Nondestructive testing method for bitter pox disease of apples in storage period and application of nondestructive testing method in early identification of bitter pox disease of apples

Abstract

The invention relates to the field of nondestructive testing of internal quality of fruits, in particular to a nondestructive testing method for bitter pox disease of apples in a storage period and application of the nondestructive testing method in early identification of the bitter pox disease of the apples. According to the method, early detection of the apple bitter pit disease is taken as a purpose, a hyperspectral imaging technology is used for acquiring images of apples in a storage period within a wave band of 400-1000 nm, and five preprocessing methods (SNV, MSC, SG, 1st-D and 2nd-D) and two characteristic wavelength selection methods (GA and RF) are used for establishing an SVM model based on spectral information; utilizing a second-order probability statistical filtering method to extract texture features of diseased and normal apples in the feature wavelength-based image, and establishing an SVM model based on the texture features; the accuracy of the model constructed by utilizing the characteristic wavelength selected by the RF algorithm is basically superior to that of the GA algorithm, and technical and theoretical references are provided for early diagnosis of the apple bitter pit disease.

Description

Technical Field

[0001] The invention belongs to the field of nondestructive detection of internal quality of fruits, and particularly relates to a nondestructive detection method for bitter pit disease of apples during storage. Background Art

[0002] Apple is one of the most popular fruits among consumers. The fruit is juicy, sweet and sour, and rich in various nutrients, such as vitamin C, antioxidants, cellulose, etc. It is an ideal choice before and after meals. However, bitter pit disease is very likely to occur during the post-harvest storage of apples. This disease seriously affects the quality of apples and consumer satisfaction. Moreover, in the early stages of bitter pit disease, the lesions only appear under the peel and cannot be directly observed. This makes it difficult to effectively identify infected fruits during the fruit screening process, causing these inferior fruits to flow into the market, thereby affecting consumer experience and brand image. However, traditional detection methods are time-consuming, labor-intensive, and destructive, which not only increases the cost of detection, but may also lead to the loss of the detected fruit, thereby reducing the overall economic benefits.

[0003] In recent years, hyperspectral imaging technology has been used to detect the quality indicators of various fruits. Hyperspectral imaging technology can realize non-destructive and efficient detection of different fruit types and multiple fruit quality indicators, such as maturity, sugar content, acidity, etc.; limited by hardware technology, its development focuses on data mining, that is, in the case of limited hardware development, accurate analysis results are obtained through targeted algorithms that are continuously updated and optimized. Patent 202311472395.0 discloses a method for detecting early hidden damage of Korla pears by integrating hyperspectral image features and spectral variables. Using hyperspectral imaging technology instruments, the spectral information and hyperspectral image texture feature information of healthy and three different degrees of damage Korla pears are obtained, and the idea of ​​spectral fusion is introduced to integrate the texture feature information with the spectral feature variable information, which is applied to the detection of early hidden damage of Korla pears. The overall accuracy of the correction set and the prediction set of the PLS-DA model based on full-wavelength spectral information are 96.7% and 93.4%, respectively; the discrimination accuracy of the RF-PLS-DA model established based on the full spectrum and feature variables exceeds 96.7%. Patent 201910468718.6 discloses a method for detecting citrus Huanglongbing based on hyperspectral imaging technology. The characteristic band reflectance, the characteristic band image texture features and the corresponding second principal component image texture features are fused to establish a least squares support vector machine discriminant model. This discriminant model has a good recognition effect for citrus Huanglongbing in different seasons, different orchards and different stages of infection. The recognition accuracy can reach 90% in months when symptoms are obvious or the carbohydrate content in the leaves is significantly different from that in healthy leaves, and can reach more than 85% in other months. Therefore, it is of great practical significance to develop an effective detection method using hyperspectral imaging technology to detect apples with bitter pit disease before consumption. Summary of the invention

[0004] In view of the above problems existing in the prior art, the present invention proposes a non-destructive detection method for bitter pit disease in apples during storage and its application in early identification of bitter pit disease in apples. The method solves the problem that bitter pit disease is prone to occur in apples during storage after harvest, and the lesions in the early stage of the disease only appear under the peel, which cannot be directly observed and is difficult to detect effectively.

