Korla pear recessive damage detection method
By using multi-source detection equipment and composite detection models, combined with deep learning and traditional machine learning, the hidden damage of Korla fragrant pears can be identified, solving the problem of low detection accuracy in existing technologies. This enables rapid and accurate identification of hidden damage and generation of detailed reports, improving product quality transparency and market circulation quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TARIM UNIV
- Filing Date
- 2026-01-13
- Publication Date
- 2026-04-21
AI Technical Summary
Existing technologies are insufficient to effectively identify hidden damage in Korla pears, resulting in low detection accuracy, poor efficiency, and an inability to adapt to large-scale industrial applications.
Multi-source detection equipment was used to collect external image data, internal structure scan data and physiological index data of pears. A composite detection model integrating deep learning and traditional machine learning was constructed. Multi-dimensional feature information was integrated through a feature fusion layer. The weight of damage influence was determined by combining the analytic hierarchy process and the entropy weight method, so as to achieve accurate identification of latent damage.
This significantly improves the comprehensiveness and reliability of detecting latent damage in Korla fragrant pears, enabling rapid and accurate identification and generating detailed test reports. This provides targeted guidance for subsequent processing, enhancing product quality transparency and market circulation quality.
Smart Images
Figure CN121899063A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural product testing technology, and in particular to a method for detecting latent damage in Korla fragrant pears. Background Technology
[0002] Korla fragrant pears are prone to hidden damage, such as mechanical and physiological damage, during harvesting, transportation, storage, and processing. These types of damage initially show no obvious external signs and are difficult to identify with the naked eye, but they accelerate the pears' decay and spoilage, not only reducing the fruit's commercial value but also causing serious economic losses.
[0003] Currently, damage detection in Korla fragrant pears relies heavily on manual sensory assessment or single physical detection methods. Manual assessment is greatly affected by subjective experience and environmental conditions, resulting in low accuracy and efficiency, and it cannot identify hidden damage. Single physical detection methods can only obtain data on the appearance or local characteristics of the fragrant pear, making it difficult to comprehensively reflect the multi-dimensional changes in structure, mechanics, and physiological metabolism caused by hidden damage, leading to one-sided detection results and a high rate of missed and false diagnoses.
[0004] Non-destructive testing methods based on technologies such as machine vision and spectral analysis are gradually being applied to fruit damage detection, but existing technologies still have obvious limitations: some methods are only designed for specific types of damage and have poor versatility; some methods do not fully integrate multi-dimensional features and have insufficient sensitivity to identify latent damage; and some methods have complex models and cumbersome detection processes, making it difficult to adapt to the needs of large-scale industrial applications. Summary of the Invention
[0005] The purpose of this invention is to provide a method for detecting latent damage in Korla fragrant pears, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting latent damage in Korla fragrant pears, comprising the following steps:
[0007] Korla fragrant pears with uniform maturity and no obvious appearance damage were selected as the test objects. The appearance image data, internal structure scan data, mechanical property data and physiological index data of the fragrant pears were collected by multi-source detection equipment to form a comprehensive test dataset.
[0008] The comprehensive detection dataset is preprocessed to obtain a standardized detection dataset;
[0009] A composite detection model integrating deep learning and traditional machine learning is constructed. A standardized detection dataset is input into the composite detection model to extract feature information related to latent damage in pears. The feature information includes texture anomaly features, structural density features, elastic modulus features, and physiological metabolic features.
[0010] Based on the aforementioned feature information, the probability value of latent damage is calculated. Combined with a preset damage judgment threshold, the system identifies whether there is latent damage in Korla pears and outputs the detection results.
[0011] Furthermore, the multi-source detection device includes:
[0012] A high-resolution industrial camera is used to acquire RGB images and infrared thermal images of the surface of the pear to obtain appearance image data;
[0013] X-ray tomography equipment is used for non-destructive scanning of pears to obtain internal structure scanning data, including pulp density distribution and cell structure integrity.
