy

[0025] By comparing the modeling effects of different models (SVM, random forest, etc.), the SVM model with the highest recognition accuracy is preferentially selected. In the prediction of external features (inner diameter, girth diameter, outer diameter), RPD>1.5 and R 2 >0.8, which is better than the models that only rely on spectral or image features in some existing studies.

[0026] 4. Potential evaluation ability of internal quality

[0027] Although the prediction effects of single fruit weight and pulp weight are limited, the SVM model shows feasibility in internal indexes such as soluble solids and acidity.

[0028] The following further describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. Description of the drawings

[0029] In the drawings:

[0030] Figure 1 It is a schematic diagram of the measurement positions of the external physical and chemical characteristics of powder bananas;

[0031] Figure 2 It is a schematic diagram of the image acquisition device;

[0032] Figure 3 In (a) is a change diagram of the average multi-spectral of powder bananas and the confidence interval; (b) is a normal distribution test diagram of the multi-spectral of powder bananas;

[0033] Figure 4 It is an apparent data distribution histogram of the lightness (L*) of the color of powder bananas and the red-green component (a*) of the color;

[0034] Figure 5 It is an apparent data distribution histogram of the yellow-blue component (b*) of the color of powder bananas and the chroma (c*) of the color;

[0035] Figure 6 It is an apparent data distribution histogram of the hue angle (h*) of the color of powder bananas and the soluble solids;

[0036] Figure 7Histogram of the apparent data distribution of the titratable acidity and single fruit weight of the powder banana color;

[0037] Figure 8 Histogram of the apparent data distribution of the outer diameter and inner diameter of the powder banana color;

[0038] Figure 9 Histogram of the apparent data distribution of the girth and peel weight of the powder banana color. Detailed implementation method

[0039] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments with reference to the accompanying drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention.

[0040] Embodiment 1:

[0041] As Figures 1 to 3 shown, a method for evaluating the internal and external physical and chemical characteristics of powder bananas based on an SVM multi-spectral model includes the following steps:

[0042] Step S1: Sample preparation. Select powder bananas with a maturity of 6 to 7. After placing the powder bananas in a constant temperature incubator at 30°C for 3 to 4 days, select powder bananas with a maturity of 8 to 9, and no pests, diseases, scars, deformities, or mechanical damage as samples;

[0043] Step S2: Determination of external physical and chemical characteristics. Use an electronic balance to measure the fresh weight of the fruit, and use a flexible tape measure to measure the outer diameter, inner diameter, and girth of the fruit; use a 3nh precision color difference meter NR20XE to measure the chromaticity at three points (upper, middle, and lower) of the selected sample, and take the average value;

[0044] Step S3: Determination of internal physical and chemical characteristics. Cut the banana into small pieces and homogenize it, then wrap it with sterile gauze and squeeze out the juice onto the induction mirror surface of the Atago PAL-BX / ACID6 special sugar-acid analyzer for bananas to measure the soluble solids content (Brix value) of the powder banana. Weigh about 1 g of banana sample with an electronic scale, put it into a 100 mL beaker, then add 50 mL of distilled water with a measuring spoon, stir evenly with a glass straw to fully dilute the sample, and suck an appropriate amount of the diluted acidity measurement sample and drop it onto the induction mirror surface of the instrument to obtain the acidity (Acid value) of the powder banana; repeat each measurement 3 times and take the average value;

[0045] Step S4: Image acquisition and processing. Connect the camera to the computer and adjust the parameters. Place the powder banana sample directly below the lens, and lay black flannel at the bottom of the dark box to avoid reflection; Image processing includes: image conversion and storage, image visual optimization, and image understanding;

[0046] Step S5: After performing feature extraction on the training samples by establishing the partial least squares discriminant method, establish a regression model between the independent variables and the dependent variables of the training samples, and then find the feature information that is most effective for classification;

[0047] Step S6: Characterize the deviation between the observed value and the true value, and the algorithm is as follows

[0048]

[0049] y i is the actual value, is the predicted value, and n is the number of samples.