[0005] The technical solution of the present invention is achieved in this way:

[0006] On the one hand, the present application proposes a non-destructive detection method for bitter pit disease in apples during storage, the steps are as follows:

[0007] (1) Collect a hyperspectral image of the bottom of the apple to be tested, a black background image, and a white calibration plate image respectively;

[0008] (2) using a black background image and a white calibration plate image to perform black-and-white correction on the acquired hyperspectral image to obtain a reflectance image, extracting the spectral information of the ROI area on the reflectance image and averaging it to obtain the average spectral information as the original spectral information;

[0009] (3) after removing the noise band from the original spectral information in step (2), preprocessing the remaining spectral information to obtain preprocessed spectral information;

[0010] (4) screening the spectral information of step (3) for characteristic wavelengths;

[0011] (5) The characteristic wavelengths in step (4) are synthesized into an image containing only the characteristic wavelengths, and the texture features of the ROI region on the image are extracted and substituted into the trained SVM model for discrimination to obtain a discrimination result.

[0012] Preferably, in the above step (1), the hyperspectral image acquisition system includes a light source, an imaging lens, an imaging spectrometer and an array detector. The light source is four adjustable 150W halogen lamps, which are respectively installed at an angle of 45° inside the dark box. The spectral range of the imaging spectrometer is 400-1000nm, and the spectral resolution is 3.5±0.5nm.

[0013] Preferably, the noise band in the above step (2) refers to a band range with larger noise at both ends.

[0014] Preferably, the preprocessing method in the above step (3) is any one of multivariate scatter correction (MSC), standard normal transformation (SNV), Savitzky-Gola filtering (SG), direct difference first-order derivative (1st-D) and direct difference second-order derivative (2nd-D).

[0015] Preferably, the characteristic wavelength in the above step (4) is screened by a genetic algorithm or a random frog leaping algorithm.

[0016] The genetic algorithm sets the total number of iterations N = 100 times, the initial population is 30, the crossover probability is 0.5, the mutation rate is 0.001, and the F test is performed based on the minimum RMSECV value. The high-frequency band with a confidence interval of more than 70% is output as the characteristic wavelength.

[0017] The random frog leaping algorithm sets the total number of iterations N = 1000, the number of principal components 10, the initial frog leaping population Q = 2, and only outputs the first 10 wavelengths as characteristic wavelengths based on the number of times each spectrum is selected during the iteration process.

[0018] Preferably, the image extraction method in the above step (5) is a second-order probability statistical filtering method; the texture features include mean, covariance, correlation, synergy, contrast, dissimilarity, second-order moment and entropy.

[0019] Preferably, the above SVM model is covariance-RF-SVM.

[0020] On the other hand, the above non-destructive testing method is used in the early identification of apple bitter pit disease.

[0021] The present invention has the following beneficial effects:

[0022] 1. This application uses hyperspectral imaging technology to perform non-destructive detection of bitter pit disease in stored apples, providing an important basis for post-harvest grading and storage quality control of bitter pit disease in apples; using images and spectral data of hyperspectral imaging to avoid the phenomenon of "same spectrum, different objects"; using characteristic wavelength selection methods to reduce data dimensions and improve detection efficiency.

[0023] 2. The best model for discriminating bitter pit in apples based on spectral information in this application is MSC-SVM, with the accuracy rates of the training set and test set being 98.1481% and 86.1111% respectively; among the eight texture-based features, the discriminating method using covariance is the best, with the accuracy rates of the training set and prediction set being 98.5507% and 93.3333% respectively. The accuracy rate of the model constructed using the characteristic wavelength selected by the RF algorithm is basically better than that of the GA algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 Flow chart of the nondestructive detection method for apple bitter pit disease.

