Nondestructive detection method and system for comprehensive quality of postharvest bananas

Abstract

The application discloses a nondestructive detection method and system for comprehensive quality of bananas after harvesting. The method obtains a hyperspectral image, spectral characteristics, an infrared radiation energy distribution image, a visible light image and dielectric characteristic parameters of bananas, thereby constructing a high-precision prediction model. The prediction model is used for subsequent nondestructive detection of target bananas, real-time monitoring of the quality of bananas after postharvest cold storage, accurate evaluation of the maturity grade of bananas after cold storage and ripening, and grading and sorting of the bananas.

Description

Technical Field

[0001] This invention relates to the field of fruit testing technology, and in particular to a non-destructive testing method and system for the comprehensive quality of bananas after harvest. Background Technology

[0002] Bananas, as a typical climacteric fruit, continue their active metabolic process even after harvesting. To maintain their freshness and quality, pre-cooling is essential. Pre-cooling allows for appropriate processing steps based on fruit ripeness levels, which helps improve fruit quality.

[0003] The quality testing of bananas after cold storage is divided into two types: external and internal quality testing.

[0004] External quality assessment refers to the evaluation of aspects such as color, freshness, size, mechanical damage, frostbite, and rot. Traditional machine vision technology, due to its low precision and complex operation, struggles to distinguish external characteristics such as mechanical damage, frostbite, rot, and freshness in the detection of fruit and vegetable external quality. Hyperspectral imaging technology, however, enables comprehensive, non-destructive testing with high precision and ease of operation, and has been increasingly used in the detection of fruit and vegetable external quality in recent years. For example, Xie et al. proposed a detection scheme entitled "Predicting Banana Color and Firmness Using Novel Wavelength Selection." This scheme uses hyperspectral imaging technology to analyze banana color to assess its freshness and ripeness. Freshly picked green bananas were ripened in a laboratory, and color data was collected over time. A partial least squares (PLS) model was used to predict freshness, yielding good prediction results and achieving non-contact, non-destructive detection of banana color and freshness.

[0005] Internal quality is an important basis for measuring the nutritional value of fruits and vegetables. It is generally judged by testing indicators such as firmness, soluble solids content (SCC), titratable acidity, moisture, ripeness, protein, and starch content. Among them, sugar content and firmness are two important indicators reflecting the internal quality of fruits and vegetables. Sugar content reflects the taste of fruits and vegetables, while firmness indirectly reflects the ripeness of the fruit. Rajkumar et al. proposed a detection scheme for "Hyperspectral Imaging Study on Banana Fruit Quality and Maturity". They used visible and near-infrared (400-1000 nm) hyperspectral imaging technology to detect the moisture content, soluble solids content, and firmness of bananas under different temperature conditions. Partial least squares analysis was used to process the spectral data, and principal component analysis was used to obtain the characteristic bands with the highest contribution rate in the entire spectrum. A prediction model based on the characteristic bands was established by combining multiple linear regression. The correlation coefficients of soluble solids content, moisture content, and firmness of bananas were found to be 0.85, 0.87, and 0.91, respectively. At the same time, it was found that the changes in soluble solids content and firmness of bananas under different temperature environments showed a polynomial relationship with their maturity, while the changes in moisture content showed a linear relationship with maturity.

[0006] However, neither of the two detection schemes proposed above constructs a comprehensive quality index for bananas. They only describe the detection of the external and internal quality of bananas. Machine vision can perform non-destructive testing of the external quality of bananas, while visible-near infrared spectroscopy and hyperspectral imaging technology can perform non-destructive testing of the internal quality of bananas. However, the prediction model only performs testing on a single quality index, and a single quality index cannot characterize the comprehensive quality of bananas after harvest. Therefore, the detection effect and significance are not great. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes a non-destructive testing method and system for the comprehensive quality of bananas after harvest, which mainly solves the problem that existing banana quality prediction models cannot characterize the comprehensive quality characteristics of samples.