[0050] As Figures 1 to 3 shown, in the specific implementation, in bananas, the banana acidity (Acid value) reflects the concentration of sugar in the banana juice. For example, a Brix value of 15% means that 100 grams of banana juice contains 15 grams of soluble solids (mainly sugar). The Acid value refers to the content of organic acids (such as citric acid, malic acid, etc.) in fruits, usually expressed as a percentage or an acidity coefficient (such as pH). In bananas, the main organic acids are citric acid and malic acid, and the acidity level affects the taste of bananas.

[0051] Furthermore, the image acquisition device mainly consists of a portable multispectral imaging camera, a light source, a dark box, and a computer. Among them, the portable multispectral imaging camera (Monarch Pro TM , Unispectral company), spectral range: 705 - 920 nm, spectral accuracy of ±2.5 nm; single-frame / spectral image cube; 60 frames per second in single-band mode, 25 frames per second in dual-band mode. Fixed in the center directly above the customized dark box for sample image acquisition, the distance between the camera lens and the sample is approximately 50 cm. The dark box is composed of black acrylic plates to block external light sources from entering, and its internal dimensions are 300×400×540 mm. The halogen cold light source (XD-302, Shanghai WenCe Optoelectronic Technology Co., Ltd.), with a built-in double-tube rigid 1-meter optical fiber and an adjustable condenser lens. The device is equipped with a computer with Monarch APP software for data acquisition. The physical diagram and schematic diagram of the device are respectively as Figure 2 shown.

[0052] Image processing software: In terms of software, use the self-written image acquisition software Monarch to acquire images of powder bananas, and use Matlab software to process and model analyze the images of powder bananas.

[0053] (1) Monarch Image Acquisition Software. This image acquisition software can connect to cameras simultaneously, obtain real-time display images, and can adjust camera parameters, and is well compatible with the Windows system. In the main interface of the software, it is divided into an image display module, a basic operation module, and a parameter adjustment module. Among them, the basic operation module includes initialization, screenshot, disconnection, and parameter update; the parameter settings are: exposure time 8000, gain 5dB, frame rate 150fps, brightness 10.

[0054] (2) Matlab Software. Matlab is developed by Mathworks, Inc. in the United States, has powerful computing functions, and comes with various toolboxes for development, matrix calculation, etc. Its built-in image processing toolbox has powerful image processing functions, and digital image processing can be achieved by writing code or using the functions in the toolbox.

[0055] Image Acquisition and Processing: Connect the camera to the computer and adjust the parameters. Place the plantain samples directly below the lens, and line the bottom of the dark box with black flannel to avoid reflection. A total of 113 images of plantain samples were collected. Images can be collected at 713, 736, 759, 782, 805, 828, 851, 874, 897, and 920 nm at one time. The resolution of each image is 1280×1024 pix, and the storage format is.png. Each sample is repeated 3 times at different positions.

[0056] Image processing can be roughly divided into three levels: (1) Image conversion and storage. It refers to image digitization, image spatial transformation, and image coding compression. (2) Image visual optimization. It is divided into two categories: image enhancement and image restoration, both of which are to improve the quality of the image. (3) Image understanding. The most critical steps in image understanding are image segmentation and feature extraction. The meaning of image segmentation is to separate the region of interest in the image from the background or other objects, which helps further analysis. Image feature extraction refers to extracting and identifying the useful information contained in the image, so as to judge the types and quantities of targets contained in the image and other effective information.

[0057] Data processing method: After extracting the features of the plantain images, it is very important to select an appropriate modeling method for data processing. The purpose of modeling is to enable the computer to automatically identify the research object and classify it, and to make the recognition rate as high as possible. In this study, two methods, support vector machine (SVM-C) and partial least squares discriminant analysis (PLS-DA), were selected. SVM-C: The support vector machine model is a binary linear classifier in machine learning methods. Its basic principle is to maximize the gap between support vectors after classification, that is, to find an optimal boundary to divide the data as accurately as possible. The support vector machine can not only be used in linear classification problems, but also be extended using kernel functions to make it applicable to non-linear classification problems. The kernel function outputs the data from a given space to a new high-dimensional space, and then uses a hyperplane to classify the data, which can solve the problems of high-dimensional feature classification and regression, and can also handle the case where the feature dimension is greater than the sample data. PLS-DA: Partial least squares discriminant method is a technique that finds the optimal classification function by minimizing the sum of squared errors. After extracting the features of the training samples, a regression model is established between the independent variables and the dependent variables of the training samples, and then the most effective feature information for classification is found.