[0026] Figure 2 The original spectral curve and five preprocessing effect diagrams; (a) is the original spectral curve, (b) is the spectral curve after standard normal transformation, (c) is the spectral curve after multivariate scattering correction, (d) is the spectral curve after Savitzky-Golay smoothing filter processing, (e) is the spectral curve after direct difference first-order derivative processing, and (f) is the spectral curve after direct difference second-order derivative processing.

[0027] Figure 3 Probability distribution plot for wavelength selection for the random frog leaping algorithm.

[0028] Figure 4 Frequency selection graph for genetic algorithm.

[0029] Figure 5 The figures are the validation results of 6 models; (a) is the validation result of the MSC-SVM model, (b) is the validation result of the MSC-GA-SVM model, (c) is the validation result of the MSC-RF-SVM model, (d) is the validation result of the contrast-RF-SVM model, (e) is the validation result of the covariance-RF-SVM model, and (f) is the validation result of the mean-RF-SVM model.

[0030] Figure 6 Apply the effects for the covariance-RF-SVM model. DETAILED DESCRIPTION

[0031] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] Unless otherwise specified, the experimental methods used in the following experimental examples are all conventional methods; the materials and reagents used are reagents and materials that can be obtained from commercial channels unless otherwise specified.

[0033] Example

[0034] A non-destructive detection method for bitter pit disease in apples during storage Figure 1 As shown, the steps are as follows:

[0035] 1. Experimental materials and methods

[0036] 150 Qincui apples of uniform size and good appearance were purchased from a Qincui apple orchard that was prone to bitter pit disease in previous years. They were placed in cartons filled with foam. Each carton had two layers, with 9 apples in each layer. The bottom of the foam box was numbered 1-150 with a black marker and stored in a 4℃ cold storage after packaging.

[0037] The hyperspectral images of the bottom of the apples were acquired using the GaiaSKy-mini-AZ drone hyperspectral imaging system equipped with a dark box and lighting system after 0, 15, 43, 58, and 73 days of storage. The hyperspectral imaging system consists of an imaging lens, an imaging spectrometer, and an array detector. The light source is four adjustable 150W halogen lamps, which are installed at 45° inside the dark box. The spectral range collected by the device is 400-1000nm, and the spectral resolution is 3.5±0.5nm. In order to reduce the influence of ambient scattered light and current, the entire shooting process was carried out in a darkroom. Note that each time the apples were taken out of the cold storage 12h in advance, put them at room temperature, and then photographed at room temperature. In each shot, it is necessary to first perform a fixed focus process, and then collect images of the black background and white calibration plate for obtaining reflectance data.

[0038] 2. Extraction of spectral information

[0039] After the five periods of shooting were completed, the skin of the diseased apples was peeled to confirm that there was no other disease information or defects that affected the spectral information. Then, the disease conditions of the apples of each number at different periods were compared to find out the period when the bitter pit spots did not appear on the skin and determine the location of the disease. The collected images of the diseased apples were corrected in black and white using SpecSight V1.2 software. The images taken at different periods were corrected using the black background and white calibration plate images collected on the same day to obtain the reflectance image. The data processing software was used to extract the spectral information of the diseased area on the corrected reflectance image and save the position. The pixels of the entire diseased area were averaged to obtain the average spectral information, which was used as the original spectral information.

[0040] 3. Spectral information preprocessing and characteristic wavelength screening

[0041] The extracted spectral information generates a spectral curve such as Figure 2 As shown, the bands with large noise at both ends are removed. The spectral information after deleting the noise band is preprocessed by multivariate scatter correction (MSC), standard normal transformation (SNV), Savitzky-Gola filtering (SG), direct difference first-order derivative (1st-D), and direct difference second-order derivative (2nd-D); the preprocessed spectral information is input into the support vector machine (SVM) model, and the five preprocessing effects are compared. The optimal preprocessing method is selected for subsequent feature selection and model establishment. The comparison results are shown in Table 1.

