Method for non-destructive detection of sugar content, acidity and maturity of fruits and vegetables using multispectral imaging

Abstract

The application discloses a kind of multispectral nondestructive testing fruit and vegetable sugar acidity maturity method, it is related to fruit and vegetable detection technical field, collection historical sample data and establish candidate band combination, according to the analysis result setting collection configuration, form real-time collection data, carry out dark current deduction, gain uniform, polar mirror suppression and view angle normalization, establish characteristic vector, carry out prediction, check consistency, carry out batch adaptive update, the application realizes the efficient, nondestructive testing fruit and vegetable sugar acidity and maturity, reduces the sample error and sampling time in traditional detection, improves data quality, and ensures accurate fruit surface information extraction, multidimensional feature extraction and multi-task modeling significantly improve the accuracy of sugar acidity, maturity prediction, especially more stable under complex environment, in addition, it also provides efficient decision support for agricultural production, generates accurate harvesting time window and sorting strategy, reduces resource waste and improves production efficiency.

Description

Technical Field

[0001] This invention relates to the field of fruit and vegetable testing technology, and in particular to a multispectral non-destructive method for detecting the sugar content, acidity, and maturity of fruits and vegetables. Background Technology

[0002] With the advancement of agriculture and the food industry, the accurate measurement of sugar content, acidity, and ripeness of fruits and vegetables has become an important means to improve product quality and reduce waste. Traditional methods for testing the quality of fruits and vegetables mainly rely on manual sampling and chemical analysis, which suffer from problems such as long processing time, large sample errors, and high costs. In order to achieve efficient, rapid, and non-destructive quality testing, detection methods based on spectral technology have gradually gained attention. However, existing multispectral detection methods often face challenges such as limited data acquisition, low accuracy, and significant light interference.

[0003] Currently, Chinese Patent Application No. CN202210490394.8 discloses a method, apparatus, air conditioning device, and storage medium for ripening fruits and vegetables. Applied to an air conditioning device, the device can controllably release ripening gas, including: responding to a user's ripening command and detecting the current maturity of the fruits and vegetables; if the current maturity is less than a preset value, obtaining a target maturity of the fruits and vegetables; determining the concentration of the ripening gas and the ripening duration based on the current maturity and the target maturity; and controlling the air conditioning device to release the ripening gas within the ripening duration to ripen the fruits and vegetables. This application adjusts the concentration of the ripening gas and the ripening duration based on both the current maturity of the fruits and vegetables and the user's desired target maturity. This ensures that the maturity of the fruits and vegetables after ripening matches the user's desired target maturity, thereby meeting the user's personalized needs for fruit and vegetable maturity.

[0004] The relevant technologies struggle to reliably and accurately predict sugar content, acidity, and maturity across varieties and batches in orchards and sorting lines, where strong light, fruit wax reflection, changing viewing angles, and batch variations coexist. They also fail to provide actionable harvesting and sorting decisions. Summary of the Invention

[0005] The technical problem solved by this invention is that existing technologies are unable to stably and accurately predict sugar content, acidity, and maturity across varieties and batches in orchard and sorting line environments where strong light, fruit wax reflection, viewing angle changes, and batch differences coexist, and output executable harvesting and sorting decisions.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A method for non-destructive testing of sugar content, acidity, and maturity of fruits and vegetables using multispectral methods includes the following steps:

[0008] Step S1: Collect historical sample data and establish candidate band combinations;

[0009] Step S2: Set the data acquisition configuration based on the analysis results to generate real-time data acquisition data;

[0010] Step S3: Perform dark current subtraction, gain unification, polarization mirror suppression and viewing angle normalization, and extract the fruit surface mask image through illumination correction;

[0011] Step S4: Establish feature vectors;

[0012] Step S5: Use the joint prediction model to make predictions and check for consistency;

[0013] Step S6: Perform batch adaptive update;

[0014] Step S4 includes the following sub-steps:

[0015] Step S401: Calculate the ratio of the reflectance difference to the sum for the sugar content sensitive band and the reference band;

[0016] For the acidity-sensitive band and the near-infrared reference band, the difference ratio is calculated;

[0017] For the short-wave infrared sensitive band and the near-infrared reference band, the ratio of energy distribution difference is calculated to form an indicator of sugar content and acidity, and the output is a chemical indicator ratio group characteristic.

[0018] Step S402: Calculate the local contrast statistic and scale-related roughness statistic under the three-angle data respectively, and calculate the consistency measure between the three angles. The output is the tissue scattering texture group feature.

[0019] Step S403: Extract the combination amount related to chlorophyll decay in the visible light sensitive segment, compare it with the combination amount related to carotenoids or anthocyanins, and cancel it with the water-related segment to output the pigment migration index group features.