[0008] To address the aforementioned technical problems, the first aspect of this invention proposes a non-destructive testing method for the comprehensive quality of bananas after harvest, applicable to bananas after cold storage, comprising the following steps:

[0009] S1, acquire several hyperspectral images of bananas, and extract the spectral features of preset points in the hyperspectral images;

[0010] S2, acquire infrared radiation energy distribution images of several bananas, and convert the infrared radiation energy distribution images into visible light images;

[0011] S3, obtain the dielectric properties of several bananas;

[0012] S4, perform data-level fusion of the hyperspectral image, the spectral features, the infrared radiation energy distribution image, the visible light image, and the dielectric property parameters, define them as sample data, associate the sample data with the maturity index, construct a prediction model, select feature variables for fusion, and define them as sample feature variables;

[0013] S5. Based on the prediction model, the sample feature variables and the maturity index are fused at the decision level. Finally, a batch of samples are selected from the sample feature variables for model training, optimization and deep analysis to complete the training task of the prediction model.

[0014] S6. For a single target banana, repeat the data collection tasks of S1-S3 to form real-time sample data and input it into the prediction model to obtain the corresponding real-time maturity index. Use the real-time maturity index as the grading condition for banana maturity.

[0015] The region of interest in the hyperspectral image is selected, and the region of interest is subjected to spectral preprocessing to extract the average spectrum as the spectral feature.

[0016] In some embodiments, the maturity indicators include at least soluble solids content, titratable acid, sugar-acid ratio, and / or hardness.

[0017] In some implementations, the infrared radiation energy distribution image is converted into a visible light image through spectral conversion and brightness enhancement. The RGB three primary color channel pixel values ​​of the visible light image are extracted, and then the hue, saturation, and brightness pixel values ​​of the visible light image are extracted using a color model. The sample texture of the visible light image is also extracted to estimate the surface area of ​​the banana.

[0018] In some embodiments, the dielectric characteristic parameters include complex impedance and complex dielectric constant, wherein the complex impedance is a physical constant of the banana, and the complex dielectric constant is used to characterize the internal quality characteristics of the banana.

[0019] In some implementations, the method further includes S7, where banana maturity levels are pre-set, and after the banana maturity levels are subjected to semi-supervised deep learning and representation learning by a prediction model, the input real-time sample data is classified according to the hierarchical structure of the banana maturity levels, and then the target bananas are transported to a designated channel for packing by a sorting mechanism.

[0020] The second aspect of this invention provides a post-harvest comprehensive quality non-destructive testing system for bananas, applied to bananas after cold storage, comprising:

[0021] A hyperspectral camera is used to acquire hyperspectral images of several bananas and extract the spectral features of preset points in the hyperspectral images;

[0022] An infrared thermal imaging camera is used to acquire infrared radiation energy distribution images of several bananas and convert the infrared radiation energy distribution images into visible light images.

[0023] Intelligent electrical parameter detection unit, used to obtain dielectric property parameters of several bananas;

[0024] The data fusion unit is used to perform data-level fusion of the hyperspectral image, the spectral features, the infrared radiation energy distribution image, the visible light image, and the dielectric property parameters, defining them as sample data; to associate the sample data with the mature index, construct a prediction model, and select feature variables for fusion, defining them as sample feature variables.

[0025] The model training unit is used to perform decision-level fusion of the sample feature variables and the maturity index based on the prediction model, and finally select a batch of samples from the sample feature variables for model training, optimization and deep analysis to complete the training task of the prediction model.

[0026] The maturity grading unit is used to re-collect data for a single target banana, generate real-time sample data, and input it into the prediction model to obtain the corresponding real-time maturity index, which is then used as the grading condition for banana maturity.

[0027] A third aspect of this invention proposes a post-harvest comprehensive quality non-destructive testing system for bananas, applied to bananas after cold storage. The system includes: a first experimental dark chamber, a second experimental dark chamber, and a conveyor belt that passes sequentially through the first and second experimental dark chambers. An intelligent electrical parameter detection unit is installed at the end of the conveyor belt. A lifting platform is installed between the first and second experimental dark chambers. A hyperspectral camera is installed on one side of the lifting platform. The hyperspectral camera is connected to the first experimental dark chamber via optical fiber. The first experimental dark chamber contains a built-in halogen light source, and the second experimental dark chamber contains a ring light source, an infrared thermal imaging camera, a bar light source, and an electrical parameter measuring station. The data output terminals of the hyperspectral camera, the infrared thermal imaging camera, and the electrical parameter measuring station are connected to the data receiving terminal of a computer.

[0028] In some embodiments, a weighing platform is provided at the bottom of the first experimental dark chamber.