[0014] A universal testing machine is used to perform compression and puncture tests on pears and collect mechanical property data, including elastic modulus, compressive strength and puncture force.
[0015] A near-infrared spectroscopy analyzer is used to detect physiological indicators of pears, including sugar content, moisture content, and polyphenol content.
[0016] Furthermore, the construction process of the composite detection model includes:
[0017] Samples of Korla fragrant pears with known latent damage were obtained, and corresponding comprehensive detection data were collected to construct training datasets, validation datasets, and test datasets, with the ratio of training datasets, validation datasets, and test datasets being 7:2:1.
[0018] A sub-model for appearance feature extraction based on convolutional neural network is built to extract texture anomaly features and color distribution features from appearance image data.
[0019] A time-series feature extraction sub-model based on a recurrent neural network is used to extract time-series variation features from internal structure scanning data.
[0020] A sub-model for classifying mechanical and physiological characteristics was built based on support vector machines, which was used to classify mechanical property data and physiological index data.
[0021] A feature fusion layer is constructed, and the features output by the three sub-models are fused using a weighted summation method to obtain a comprehensive feature vector;
[0022] A fully connected layer and a Softmax classifier are connected after the feature fusion layer to construct a complete composite detection model;
[0023] The training dataset is input into the composite detection model for training. The learning rate and regularization coefficient of the model are adjusted using the validation dataset. The model performance is verified using the test dataset. The model construction is completed when the detection accuracy of the model reaches more than 95%.
[0024] Furthermore, the feature information extraction process includes:
[0025] Texture features, color features, and infrared thermal imaging temperature distribution features of the surface of the pear were extracted from the appearance image data. The texture features include gray-level co-occurrence matrix and local binary mode. The color features include RGB color space components and HSV color space components. Texture anomaly features and temperature anomaly features related to latent damage were screened out.
[0026] Structural density features are extracted from internal structural scanning data, including pulp density distribution parameters, intercellular space size, and the presence or absence of cavities or browning areas.
[0027] Extract mechanical anomaly features from mechanical property data, including peak values and variation curves of elastic modulus, compressive strength, and puncture force;
[0028] Physiological metabolic features are extracted from physiological index data, including the values and rates of change of sugar content, water content, and polyphenol content.
[0029] By integrating texture features, color features, infrared thermal imaging temperature distribution features, structural density features, mechanical anomaly features, and physiological metabolic features, complete latent damage-related feature information is obtained.
[0030] Furthermore, based on the aforementioned feature information, identifying whether Korla fragrant pears have latent damage includes:
[0031] Calculate the damage impact weight corresponding to each latent damage-related feature information, and the damage impact weight is determined by combining the analytic hierarchy process (AHP) with the entropy weight method.
[0032] The relevant feature information of each latent injury is weighted and fused according to the injury impact weight to obtain a comprehensive feature value; the comprehensive feature value is compared with a preset injury judgment threshold, which is determined to be the optimal critical value through the receiver operating characteristic curve;
[0033] If the comprehensive feature value is greater than or equal to the preset damage judgment threshold, then the Korla fragrant pear is judged to have latent damage; otherwise, the Korla fragrant pear is judged to have no latent damage.
[0034] Furthermore, the process for determining the damage impact weight includes:
[0035] A feature hierarchy structure is constructed, with the target layer representing the accuracy of latent damage detection, the criterion layer representing appearance features, structural features, mechanical features, and physiological features, and the scheme layer representing the feature information related to each specific latent damage.
[0036] Invite 5-8 experts in the field of agricultural testing to conduct pairwise comparisons of features at each feature level, construct a judgment matrix, calculate the maximum eigenvalue and corresponding eigenvector of the judgment matrix, and conduct a consistency test.
[0037] The subjective weights of each latent injury-related feature information are calculated based on the effective judgment matrix; the information entropy of each latent injury-related feature information is calculated, and the objective weights are determined based on the information entropy.
[0038] By combining subjective and objective weights using a weighted average method, the damage impact weights of each latent damage-related feature information are obtained.