[0020] Step S404: The embedded representation of temperature drift correction, illumination position correction, polarization balance correction and encoding information is used as features, and the output is a robust reference correction group feature.

[0021] Step S405: The chemical indicator ratio group features, tissue scattering texture group features, pigment migration index group features and robust benchmark correction group features are concatenated into a feature vector and aligned with the unique identifier of a single fruit.

[0022] Step S5 includes the following sub-steps:

[0023] Step S501: Using the feature vector as input, the maturity level as the main task output, and the sugar content prediction value and acidity prediction value as the auxiliary task output, a first joint prediction model is constructed.

[0024] Step S502: Establish a correspondence between the first joint prediction model and the encoded information, and output the second joint prediction model.

[0025] Step S503: Obtain real-time encoding information, find the corresponding second joint prediction model, reason about the feature vector to obtain the maturity level, sugar content prediction value and acidity prediction value, and check whether the improvement of maturity level is accompanied by the increase of sugar content prediction value and the decrease of acidity prediction value. If they are consistent, mark it as stable. If they are inconsistent, mark it as needing to be reviewed, and form the second analysis result.

[0026] Step S6 includes the following sub-steps:

[0027] Step S601: When the second analysis result is stable and the maturity level reaches the target level, generate a list of regional harvesting time windows in the orchard under the tree detection scenario. The list of regional harvesting time windows includes a regional identifier, a recommended harvesting date range and a target maturity coverage ratio.

[0028] When the second analysis result is stable and the sorting index reaches the target range, the sorting threshold and sorting frequency are output in the online grading scenario of the sorting line.

[0029] Step S602: Calculate the difference between the feature vector distribution of the target batch and the feature vector distribution of the historical baseline batch, and adjust the distribution of the shared representation layer;

[0030] Step S603: When the prediction confidence is high and consistent with neighboring samples, the sample is cached as a pseudo sample; when the sampling calibration sample is obtained, the second joint prediction model is updated with a small step size.

[0031] If the verification error increases after the update, then roll back to the previous stable checkpoint;

[0032] Step S604: Record the maturity level, sugar content prediction, acidity prediction, confidence level, correction amount, and model version for each inference, and generate a traceable report.

[0033] Preferably, step S1 includes the following sub-steps:

[0034] Step S101: Collect historical samples covering different scenarios, tree positions, batches, and maturity stages, obtain multispectral and multiband data, parallel polarization channel data, vertical polarization channel data, and three-angle acquisition data, and establish a one-to-one correspondence between these data and historical sugar content calibration values, historical acidity calibration values, and historical maturity levels.

[0035] Step S102: Establish a reflectivity reference using a standard white board and dark field sequence; establish a temperature drift reference by recording the correspondence between detector temperature and baseline of each band; establish a polarization balance reference by recording the response difference between parallel polarization channel and vertical polarization channel using a sampleless scene; and establish an illumination position reference by collecting empty background and blade reference data.

[0036] Step S103: Establish candidate band combinations and perform significance tests. For each candidate band combination and each three-angle data collection, compare the distinguishing ability and correlation degree between them and historical sugar content calibration values, historical acidity calibration values ​​and historical maturity levels. Output sensitive band combinations and angle weights, and form the first analysis result.

[0037] Preferably, the significance test includes:

[0038] Multispectral multiband data, parallel polarization channel data, and vertical polarization channel data are divided into several candidate band groups according to the historical load correspondence. The historical load includes variety, scene, and batch.

[0039] For each candidate band combination and each three-angle acquisition data, a comparison index for intra-group and inter-group differences is calculated, and the comparison index is used to measure sensitivity.

[0040] The correlation between the comparative indicators and historical sugar content calibration values, historical acidity calibration values ​​and historical maturity levels was summarized to obtain the stable sensitive band combinations and angle weights under various historical loads, which were used as the first analysis results.

[0041] Preferably, step S2 includes the following sub-steps:

[0042] Step S201: Based on the first analysis results, set the acquisition configuration and determine the sensitive band set, three-angle set and dual polarization acquisition configuration;

[0043] Step S202: Output the self-propelled platform's row-direction movement signal to the orchard tree inspection end, and trigger the acquisition signal sequentially at the front upper angle, side front angle and top angle;

[0044] Output the single-fruit conveyor belt rolling signal to the online grading end of the sorting line;

[0045] Step S203: Synchronously record the position code, attitude code, and job batch code and output them as encoding information;

[0046] Step S204: The multi-band parallel polarization channel data, vertical polarization channel data, three-angle acquisition data and encoding information are stored together in the database to form real-time acquisition data.

[0047] Preferably, step S3 includes the following sub-steps:

[0048] Step S301: Dark current subtraction and gain are unified, and real-time acquired data is converted into preprocessed reflectivity data based on reflectivity reference.