[0029] In some embodiments, a weighing device is also provided at the end of the conveyor belt.

[0030] The beneficial effects of this invention are as follows: by acquiring the hyperspectral image, spectral characteristics, infrared radiation energy distribution image, visible light image, and dielectric property parameters of bananas, a high-precision prediction model can be constructed. This prediction model can then be used to perform non-destructive testing on the target bananas, monitor the quality of bananas after cold storage in real time, accurately assess the maturity level of banana samples after cold storage ripening, and grade and sort them. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the non-destructive testing method for comprehensive post-harvest quality of bananas disclosed in Embodiment 1 of the present invention.

[0032] Figure 2 This is a schematic diagram of the structure of the banana post-harvest comprehensive quality non-destructive testing system disclosed in Embodiment 3 of the present invention;

[0033] Figure 3 for Figure 2 Schematic diagram of the cross-sectional structure of the inner AA region;

[0034] Figure 4 for Figure 2 Schematic diagram of the cross-sectional structure of the inner BB region;

[0035] Figure 5A schematic diagram of the spectral preprocessing results of the region of interest in the hyperspectral image in Embodiment 1 or 3 of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the content of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to this invention are shown in the accompanying drawings, not all of them.

[0037] Example 1

[0038] This embodiment proposes a non-destructive testing method for comprehensive post-harvest quality of bananas, applied to bananas after cold storage. By acquiring hyperspectral images, spectral characteristics, infrared radiation energy distribution images, visible light images, and dielectric property parameters of bananas, a high-precision prediction model is constructed. This prediction model is then used to perform non-destructive testing on the target bananas, monitor the quality of bananas after cold storage in real time, accurately assess the maturity level of banana samples after cold storage ripening, and grade and sort them.

[0039] like Figure 1 As shown, this method includes the following steps:

[0040] S1. Collect several hyperspectral images of bananas and extract the spectral features of preset points in the hyperspectral images.

[0041] Select the region of interest (ROI) from the hyperspectral image, perform spectral preprocessing on the ROI, and extract the average spectrum as spectral features, such as... Figure 5 As shown.

[0042] S2, acquire infrared radiation energy distribution images of several bananas, and convert the infrared radiation energy distribution images into visible light images;

[0043] Specifically, the system relies on the near-infrared radiation emitted by near-infrared radiation sources (such as the sun or halogen lamps) reflected by the target object. A near-infrared camera then receives this radiation, forming an infrared radiation energy distribution image of the target object. This image undergoes spectral conversion and brightness enhancement to convert it into a visible light image. This allows for real-time acquisition of both infrared and true-color images of bananas after cold storage or cold-ripening. A machine vision system then extracts the RGB (Red, Green, Blue) primary color channel pixel values ​​from the visible light image. The HSV (Hue, Saturation, Value) color model is then used to extract the hue, saturation, and brightness pixel values ​​from the visible light image, along with sample texture extraction to estimate the banana's surface area. After threshold binarization and edge contour extraction, the image's grayscale value is obtained.

[0044] S3, obtain the dielectric properties of several bananas.

[0045] Dielectric properties include complex impedance and complex dielectric constant. Complex impedance is a physical constant of bananas, while complex dielectric constant (which integrates current, voltage, and power factor) is used to characterize the internal quality characteristics of bananas.

[0046] S4 involves data-level fusion of hyperspectral images, spectral features, infrared radiation energy distribution images, visible light images, and dielectric property parameters, defining them as sample data. The sample data is then correlated with mature indicators to construct a prediction model. Feature variables are selected and fused, defining them as sample feature variables.

[0047] In this embodiment, the data-level fusion refers to fusing parameters such as the RGB primary colors, HSV color model (brightness, saturation, and sharpness), image texture and grayscale values, surface area, perimeter, and estimated volume of the sample's region of interest with thermal imaging temperature field, 400-1000nm hyperspectral band features, and dielectric properties including tube current and tube voltage. The aforementioned maturity indicators include at least soluble solids content, titratable acidity, sugar-acid ratio, and / or hardness.

[0048] S5, based on the prediction model, performs decision-level fusion of sample feature variables and mature indicators, and finally selects a batch of samples from the sample feature variables for model training, optimization and deep analysis to complete the training task of the prediction model.