[0058] After the model is established, some evaluation parameters are needed to measure the accuracy and precision of the model. In this study, the root mean square error (RMSE) is used to determine the model effect after feature parameter screening. It can be used to characterize the deviation between the observed value and the true value. Therefore, the smaller the RMSE, the better the model stability and the higher the precision.

[0059] External and internal physical and chemical characteristics: The external physical and chemical characteristics (single fruit weight, outer diameter, inner diameter, girth, chromaticity) and internal physical and chemical characteristics (soluble solids, acidity) of 113 plantains were measured. The specific values are shown in Table 2.

[0060] Image data processing results: The physical and chemical numerical distributions of all samples showed a strong central tendency, close to a normal distribution ( Figure 3 ), indicating that the obtained physical and chemical parameter data can effectively represent the physical and chemical states of the plantain varieties, and the construction of the next physical and chemical regression model can be carried out.

[0061] From Figure 4 (a), it can be seen that the average reflectance of the plantain varies in the range of 0.43 to 0.48, and shows an upward trend in the wavelength range of 713 nm to 851 nm, and has a slight decrease in the range of 851 nm to 920 nm. From Figure 5(b) It can be seen from the normal distribution diagrams of reflectance at different bands that the reflectance at all bands is concentrated around 0.48, with a variation range between 0.1 and 0.9, and the reflectance at all bands shows a good normal distribution, indicating that there is reasonable individual difference among the multispectral data of powder bananas, indicating that the next step of regression model construction can be carried out.

[0062] Experimental example:

[0063] Table 1 Comparison of the results of different regression models in evaluating the chromaticity of powder bananas

[0064]

[0065]

[0066] Compared with the traditional linear model - Partial Least Squares (PLS), the prediction result is poor in predicting the chromaticity of powder bananas, and the Residual Prediction Deviation (RPD) < 1.5. The non-linear machine learning methods and deep learning algorithms have better results in evaluating the chromaticity of powder bananas (Table 2). Among all the prediction results of powder banana chromaticity, the regression model constructed based on Support Vector Machine (SVM) obtained the best prediction result, with the prediction accuracy (Rv2) > 0.82 and RPD > 2.36. Among them, satisfactory results were obtained for the prediction of L*, a*, b* and h*, with prediction accuracies of 0.91, 0.85, 0.87 and 0.92 respectively, and the RPD values reached 3.40, 2.58, 2.82 and 3.46 respectively. The prediction result for c* is acceptable (Rv 2 = 0.82, RPD = 2.36).

[0067] The comparison of the results of different regression models in evaluating the sugar acidity and appearance of powder bananas is shown in Table 4. The results show that: in evaluating the soluble solids of powder bananas, the SVM model shows a certain feasibility (RPD > 1.5). However, the results of all models in predicting the titratable acidity of powder bananas are not ideal (RPD < 1.5). On the one hand, this is because the multispectral data obtained is from the surface of the powder banana peel, and the response of the peel to the spectrum and internal physical and chemical properties causes interference. On the other hand, it is because there is no significant C-H absorption peak in the range of 713 - 920 nm, resulting in unsatisfactory prediction results. In addition, the SVM model also shows a certain feasibility in predicting the size dimensions (inner diameter, perimeter) of powder bananas (RPD > 1.5), and a good result is obtained in predicting the outer diameter of powder bananas (Rv 2= 0.81, RPD = 2.30). However, in predicting the single fruit weight and pulp weight of plantains, the results were poor. This indicates that the SVM-based multispectral model can effectively evaluate the external physical and chemical characteristics (chromaticity, outer diameter) of plantains, and has a certain potential feasibility in evaluating the internal physical and chemical characteristics (soluble solids, single fruit weight) of plantains.