[0049] Step S302: Based on the difference between the parallel polarization channel data and the vertical polarization channel data, the specular reflection component is separated, and the remaining component is used as the diffuse reflection component for polarization specular suppression to obtain specular suppression reflectivity data.

[0050] Step S303: Estimate the scattering intensity difference based on the brightness change and texture contrast change of the three-angle acquired data, perform intensity unification and contrast unification on the three-angle data, and output the viewpoint normalized reflectivity data.

[0051] Step S304: Use the illumination position reference to perform position-related illumination correction on the normalized reflectance data of the viewing angle, and extract the fruit surface mask image under the green background reference, and obtain the normalized reflectance data based on the fruit surface mask image.

[0052] Preferably, the polarization mirror suppression in step S302 involves performing a difference analysis on the parallel polarization channel data and the vertical polarization channel data to identify the energy ratio of the mirror reflection component and remove the energy ratio from the preprocessed reflectivity data.

[0053] Preferably, the viewpoint normalization in step S303 is achieved by comparing the brightness range and texture contrast range of the three angle acquisition data within the same fruit surface area, calculating the difference range between the three angles, and minimizing the difference range.

[0054] The beneficial effects of this invention are as follows: This invention enables efficient and non-destructive detection of sugar content, acidity, and maturity of fruits and vegetables, reduces sample error and sampling time in traditional detection, improves data quality, and ensures accurate extraction of fruit surface information. Multidimensional feature extraction and multi-task modeling significantly improve the accuracy of sugar content, acidity, and maturity prediction, and are particularly stable in complex environments. In addition, it provides efficient decision support for agricultural production, generates precise harvesting windows and sorting strategies, reduces resource waste, and improves production efficiency. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the steps of a method for non-destructive testing of sugar content, acidity, and maturity of fruits and vegetables using multispectral methods, as provided in one embodiment of the present invention. Detailed Implementation

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0057] Example, refer to Figure 1 This paper provides a multispectral non-destructive method for detecting the sugar content, acidity, and maturity of fruits and vegetables, comprising the following steps:

[0058] Step S1: Collect historical sample data and establish candidate band combinations.

[0059] Step S2: Set the acquisition configuration based on the analysis results to generate real-time acquisition data.

[0060] Step S3 involves performing dark current subtraction, gain unification, polarization mirror suppression, and viewing angle normalization, and extracting the fruit surface mask image through illumination correction.

[0061] Step S4: Establish feature vectors.

[0062] Step S5: Use the joint prediction model to make predictions and check for consistency.

[0063] Step S6: Perform batch adaptive update.

[0064] This invention enables efficient and non-destructive detection of sugar content, acidity, and maturity of fruits and vegetables, reducing sample errors and sampling time in traditional detection methods, improving data quality, and ensuring accurate extraction of fruit surface information. Multidimensional feature extraction and multi-task modeling significantly improve the accuracy of sugar content, acidity, and maturity prediction, and are particularly stable in complex environments. In addition, it provides efficient decision support for agricultural production, generating precise harvesting windows and sorting strategies, reducing resource waste and improving production efficiency.

[0065] Step S1 includes the following sub-steps:

[0066] Step S101: Collect historical samples covering different scenarios, tree positions, batches, and maturity stages, obtain multispectral and multiband data, parallel polarization channel data, vertical polarization channel data, and three-angle acquisition data, and establish a one-to-one correspondence between these data and historical sugar content calibration values, historical acidity calibration values, and historical maturity levels.

[0067] Step S101 collects data from multiple scenarios, tree positions, batches, and maturity stages to ensure the diversity and representativeness of historical samples, providing a comprehensive and reliable data source for subsequent analysis. The correspondence with historical calibration values ​​ensures the accuracy and reliability of subsequent predictions.

[0068] Step S102: Establish a reflectivity reference using a standard white board and a dark field sequence; establish a temperature drift reference by recording the correspondence between detector temperature and baselines of each band; establish a polarization balance reference by recording the response difference between the parallel polarization channel and the vertical polarization channel using a sampleless scene; and establish an illumination position reference by collecting empty background and blade reference data.

[0069] Step S102 establishes a benchmark to ensure standardized processing of various types of collected data under different environmental conditions, eliminates errors caused by environmental changes or equipment differences, and ensures data consistency at each collection stage.

[0070] Step S103: Establish candidate band combinations and perform significance tests. For each candidate band combination and each three-angle data collection, compare the distinguishing ability and correlation degree between them and historical sugar content calibration values, historical acidity calibration values ​​and historical maturity levels. Output sensitive band combinations and angle weights, and form the first analysis result.

[0071] Significance tests include:

[0072] Multispectral multiband data, parallel polarization channel data, and vertical polarization channel data are divided into several candidate band groups according to the historical load correspondence. The historical load includes variety, scene, and batch.