[0049] In this embodiment, the decision-level fusion refers to using methods such as SPA (continuous projection method) or GA (genetic algorithm) to select the sample feature variables after data fusion and perform quantitative analysis on the correlation between the mature indicators (soluble solids content, titratable acid, water content and edible rate, etc.). The purpose is to reduce the dimensionality of the sample feature variables after data-level fusion in order to find data that are strongly correlated with the corresponding mature indicators.

[0050] The aforementioned model training, optimization, and deep analysis address the fact that changes in soluble solids content, titratable acidity, moisture content, and edibility during sample storage inevitably affect overall quality, including maturity. In this embodiment, the CNN-LSTM-Attention_Adaboost prediction model is used to predict sample maturity through multivariate loading and to perform quality grading and sorting based on maturity. Specifically, the CNN-LSTM-Attention-Adaboost model combines the attention mechanism of a convolutional long short-term memory neural network with AdaBoost multivariate time series prediction (loading prediction). This attention mechanism, spatiotemporal feature fusion, and ensemble model learning prediction constitute the CNN-LSTM-Attention-Adaboost multivariate loading prediction. CNN-LSTM-Attention-AdaBoost is a method that combines CNN-LSTM-Attention and AdaBoost machine learning techniques to improve model performance and robustness. Specifically, AdaBoost is an ensemble learning method that combines multiple weak learners into a strong learner, where each learner is trained on different datasets and feature representations. The basic idea of ​​the CNN-LSTM-Attention-AdaBoost algorithm is to use CNN-LSTM-Attention as a base model and enhance it using the AdaBoost algorithm. Specifically, multiple CNN-LSTM-Attention models can be trained, each using a different dataset and feature representations, and then their predictions can be combined to form a more accurate and robust model.

[0051] S6 involves repeating the data collection tasks from S1 to S3 for a single target banana, generating real-time sample data, which is then input into the prediction model to obtain the corresponding real-time maturity index. This real-time maturity index is used as the grading condition for banana maturity.

[0052] It also includes S7, which pre-sets banana maturity levels. After the banana maturity levels are semi-supervised by the prediction model for deep learning and representation learning, the input real-time sample data is classified according to the hierarchical structure of banana maturity levels. Then, the target bananas are transported to the designated channel for packing by the sorting mechanism.

[0053] In summary, this embodiment provides a rapid, non-destructive, and stable banana quality grading and sorting detection solution, enabling intelligent grading and sorting of bananas after harvest, thereby enhancing their market competitiveness and added value. Simultaneously, it saves significant labor and economic costs, and the accuracy, stability, and specificity of post-harvest banana quality grading and sorting are superior to manual grading and sorting. As the model database sample size increases, the model algorithm continuously trains, optimizes, and validates the training set, improving the accuracy, stability, and specificity of the model's prediction, grading, and sorting of different banana varieties, thus enhancing the model's versatility. Furthermore, the model's ability to predict the overall quality of samples far surpasses previous methods for detecting single samples and single qualities, resulting in more scientific, effective, and comprehensive detection results.

[0054] Example 2

[0055] This embodiment proposes a post-harvest comprehensive quality non-destructive testing system for bananas, applied to bananas after cold storage, including:

[0056] A hyperspectral camera is used to acquire hyperspectral images of several bananas and extract the spectral features of preset points in the hyperspectral images.

[0057] An infrared thermal imaging camera is used to acquire infrared radiation energy distribution images of several bananas and convert the infrared radiation energy distribution images into visible light images.

[0058] Intelligent electrical parameter detection unit, used to obtain dielectric property parameters of several bananas;

[0059] The data fusion unit is used to perform data-level fusion of hyperspectral images, spectral features, infrared radiation energy distribution images, visible light images, and dielectric property parameters, which are defined as sample data. The sample data is then correlated with mature indicators to build a prediction model. Feature variables are selected and fused, and defined as sample feature variables.

[0060] The model training unit is used to perform decision-level fusion of sample feature variables and maturity indicators based on the prediction model, and finally select a batch of samples from the sample feature variables for model training, optimization and deep analysis to complete the training task of the prediction model.

[0061] The maturity grading unit is used to re-collect data for a single target banana, generate real-time sample data, and input it into the prediction model to obtain the corresponding real-time maturity index, which is then used as the grading condition for banana maturity.

[0062] The components in this embodiment can be referred to as S1-S6 in Embodiment 1, and will not be repeated here.