[0039] Furthermore, it also includes:
[0040] When it is determined that there is latent damage in Korla fragrant pears, the specific situation of the latent damage is further analyzed based on the extracted feature information, including the damage type, damage location and damage degree. The damage type includes mechanical damage, physiological damage and disease damage. The damage location includes shallow pulp and deep pulp. The damage degree includes mild, moderate and severe.
[0041] A test report is generated based on the type, location and extent of damage. The test report includes basic information about the pear, damage assessment results and corresponding treatment suggestions. The basic information includes the test number, test time and maturity level.
[0042] The test report is stored in a cloud database, and a traceable QR code is generated.
[0043] Furthermore, determining the extent of occult lesions includes:
[0044] A damage assessment system was constructed based on the percentage of damaged area, the degree of structural damage, and the degree of abnormality in physiological indicators. Mild damage was defined as ≤5% of the damaged area, structural damage involving only the shallow pulp, and an abnormality rate of physiological indicators ≤10%; moderate damage was defined as 5%-15% of the damaged area, structural damage involving the middle pulp, and an abnormality rate of physiological indicators 10%-20%; severe damage was defined as ≥15% of the damaged area, structural damage involving the deep pulp or the presence of cavities, and an abnormality rate of physiological indicators ≥20%.
[0045] Extract the damage feature parameters corresponding to latent damage, including the damage area ratio, the decrease in pulp density, the decrease in elastic modulus ratio, and the abnormality rate of physiological indicators;
[0046] Input the damage feature parameters into the damage degree evaluation system, and calculate the quantified value D of the damage degree by using the weighted summation method;
[0047] Determine the damage level according to the quantified value D of the damage degree. When D ≤ 0.05, it is a minor injury; when 0.05 < D ≤ 0.15, it is a moderate injury; when D > 0.15, it is a severe injury.
[0048] Furthermore, the preprocessing of the comprehensive detection dataset includes:
[0049] Detect the abnormal data in the detection dataset, mark the data beyond ±3 times the standard deviation as abnormal data and eliminate it; perform standardization processing on the dataset after eliminating abnormal data, convert different types of data to the [0,1] interval, and eliminate the dimension difference; perform data augmentation on the standardized dataset by using random flipping, rotation, noise addition and data interpolation to expand the dataset scale and generate a standardized detection dataset.
[0050] Furthermore, it also includes the update and optimization of the composite detection model:
[0051] Regularly collect new Korla fragrant pear detection data and the corresponding hidden damage verification results, and the verification results are confirmed through manual dissection and professional instrument detection;
[0052] Add the newly collected detection data and verification results to the training dataset, and retrain the composite detection model; adjust the model structure and parameters based on the performance indicators of the retrained model to optimize the detection performance of the model.
[0053] Compared with the prior art, the beneficial effects of the present invention are:
[0054] 1. The present invention uses multi-source detection devices to collect multi-dimensional data of the appearance, internal structure, mechanical properties and physiological indicators of fragrant pears, combined with a standardized preprocessing process, breaking the limitation of a single detection dimension, comprehensively capturing visual, structural, physical and metabolic signals related to hidden damage, avoiding missed and misjudgments caused by one-sided data, and significantly improving the comprehensiveness and reliability of the detection of hidden damage of Korla fragrant pears.