[0073] For each candidate band combination and each triangular acquisition data point, comparison indices for intra-group and inter-group differences are calculated. These indices are used to measure sensitivity.

[0074] The correlation between the comparative indicators and historical sugar content calibration values, historical acidity calibration values ​​and historical maturity levels was summarized to obtain the stable sensitive band combinations and angle weights under various historical loads, which were used as the first analysis results.

[0075] Step S103 uses a significance test to select band combinations and acquisition angles that are highly sensitive to sugar content, acidity, and maturity. This process ensures the validity of the data, avoids interference from invalid data, and provides stable and high-quality feature inputs for subsequent model training. Testing under multiple historical loads further improves the model's stability and generalization ability.

[0076] The purpose of step S1 is to ensure the accuracy of subsequent data collection and analysis by collecting historical sample data and establishing benchmarks. By comprehensively covering historical data from different scenarios, tree locations, batches, and maturity stages, and combining multispectral, polarization channel, and three-angle data, the data is ensured to be comprehensive and multidimensional, accurately reflecting the quality indicators of fruits and vegetables under different conditions. The significance test process ensures that the band combination and the selection of collection angles have high discriminative power, providing a stable and effective foundation for subsequent data processing.

[0077] Step S2 includes the following sub-steps:

[0078] Step S201: Based on the first analysis results, set the acquisition configuration and determine the sensitive band set, three-angle set and dual polarization acquisition configuration.

[0079] Step S201 sets the acquisition configuration based on the first analysis results, accurately selecting the sensitive band set, three-angle set, and dual polarization acquisition configuration, ensuring that the acquired data can fully cover the spectral characteristics of fruits and vegetables, thereby improving the data quality and the effectiveness of subsequent analysis.

[0080] Step S202: Output the self-propelled platform's row-direction movement signal to the orchard tree inspection end, and trigger the acquisition signal sequentially at the front-up angle, side-front angle, and top angle.

[0081] Output the single-fruit conveyor belt rolling signal to the online grading end of the sorting line.

[0082] Step S202 synchronously outputs acquisition signals at both the orchard tree inspection end and the sorting line end, ensuring stable acquisition by the self-propelled platform. Through sequential triggering at different angles, it maximizes the capture of multi-angle information from the fruits and vegetables. At the sorting line end, the coordination of the conveyor belt rolling signal effectively synchronizes data acquisition with fruit and vegetable movement, improving the efficiency and accuracy of data acquisition.

[0083] Step S203: Synchronously record the position code, attitude code, and job batch code and output them as coded information.

[0084] Step S203 synchronously records the position code, attitude code, and job batch code, providing a unique identifier for each acquisition. This allows subsequent data processing, analysis, and model training to be accurately traced back to the specific acquisition conditions, ensuring data consistency and repeatability.

[0085] Step S204: The multi-band parallel polarization channel data, vertical polarization channel data, three-angle acquisition data and encoding information are stored together in the database to form real-time acquisition data.

[0086] Step S204 ensures efficient integration and management of the acquired data by storing multi-band, parallel polarization channel, vertical polarization channel, three-angle data, and encoding information in the database. Real-time data storage not only improves data security and accessibility but also provides rich feature information for subsequent analysis, facilitating feature extraction, modeling, and decision support.

[0087] Step S2 achieves precise real-time data acquisition by setting the acquisition configuration based on the first analysis results. Optimization of the sensitive band set, three-angle set, and dual-polarization acquisition configuration ensures that the spectral information acquired during data acquisition is comprehensive and highly discriminative, providing accurate input data for subsequent fruit and vegetable quality assessment. Synchronous recording of real-time acquisition and positioning signals helps ensure the spatiotemporal consistency of the data and provides accurate reference for subsequent data analysis and sorting. Finally, the multi-dimensional acquired data in the database provides the foundation for subsequent feature extraction, modeling, and prediction, ensuring data traceability and operability.

[0088] Step S3 includes the following sub-steps:

[0089] Step S301: Dark current subtraction and gain are unified, and real-time acquired data is converted into preprocessed reflectivity data based on reflectivity reference.

[0090] Step S301 eliminates reflectivity fluctuations caused by equipment errors and environmental factors through dark current subtraction and gain unification, ensuring more stable and reliable data. Preprocessing based on a reflectivity benchmark ensures that all collected data are compared and analyzed under the same benchmark, resulting in more consistent subsequent processing results.

[0091] Step S302: Based on the difference between the parallel polarization channel data and the vertical polarization channel data, the specular reflection component is separated, and the remaining component is used as the diffuse reflection component for polarization specular suppression to obtain specular suppression reflectivity data. Polarization specular suppression identifies the energy ratio of the specular reflection component by performing difference analysis on the parallel polarization channel data and the vertical polarization channel data, and removes the energy ratio from the preprocessed reflectivity data.