[0063] Example 3

[0064] This embodiment proposes a non-destructive testing system for comprehensive post-harvest quality of bananas, applicable to bananas after cold storage, such as... Figure 2 As shown, the system includes: a first experimental dark box 13, a second experimental dark box 20, and a conveyor belt 7 that passes through the first experimental dark box 13 and the second experimental dark box 20 in sequence. An intelligent electrical parameter detection unit 9 is also provided at the end of the conveyor belt 7. A lifting platform 1 is provided between the first experimental dark box 13 and the second experimental dark box 20. A hyperspectral camera 2 is provided on one side of the lifting platform 1. The hyperspectral camera 2 is connected to the first experimental dark box 13 through an optical fiber. The first experimental dark box 13 has a built-in halogen light source 15. The second experimental dark box 20 has a built-in ring light source 18, an infrared thermal imaging camera 19, a strip light source 21, and an electrical parameter measuring station 22. The data output terminals of the hyperspectral camera 2, the infrared thermal imaging camera 19, and the electrical parameter measuring station 22 are connected to the data receiving terminal of the computer 3.

[0065] Furthermore, a weighing platform 16 is provided at the bottom of the first experimental dark box 13, and a weighing device 8 is also provided at the end of the conveyor belt 7.

[0066] Instructions for use: Turn on the power supply 5. The hyperspectral camera 2 and conveyor belt 7 are connected to the computer 3 via USB cable 4. The conveyor belt 7 is driven by motor 6, transporting the banana sample 17 to the first experimental dark chamber 13. The first experimental dark chamber 13 has a built-in weighing platform 16, which has a built-in weighing sensor below the conveyor belt 7. The weighing sensor is connected to an external weighing device 8, which can transmit the mass information of the banana sample 17 to the computer 3. The intelligent electrical parameter detection unit 9 is connected to the conveyor belt 7 and is used in conjunction with the electrical parameter measuring platform 22 in the second experimental dark chamber 20.

[0067] In a specific implementation plan, such as Figure 3 As shown, the first experimental dark chamber 13 contains a cooling fan 14, a halogen light source 15, and a weighing platform 16. The halogen light source 15 is arranged inside the first experimental dark chamber 13 at a 45° illumination angle. The cooling fan 14 is located on both sides of the bottom of the first experimental dark chamber 13. The experimental frame 11 is threadedly connected to the lifting platform 1, and the lifting platform 1 is threadedly connected to the hyperspectral camera 2. According to the experimental requirements, the lifting platform 1 is used to control the height of the hyperspectral camera 2. The hyperspectral camera 2 is used in conjunction with the hyperspectral camera lens 10 to adjust parameters such as focal length. An optical fiber 12 is used to connect to the interior of the first experimental dark chamber 13 to obtain a hyperspectral image of the banana sample 17.

[0068] In a specific implementation plan, such as Figure 4As shown, the second experimental darkroom 20 contains a ring light source 18, an infrared thermal imaging camera 19, a bar light source 21, and an electrical parameter measuring station 22. The infrared thermal imaging camera 19 is connected to the computer 3 via a USB cable and can acquire information such as infrared radiation energy distribution images and visible light images of the banana sample 17. The electrical parameter measuring station 22, in conjunction with the intelligent electrical parameter detection unit 9, acquires the dielectric property parameters of the banana sample 17 and transmits them to the computer 3.

[0069] Computer 3 was equipped with FigSpec spectral acquisition software to acquire the spectral information of banana sample 17 in the 400-1000nm wavelength range, i.e., the hyperspectral image, and the spectral data of any point in the image in the 400-1000nm range. The region of interest in the hyperspectral image was selected, and through spectral preprocessing, the average spectrum was extracted as the sample's spectral feature information, such as... Figure 5 As shown.

[0070] For more detailed testing steps, please refer to Example 1.

[0071] This solution not only accurately classifies cold-stored bananas but also reduces the workload of operators, lowers labor costs, and enables batch post-harvest quality testing of bananas, thereby enhancing their market competitiveness. Furthermore, the testing methods employed in this solution are universal and can be used for non-destructive testing of the overall quality of various types of bananas on the market; they are effective not only for fresh bananas but also for cooked bananas.