[0042] Table 1 Results of full-spectrum SVM classification model established based on five preprocessing methods

[0043]

[0044] Table 2 Characteristic wavelengths selected based on two characteristic wavelength methods

[0045]

[0046] Genetic algorithm (GA) and random frog leaping (RF) algorithms are used to screen characteristic wavelengths to reduce redundant information and improve computational efficiency. For the RF algorithm, the total number of iterations N = 1000, the number of principal components 10, the initial frog leaping population Q = 2, and the number of times each spectrum is selected during the iteration process is used as the basis, and only the first 10 wavelengths are output as characteristic wavelengths, such as Figure 3 As shown, the part above the red line is the characteristic wavelength used in the present invention. For the genetic algorithm, the total number of iterations N = 100, the initial population is 30, the crossover probability is 0.5, the mutation rate is 0.001, and the F test is performed based on the minimum RMSECV value as the standard. The high-frequency band with a confidence interval of more than 70% is output as the characteristic wavelength. The frequency selection diagram of the genetic algorithm is shown in FIG. Figure 4 As shown in Figure 2, the green line is the characteristic wavelength selected based on the minimum RMSECV standard, and the red line is the characteristic wavelength selected based on the F test. The characteristic wavelength results selected by the two preprocessing methods are shown in Table 2.

[0047] 4. Extract texture features

[0048] According to the characteristic wavelengths screened out by the above two characteristic wavelengths, an image containing only the characteristic wavelengths is generated, and the texture features of the region of interest saved when the spectral information is extracted are extracted. The second-order probability statistical filtering method is used for extraction, and all the texture features contained therein are extracted, including mean, covariance, correlation, homogeneity, contrast, dissmilanity, second moment, and entropy.

[0049] 5. Division of sample sets

[0050] During the whole shooting period, 5 periods of images were acquired, with 150 images in each period, for a total of 750 images. After comparison, the spectral information of 90 samples (45 diseased apple samples and 45 non-diseased apple samples) and the texture features of 114 samples (57 diseased apple samples and 57 non-diseased apple samples) were extracted. The training set and the test set were divided into a 3:2 ratio by random partitioning, where the ratio of samples used to construct the spectral information was training set: test set = 54:36; the ratio of samples used to construct the characteristic wavelength-texture feature was training set: test set = 69:45, and the diseased samples were classified as 0 and the normal samples were classified as 1 for the calculation of the accuracy of the model classification results.

[0051] 6. Construction of classification model

[0052] The spectral information screened by the optimal preprocessing-characteristic wavelength and the texture information based on the characteristic wavelength image are input into the SVM classification model for discrimination, and the radial basis function (RBF) kernel function is selected to construct the SVM model. Then, the optimal combination of c and g is determined through grid optimization and multiple cross-validation results.

[0053] Table 3 Classification results of SVM model based on spectral and texture features

[0054]

[0055] The results of the classification models established based on the two types of information are shown in Table 3. The best model for identifying apple bitter pit based on spectral information is MSC-SVM, with accuracies of 98.1481% and 86.1111% for the training set and test set, respectively. Among the eight texture-based features, the use of covariance to distinguish diseased and normal apples is the best, with accuracies of 98.5507% and 93.3333% for the training set and prediction set, respectively. Six models with higher accuracies, namely MSC-SVM, MSC-GA-SVM, MSC-RF-SVM, contrast-RF-SVM, covariance-RF-SVM and mean-RF-SVM, were selected for verification. The verification results of the six models are shown in Table 3. Figure 5 ; The accuracy of the model constructed using the characteristic wavelength selected by the RF algorithm is basically better than that of the GA algorithm.

[0056] Application Examples

[0057] In order to verify the robustness and accuracy of the constructed model, 50 Qincui apples were picked on September 25, 2024, numbered and stored in a 4℃ cold storage, and hyperspectral images were collected on September 26, October 11, and November 11, 2024, respectively. The instruments and collection environment were consistent with the previous year. The spectral information and texture information of all samples were extracted as a validation set to evaluate the accuracy of the model. The specific steps are as follows:

[0058] The collected spectral information is preprocessed by multivariate scatter correction (MSC), and the preprocessed spectral information is input into the random frog leaping (RF) algorithm to screen the characteristic wavelengths.