[0092] Step S302 successfully separated the specular reflection component by analyzing the difference between the parallel polarization channel and the vertical polarization channel data, and removed it from the reflectance data. This step effectively removed the spectral interference caused by specular reflection, making the subsequent reflectance data more accurately reflect the true spectral characteristics of the substances on the surface of fruits and vegetables.

[0093] Step S303: Estimate the scattering intensity difference based on the brightness change and texture contrast change of the three-angle acquired data, perform intensity unification and contrast unification on the three-angle data, and output viewpoint normalized reflectivity data. Viewpoint normalization is achieved by comparing the brightness range and texture contrast range of the three-angle acquired data in the same fruit surface area, calculating the difference range between the three angles, and minimizing the difference range.

[0094] Step S303 addresses the issues of brightness differences and texture contrast variations under different acquisition angles through viewpoint normalization, resulting in more consistent data across all angles. By minimizing the range of differences between the three-angle data, data deviations caused by angle variations are reduced, enhancing cross-angle consistency of the data.

[0095] Step S304: Use the illumination position reference to perform position-related illumination correction on the normalized reflectance data of the viewing angle, and extract the fruit surface mask image under the green background reference, and obtain the normalized reflectance data based on the fruit surface mask image.

[0096] Step S304, through the combination of illumination correction and fruit surface mask extraction, ensures that the data comes only from the surface area of ​​the fruits and vegetables, effectively removing the influence of the background. Illumination correction using an illumination position reference eliminates the interference of ambient light differences, making the final reflectance data more accurate and reliable, further enhancing the accuracy of fruit and vegetable quality testing.

[0097] Step S3 significantly improves the quality and accuracy of the real-time acquired data through a series of preprocessing and normalization operations. Dark current subtraction, gain unification, and specular suppression eliminate interference from factors such as ambient light and equipment drift, ensuring high stability and consistency of the acquired reflectivity data. Viewpoint normalization and illumination correction further enhance the data's adaptability under different acquisition angles and ambient lighting conditions, ensuring the consistency and reliability of fruit and vegetable surface information. Finally, fruit surface mask extraction ensures that the data originates only from the surface area of ​​the fruits and vegetables, avoiding the influence of background noise.

[0098] Step S4 includes the following sub-steps:

[0099] Step S401: Calculate the ratio of reflectance difference to sum for the sugar content sensitive band and the reference band.

[0100] The difference ratio is calculated for the acidity-sensitive band and the near-infrared reference band.

[0101] For the short-wave infrared sensitive band and the near-infrared reference band, the ratio of the energy distribution difference is calculated to form an indicator of sugar content and acidity, and the output is a chemical indicator ratio group characteristic.

[0102] Step S401 generates indicators for sugar content and acidity by calculating the ratio of the reflectance difference between the sugar content-sensitive band and the reference band to the sum of their values, and the difference ratio between the acidity-sensitive band and the near-infrared reference band. This process provides direct spectral characteristics for subsequent sugar content and acidity prediction, ensuring high sensitivity and accuracy of the data.

[0103] Step S402: Calculate the local contrast statistics and scale-related roughness statistics under the three-angle data, and calculate the consistency measure among the three angles. The output is the tissue scattering texture group feature.

[0104] Step S402 further extracts the texture features of fruit and vegetable surfaces by calculating local contrast statistics and scale-related roughness statistics. Furthermore, calculating the consistency measure among the three-dimensional data helps assess the structural consistency of fruit and vegetable surfaces, providing strong support for maturity prediction.

[0105] Step S403: Extract the combination amount related to chlorophyll decay in the visible light sensitive segment, compare it with the combination amount related to carotenoids or anthocyanins, and cancel it with the water-related segment to output the pigment migration index group features.

[0106] Step S403 extracts the compositional amounts associated with chlorophyll decay and compares them with those associated with carotenoids or anthocyanins, further extracting features related to pigment changes. These features play a crucial role in assessing the maturity and quality changes of fruits and vegetables, effectively reflecting their physiological state.

[0107] Step S404: The embedded representation of temperature drift correction, illumination position correction, polarization balance correction and encoding information is used as features, and the output is a robust reference correction group feature.

[0108] Step S404 enhances the robustness of the features by embedding corrections for temperature drift, illumination position, and polarization balance with the encoded information, ensuring high stability and accuracy even under different acquisition environments or conditions. This step helps reduce the impact of the external environment on the data and improves the model's adaptability.

[0109] Step S405: The chemical indicator ratio group features, tissue scattering texture group features, pigment migration index group features and robust benchmark correction group features are concatenated into a feature vector and aligned with the unique identifier of a single fruit.