[0072] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A non-destructive testing method for comprehensive post-harvest quality of bananas, applied to bananas after cold storage, characterized in that, Includes the following steps: S1, acquire several hyperspectral images of bananas, and extract the spectral features of preset points in the hyperspectral images; Select the region of interest in the hyperspectral image, perform spectral preprocessing on the region of interest, and extract the average spectrum as the spectral feature; S2, acquire several infrared radiation energy distribution images of bananas, and convert the infrared radiation energy distribution images into visible light images; the infrared radiation energy distribution images are converted into visible light images after spectral conversion and brightness enhancement, the RGB three primary color channel pixel values ​​of the visible light images are extracted, and then the hue, saturation and brightness pixel values ​​of the visible light images are extracted using a color model, and the sample texture of the visible light images is extracted to estimate the surface area of ​​the bananas; S3, Obtain several dielectric property parameters of bananas; the dielectric property parameters include complex impedance and complex dielectric constant, wherein the complex impedance is a physical constant of bananas, and the complex dielectric constant is used to characterize the internal quality characteristics of bananas. S4, the hyperspectral image, the spectral features, the infrared radiation energy distribution image, the visible light image, and the dielectric property parameters are fused at the data level and defined as sample data. The sample data is then correlated with maturity indicators to construct a prediction model. Feature variables are selected and fused and defined as sample feature variables. The prediction model is a CNN-LSTM-Attention_Adaboost prediction model. The maturity indicators include at least soluble solids content, titratable acid, sugar-acid ratio, and / or hardness. S5. Based on the prediction model, the sample feature variables and the maturity index are fused at the decision level. Finally, a batch of samples are selected from the sample feature variables for model training, optimization and deep analysis to complete the training task of the prediction model. S6. For a single target banana, repeat the data collection tasks of S1-S3 to form real-time sample data and input it into the prediction model to obtain the corresponding real-time maturity index. Use the real-time maturity index as the grading condition for banana maturity.

2. The non-destructive testing method for comprehensive post-harvest quality of bananas as described in claim 1, characterized in that, It also includes S7, which pre-sets banana maturity levels. After the banana maturity levels are subjected to semi-supervised deep learning and representation learning by the prediction model, the input real-time sample data is classified according to the hierarchical structure of the banana maturity levels. Then, the target bananas are transported to the designated channel for packing by the sorting mechanism.

3. A non-destructive testing system for comprehensive post-harvest quality of bananas, applied to bananas after cold storage, characterized in that, include: A hyperspectral camera is used to acquire hyperspectral images of several bananas and extract the spectral features of preset points in the hyperspectral images; Select the region of interest in the hyperspectral image, perform spectral preprocessing on the region of interest, and extract the average spectrum as the spectral feature; An infrared thermal imaging camera is used to acquire infrared radiation energy distribution images of several bananas, convert the infrared radiation energy distribution images into visible light images; the infrared radiation energy distribution images are converted into visible light images through spectral conversion and brightness enhancement, the RGB three primary color channel pixel values ​​of the visible light images are extracted, and then the hue, saturation and brightness pixel values ​​of the visible light images are extracted using a color model, and the sample texture of the visible light images is extracted to estimate the surface area of ​​the bananas; An intelligent electrical parameter detection unit is used to acquire several dielectric characteristic parameters of bananas; the dielectric characteristic parameters include complex impedance and complex dielectric constant, wherein the complex impedance is a physical constant of the banana, and the complex dielectric constant is used to characterize the internal quality characteristics of the banana. The data fusion unit is used to perform data-level fusion of the hyperspectral image, the spectral features, the infrared radiation energy distribution image, the visible light image, and the dielectric property parameters, defining them as sample data. The sample data is then correlated with mature indicators to construct a prediction model. Selected feature variables are fused and defined as sample feature variables. The prediction model is a CNN-LSTM-Attention_Adaboost prediction model. The mature indicators include at least soluble solids content, titratable acid, sugar-acid ratio, and / or hardness. The model training unit is used to perform decision-level fusion of the sample feature variables and the maturity index based on the prediction model, and finally select a batch of samples from the sample feature variables for model training, optimization and deep analysis to complete the training task of the prediction model. The maturity grading unit is used to re-collect data for a single target banana, generate real-time sample data, and input it into the prediction model to obtain the corresponding real-time maturity index, which is then used as the grading condition for banana maturity.

Images