[0059] The screened characteristic wavelengths are synthesized into an image containing only the characteristic wavelengths, and the texture features of the ROI area on the image are extracted and input into the SVM model to discriminate the samples; based on the accuracy of the model constructed in the previous year, the covariance-RF-SVM model is selected to discriminate the samples.

[0060] Table 4 Evaluation and verification results

[0061]

[0062] The true class results of this application were obtained through phenotypic observations at the late stage of disease onset ( Figure 6 ), the evaluation results of this application model are shown in Figure 6 ,The results show that there are 25 diseased samples and 10 healthy samples, ,and the model recognition accuracy is 82.8571%, ,which proves that the detection method of this application is feasible.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A non-destructive detection method for bitter pit disease in apples during storage, characterized in that: Here are the steps: (1) Collect the hyperspectral image of the bottom of the apple to be tested, the black background image, and the white calibration plate image respectively; (2) Use the black background image and the white calibration plate image to perform black and white correction on the acquired hyperspectral image to obtain a reflectance image, extract the spectral information of the ROI area on the reflectance image, and average it to obtain the average spectral information as the original spectral information; (3) After removing the noise band from the original spectral information in step (2), preprocessing the remaining spectral information is performed to obtain preprocessed spectral information; (4) screening the spectral information of step (3) for characteristic wavelengths; (5) The characteristic wavelengths in step (4) are synthesized into an image containing only the characteristic wavelengths, and the texture features of the ROI area on the image are extracted. The texture features are substituted into the trained SVM model for discrimination to obtain the discrimination result.

2. The nondestructive detection method for bitter pit disease of apples during storage according to claim 1, characterized in that: In step (1), the hyperspectral image acquisition system includes a light source, an imaging lens, an imaging spectrometer and an array detector. The light source is four adjustable 150 W halogen lamps, which are installed at 45° in the dark box. The spectral range of the imaging spectrometer is 400-1000 nm, and the spectral resolution is 3.5 ± 0.5 nm.

3. The nondestructive detection method for bitter pit disease of apples during storage according to claim 1, characterized in that: The noise band in step (3) refers to the band range with larger noise at both ends.

4. The nondestructive detection method for bitter pit disease of apples during storage according to claim 1, characterized in that: The preprocessing method in step (3) is any one of multivariate scatter correction, standard normal transformation, Savitzky-Gola filtering, direct difference first-order derivative and direct difference second-order derivative.

5. The nondestructive detection method for bitter pit disease of apples during storage according to claim 1, characterized in that: In the step (4), the characteristic wavelength is selected by a genetic algorithm or a random frog leaping algorithm.

6. The nondestructive detection method for bitter pit disease of apples during storage according to claim 5, characterized in that: The genetic algorithm sets the total number of iterations N=100 times, the initial population is 30, the crossover probability is 0.5, the mutation rate is 0.001, and the F test is performed based on the minimum RMSECV value as the standard, and the high-frequency wavelength with a confidence interval of more than 70% is output as the characteristic wavelength.

7. The nondestructive detection method for bitter pit disease of apples during storage according to claim 5, characterized in that: The random frog leaping algorithm sets the total number of iterations N=1000, the number of principal components 10, the initial frog leaping population Q=2, and outputs only the first 10 wavelengths as characteristic wavelengths based on the number of times each spectrum is selected during the iteration process.

8. The nondestructive detection method for bitter pit disease of apples during storage according to claim 1, characterized in that: The image extraction method in step (5) is a second-order probability statistical filtering method; the texture features include mean, covariance, correlation, synergy, contrast, dissimilarity, second-order moment and entropy.

9. The nondestructive detection method for bitter pit disease of apples during storage according to claim 1, characterized in that: The SVM model is covariance-RF-SVM.

10. Use of the nondestructive testing method according to any one of claims 1 to 9 in early identification of apple bitter pit disease.

Images