[0110] Step S405 ensures data integrity and consistency by concatenating the various feature groups into a feature vector and aligning it with the unique identifier of each fruit. This allows subsequent analysis, modeling, and prediction processes to be based on a unified and stable dataset, improving the accuracy and reliability of fruit and vegetable quality assessment.

[0111] Step S4 extracts multi-level features from reflectance data, forming rich feature vectors that provide high-quality data support for predicting the sugar content, acidity, and maturity of fruits and vegetables. By constructing features from chemical indicator ratio groups, tissue scattering texture groups, pigment migration index groups, and robust benchmark correction groups, this step effectively characterizes the intrinsic quality features of fruits and vegetables, ensuring a comprehensive assessment of their quality. Finally, these features are integrated into feature vectors, combined with the unique identifiers of fruits and vegetables, ensuring data consistency and traceability, and providing reliable input for subsequent modeling and prediction.

[0112] Step S5 includes the following sub-steps:

[0113] Step S501: Using the feature vector as input, the maturity level as the main task output, and the sugar content prediction value and acidity prediction value as the auxiliary task output, a first joint prediction model is constructed.

[0114] Step S501 uses the feature vector as input to establish a first joint prediction model with maturity level as the main task and sugar content prediction and acidity prediction as auxiliary tasks, forming a shared representation to avoid information fragmentation caused by single-task fitting. The main and auxiliary tasks are jointly constrained by the trend consistency relationship, reducing the impact of abnormal samples on a single task and improving the overall prediction stability.

[0115] Step S502: Establish a correspondence between the first joint prediction model and the encoded information, and output the second joint prediction model.

[0116] Step S502 establishes a correspondence between the first joint prediction model and the coding information to obtain several second joint prediction models, which bind the model parameters to the scenario, variety and batch conditions, select the corresponding model for different coding information, and reduce the prediction bias caused by the distribution differences across scenarios and batches.

[0117] Step S503: Obtain real-time encoding information, find the corresponding second joint prediction model, perform reasoning on the feature vector to obtain maturity level, sugar content prediction value and acidity prediction value, and check whether the improvement of maturity level is accompanied by an increase in sugar content prediction value and a decrease in acidity prediction value. If they are consistent, mark it as stable; if they are inconsistent, mark it as needing to be reviewed, and form the second analysis result.

[0118] Step S503 reads the real-time encoding information, selects the corresponding second joint prediction model, performs inference on the feature vector to obtain the maturity level, sugar content prediction value and acidity prediction value, checks whether the improvement of the maturity level is accompanied by the increase of the sugar content prediction value and the decrease of the acidity prediction value. If consistent, it is marked as stable; if inconsistent, it is marked as needing to be reviewed. The second analysis result is output and directly used for strategy selection and model update in S6.

[0119] Step S5 outputs maturity level, sugar content prediction, and acidity prediction simultaneously using feature vectors. Results that do not conform to physical relationships are filtered out by trend consistency checks, forming a second analysis result for decision-making. By matching the first joint prediction model with the encoded information, a matching second joint prediction model is selected online to reduce the bias caused by differences in variety, scenario, and batch. The main task and auxiliary task share a representation, and the output distribution is stabilized by inter-task constraints, providing reliable input for the harvesting / sorting strategy and batch domain adaptation in S6.

[0120] Step S6 includes the following sub-steps:

[0121] Step S601: When the second analysis result is stable and the maturity level reaches the target level, generate a list of regional harvesting time windows in the orchard under the tree detection scenario. The list of regional harvesting time windows includes the regional identifier, the recommended harvesting date range and the target maturity coverage ratio.

[0122] When the second analysis result is stable and the sorting index reaches the target range, the sorting threshold and sorting frequency are output in the online grading scenario of the sorting line.

[0123] Step S601: When the second analysis result is stable and the maturity level meets the standard, output a list of zonal harvesting time windows to arrange zonal harvesting. When the second analysis result is stable and the sorting index meets the standard, output the sorting threshold and sorting frequency to make the sorting execution consistent with the model output.

[0124] Step S602: Calculate the difference between the feature vector distribution of the target batch and the feature vector distribution of the historical benchmark batch, and adjust the distribution of the shared representation layer.

[0125] Step S602 quantifies the difference in feature vector distribution between the target batch and the historical baseline batch, and adjusts the distribution for the shared representation layer to ensure that the second joint prediction model maintains consistent output under the new batch conditions.

[0126] Step S603: When the prediction confidence is high and consistent with neighboring samples, the sample is cached as a pseudo-sample. The second joint prediction model is updated with small steps when the sampled calibration samples are obtained.

[0127] If the verification error increases after the update, then roll back to the previous stable checkpoint.

[0128] Step S603 performs pseudo-sample caching when the prediction confidence is high and consistent with neighboring samples; after obtaining the sampled calibration, the second joint prediction model is updated with a small step size; if the verification error increases, it immediately reverts to the previous stable checkpoint to suppress error accumulation.