[0068] The interpretation of the characteristic bands based on the SVM model is as Figure 5 shown. It shows that using SHAP Values to assign weights to the characteristic bands of the SVM model with good prediction results, for the chromaticity data of plantains, the bands that have a greater impact on the model prediction results mainly concentrate on 713 nm, 736 nm, 782 nm, and 897 nm. Under these bands, the SHAP Value shows a large range of changes, indicating good discrimination and prediction effects on samples. In addition, the band that contributes the most to the prediction result of the outer diameter of plantains is 713 nm, and the range of its SHAP Values is -30 - 30, which is greater than the other bands, indicating that 713 nm can better distinguish and predict the outer diameter of plantains.

[0069] Table 2 Comparison of the results of different regression models in evaluating the sugar acidity and appearance of plantains

[0070]

[0071]

[0072]

[0073] Introduced the advanced technical means of Sylvie Bureau and her team from the French Institute for the Quality and Safety of Phytogenic Processing Products. Taking the main tropical fruit, plantain, as the research object, detected its external physical and chemical characteristics (single fruit weight, outer diameter, inner diameter, perimeter, chromaticity) and internal physical and chemical characteristics (soluble solids, acidity) quality indicators, and established a total of 2 grade discrimination models including partial least squares regression, support vector machine, random forest, decision tree, and one-dimensional convolutional neural network models. Compared the modeling effects of different models, and selected the model with the highest recognition accuracy to import into the software. The SVM model also showed a certain feasibility in predicting the size dimensions (inner diameter, perimeter) of plantains (RPD > 1.5), and obtained good results in predicting the outer diameter of plantains (Rv2 = 0.81, RPD = 2.30). However, in predicting the single fruit weight and pulp weight of plantains, the results were poor. This indicates that the SVM-based multispectral model can effectively evaluate the external physical and chemical characteristics (chromaticity, outer diameter) of plantains, and has a certain potential feasibility in evaluating the internal physical and chemical characteristics (soluble solids, single fruit weight) of plantains.

Claims

1. A method for evaluating the internal and external physical and chemical characteristics of banana based on a multispectral model of SVM, characterized in that: The steps include: Step S1: Sample preparation, select pink bananas with a maturity of 60% to 70%; place the pink bananas in a constant temperature box at 30°C for 3 to 4 days, and select pink bananas with a maturity of 80% to 90% and no pests, scars, deformities, or mechanical damage as samples; Step S2: Determination of external physical and chemical characteristics: using an electronic balance to measure the fresh weight of the fruit, using a soft tape measure to measure the outer diameter, inner diameter and circumference of the fruit; using a 3nh precision colorimeter NR20XE to measure the chromaticity of the upper, middle and lower points of the selected sample, and taking the average value; Step S3: determination of internal physical and chemical characteristics, the banana is cut into pieces and homogenized and wrapped with sterile gauze, the juice is squeezed out and the soluble solid content of the banana powder is determined on the sensing mirror of the Aituo PAL-BX / ACID6 banana-specific sugar-acid all-in-one machine; Step S4: Image acquisition and processing: connect the camera to the computer and adjust the parameters. The banana sample is placed directly below the lens. The bottom of the dark box is covered with black velvet to avoid reflection. Image processing includes: image conversion and storage, image visual optimization and image understanding. Step S5: After extracting features from the training samples by establishing a partial least squares discriminant method, a regression model between the independent variables and the dependent variables of the training samples is established, thereby finding the most effective feature information for classification; Step S6: Characterize the deviation between the observed value and the true value. The algorithm is as follows: y i is the actual value, is the predicted value, and n is the number of samples.

2. The method for evaluating the internal and external physical and chemical characteristics of banana based on SVM multispectral model according to claim 1 is characterized in that: In step S3, about 1 g of banana sample is weighed with an electronic scale and put into a 100 mL beaker, and then 50 mL of distilled water is added with a measuring spoon, and stirred evenly with a glass pipette to fully dilute the sample, and a proper amount of the diluted acidity measurement sample is dripped onto the induction mirror of the instrument to obtain the banana acidity; each is repeated 3 times and the average value is taken.

Images