[0129] Step S604: Record the maturity level, sugar content prediction, acidity prediction, confidence level, correction amount, and model version for each inference, and generate a traceable report.

[0130] Step S604 records the maturity level, sugar content prediction, acidity prediction, confidence level, correction amount, and model version for each inference, generating a traceable report to provide a basis for result verification, strategy evaluation, and model version management.

[0131] Step S6, based on the second analysis results, generates a list of zoning harvesting windows in the orchard tree detection scenario, and outputs the sorting threshold and sorting frequency in the online grading scenario of the sorting line to ensure that the prediction results directly drive the operation. By statistically analyzing the difference in feature vector distribution between the target batch and the historical benchmark batch and adjusting the shared representation layer, the bias caused by batch difference is reduced. The model freshness is maintained by using confidence-gated pseudo-sample caching and small-step updates. Validation error monitoring and stability checkpoint rollback are introduced to control drift risk. The predicted value, confidence level, correction amount and model version are recorded to form a traceable report to support review and audit.

[0132] This invention's method, through the simultaneous acquisition of multispectral, polarization channel, and multi-angle data, enables rapid, accurate, and non-destructive detection of sugar content, acidity, and maturity in fruits and vegetables. It reduces sample error and sampling time in traditional detection methods. Data preprocessing steps, including dark current subtraction, gain unification, polarization mirror suppression, and viewing angle normalization, effectively mitigate the impact of environmental factors such as light and angle, and extract accurate fruit surface information, ensuring high-quality data. Multi-dimensional feature extraction and multi-task joint modeling effectively improve the accuracy of sugar content, acidity, and maturity prediction, exhibiting greater stability, especially in complex natural environments. The batch-domain adaptive technique handles differences between batches, ensuring continuous model optimization and adaptability, and preventing increased errors due to batch variations. Accurate fruit and vegetable quality assessment generates efficient harvesting windows and sorting strategies, providing real-time decision support for agricultural production, reducing resource waste, and improving production efficiency.

[0133] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0134] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for non-destructive testing of sugar content, acidity, and maturity of fruits and vegetables using multispectral methods, characterized in that... Includes the following steps: Step S1: Collect historical sample data and establish candidate band combinations; Step S2: Set the data acquisition configuration based on the analysis results to generate real-time data acquisition data; Step S3: Perform dark current subtraction, gain unification, polarization mirror suppression and viewing angle normalization, and extract the fruit surface mask image through illumination correction; Step S4: Establish feature vectors; Step S5: Use the joint prediction model to make predictions and check for consistency. Step S6: Perform batch adaptive update; Step S1 includes the following sub-steps: Step S101: Collect historical samples covering different scenarios, tree positions, batches, and maturity stages, obtain multispectral and multiband data, parallel polarization channel data, vertical polarization channel data, and three-angle acquisition data, and establish a one-to-one correspondence between these data and historical sugar content calibration values, historical acidity calibration values, and historical maturity levels. Step S102: Establish a reflectivity reference using a standard white board and dark field sequence; establish a temperature drift reference by recording the correspondence between detector temperature and baseline of each band; establish a polarization balance reference by recording the response difference between parallel polarization channel and vertical polarization channel using a sampleless scene; and establish an illumination position reference by collecting empty background and blade reference data. Step S103: Establish candidate band combinations and perform significance tests. For each candidate band combination and each three-angle data collection, compare the distinguishing ability and correlation degree between them and historical sugar content calibration values, historical acidity calibration values ​​and historical maturity levels. Output sensitive band combinations and angle weights, and form the first analysis result. Step S4 includes the following sub-steps: Step S401: Calculate the ratio of the reflectance difference to the sum for the sugar content sensitive band and the reference band; For the acidity-sensitive band and the near-infrared reference band, the difference ratio is calculated; For the short-wave infrared sensitive band and the near-infrared reference band, the ratio of energy distribution difference is calculated to form an indicator of sugar content and acidity, and the output is a chemical indicator ratio group characteristic. Step S402: Calculate the local contrast statistic and scale-related roughness statistic under the three-angle data respectively, and calculate the consistency measure between the three angles. The output is the tissue scattering texture group feature. Step S403: Extract the combination amount related to chlorophyll decay in the visible light sensitive segment, compare it with the combination amount related to carotenoids or anthocyanins, and cancel it with the water-related segment to output the pigment migration index group features. Step S404: The embedded representation of temperature drift correction, illumination position correction, polarization balance correction and encoding information is used as features, and the output is a robust reference correction group feature. Step S405: The chemical indicator ratio group features, tissue scattering texture group features, pigment migration index group features and robust benchmark correction group features are concatenated into a feature vector and aligned with the unique identifier of a single fruit. Step S5 includes the following sub-steps: Step S501: Using the feature vector as input, the maturity level as the main task output, and the sugar content prediction value and acidity prediction value as the auxiliary task output, a first joint prediction model is constructed. Step S502: Establish a correspondence between the first joint prediction model and the encoded information, and output the second joint prediction model. Step S503: Obtain real-time encoding information, find the corresponding second joint prediction model, reason about the feature vector to obtain the maturity level, sugar content prediction value and acidity prediction value, and check whether the improvement of maturity level is accompanied by the increase of sugar content prediction value and the decrease of acidity prediction value. If they are consistent, mark it as stable. If they are inconsistent, mark it as needing to be reviewed, and form the second analysis result. Step S6 includes the following sub-steps: Step S601: When the second analysis result is stable and the maturity level reaches the target level, generate a list of regional harvesting time windows in the orchard under the tree detection scenario. The list of regional harvesting time windows includes a regional identifier, a recommended harvesting date range and a target maturity coverage ratio. When the second analysis result is stable and the sorting index reaches the target range, the sorting threshold and sorting frequency are output in the online grading scenario of the sorting line. Step S602: Calculate the difference between the feature vector distribution of the target batch and the feature vector distribution of the historical baseline batch, and adjust the distribution of the shared representation layer; Step S603: When the prediction confidence is high and consistent with neighboring samples, the sample is cached as a pseudo sample; when the sampling calibration sample is obtained, the second joint prediction model is updated with a small step size. If the verification error increases after the update, then roll back to the previous stable checkpoint; Step S604: Record the maturity level, sugar content prediction, acidity prediction, confidence level, correction amount, and model version for each inference, and generate a traceable report.

2. The method for non-destructive testing of sugar content, acidity, and maturity of fruits and vegetables using multispectral methods as described in claim 1, characterized in that... The significance test includes: Multispectral multiband data, parallel polarization channel data, and vertical polarization channel data are divided into several candidate band groups according to the historical load correspondence. The historical load includes variety, scene, and batch. For each candidate band combination and each three-angle acquisition data, a comparison index for intra-group and inter-group differences is calculated, and the comparison index is used to measure sensitivity. The correlation between the comparative indicators and historical sugar content calibration values, historical acidity calibration values ​​and historical maturity levels was summarized to obtain the stable sensitive band combinations and angle weights under various historical loads, which were used as the first analysis results.

3. The method for non-destructive testing of sugar content, acidity, and maturity of fruits and vegetables using multispectral methods as described in claim 2, characterized in that... Step S2 includes the following sub-steps: Step S201: Based on the first analysis results, set the acquisition configuration and determine the sensitive band set, three-angle set and dual polarization acquisition configuration; Step S202: Output the self-propelled platform's row-direction movement signal to the orchard tree inspection end, and trigger the acquisition signal sequentially at the front upper angle, side front angle and top angle; Output the single-fruit conveyor belt rolling signal to the online grading end of the sorting line; Step S203: Synchronously record the position code, attitude code, and job batch code and output them as encoding information; Step S204: The multi-band parallel polarization channel data, vertical polarization channel data, three-angle acquisition data and encoding information are stored together in the database to form real-time acquisition data.

4. The method for non-destructive testing of sugar content, acidity, and maturity of fruits and vegetables using multispectral methods as described in claim 3, characterized in that... Step S3 includes the following sub-steps: Step S301: Dark current subtraction and gain are unified, and real-time acquired data is converted into preprocessed reflectivity data based on reflectivity reference. Step S302: Based on the difference between the parallel polarization channel data and the vertical polarization channel data, the specular reflection component is separated, and the remaining component is used as the diffuse reflection component for polarization specular suppression to obtain specular suppression reflectivity data. Step S303: Estimate the scattering intensity difference based on the brightness change and texture contrast change of the three-angle acquired data, perform intensity unification and contrast unification on the three-angle data, and output the viewpoint normalized reflectivity data. Step S304: Use the illumination position reference to perform position-related illumination correction on the normalized reflectance data of the viewing angle, and extract the fruit surface mask image under the green background reference, and obtain the normalized reflectance data based on the fruit surface mask image.

5. The method for non-destructive testing of sugar content, acidity, and maturity of fruits and vegetables using multispectral methods as described in claim 4, characterized in that... The polarization mirror suppression in step S302 involves analyzing the differences between the parallel polarization channel data and the vertical polarization channel data to identify the energy ratio of the mirror reflection component and then removing the energy ratio from the preprocessed reflectivity data.

6. The method for multispectral nondestructive testing of sugar content, acidity, and maturity of fruits and vegetables as described in claim 5, characterized in that, The viewpoint normalization in step S303 calculates the difference range between the three angles by comparing the brightness range and texture contrast range of the three angle acquisition data in the same fruit surface area, and minimizes the difference range.

Images