[0055] 2. The present invention constructs a composite detection model that integrates deep learning and traditional machine learning. By dividing the work of multiple sub-models to extract features and weighted fusion, it gives full play to the technical advantages of different algorithms, accurately mines the damage features in each dimension of data, and solves the problem that the hidden damage features are complex and difficult to accurately identify with a single model; the feature fusion layer integrates the outputs of each sub-model, strengthens the synergistic effect of key damage information, and配合科学的权重确定方法和最优判定阈值,使检测准确率大幅提升,实现对库尔勒香梨隐性损伤的快速、精准识别。
[0056] It should be noted that there is an incomplete sentence in the translation of item . You may need to check and correct it according to the original Chinese text. The correct translation of this sentence should be: 2. The present invention constructs a composite detection model that integrates deep learning and traditional machine learning. By dividing the work of multiple sub-models to extract features and weighted fusion, it gives full play to the technical advantages of different algorithms, accurately mines the damage features in each dimension of data, and solves the problem that the hidden damage features are complex and difficult to accurately identify with a single model; the feature fusion layer integrates the outputs of each sub-model, strengthens the synergistic effect of key damage information, and cooperates with a scientific weight determination method and an optimal decision threshold, so that the detection accuracy is greatly improved, and the rapid and accurate identification of hidden damage of Korla fragrant pears is realized.3. Based on damage identification, this invention further analyzes the type, location, and extent of damage, generating a detailed inspection report with traceability capabilities. This breaks through the limitations of basic inspections that only determine whether damage has occurred, providing targeted guidance for subsequent processing of Korla pears, assisting in quality grading, storage scheme optimization, and processing and utilization decisions. Cloud storage and QR code traceability enable full-process control of inspection data, improving product quality transparency and credibility, and effectively ensuring the market circulation quality of Korla pears. Attached Figure Description
[0057] Figure 1 This is a schematic diagram of the detection method for latent damage in Korla fragrant pears according to the present invention. Detailed Implementation
[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] Please see Figure 1 The present invention provides the following technical solutions:
[0060] A method for detecting latent damage in Korla fragrant pears includes the following steps:
[0061] Korla fragrant pears with uniform maturity and no obvious appearance damage were selected as the test objects. The appearance image data, internal structure scan data, mechanical property data and physiological index data of the fragrant pears were collected by multi-source detection equipment to form a comprehensive test dataset.
[0062] The comprehensive detection dataset is preprocessed to obtain a standardized detection dataset;
[0063] A composite detection model integrating deep learning and traditional machine learning is constructed. A standardized detection dataset is input into the composite detection model to extract feature information related to latent damage in pears. The feature information includes texture anomaly features, structural density features, elastic modulus features, and physiological metabolic features.
[0064] Based on the aforementioned feature information, the probability value of latent damage is calculated. Combined with a preset damage judgment threshold, the system identifies whether there is latent damage in Korla pears and outputs the detection results.
[0065] Multi-source detection equipment includes:
[0066] A high-resolution industrial camera is used to acquire RGB images and infrared thermal images of the surface of the pear to obtain appearance image data;
[0067] X-ray tomography equipment is used for non-destructive scanning of pears to obtain internal structure scanning data, including pulp density distribution and cell structure integrity.
[0068] A universal testing machine is used to perform compression and puncture tests on pears and collect mechanical property data, including elastic modulus, compressive strength and puncture force.
[0069] A near-infrared spectroscopy analyzer is used to detect physiological indicators of pears, including sugar content, moisture content, and polyphenol content.
[0070] In the above embodiments, the targeted selection and functional settings of multi-source detection equipment enabled the precise acquisition of different dimensional characteristics of Korla pears. The high-resolution industrial camera simultaneously acquired RGB images and infrared thermal images, capturing not only visible surface texture and color information but also identifying hidden damage areas that are difficult to detect with the naked eye through temperature distribution differences. The non-destructive scanning method of the X-ray tomography equipment clearly presented the pulp density distribution and cell structure integrity without damaging the pear, effectively capturing internal hidden damage. The universal material testing machine accurately acquired mechanical property parameters through standardized compression and puncture tests, objectively reflecting the impact of damage on the physical properties of the pear. The near-infrared spectroscopy analyzer rapidly detected physiological index data and sensitively captured the changes in component metabolism caused by damage.
[0071] The process of constructing the composite detection model includes:
[0072] Samples of Korla fragrant pears with known latent damage were obtained, and corresponding comprehensive detection data were collected to construct training datasets, validation datasets, and test datasets, with the ratio of training datasets, validation datasets, and test datasets being 7:2:1.
[0073] A sub-model for appearance feature extraction based on convolutional neural network is built to extract texture anomaly features and color distribution features from appearance image data.
[0074] A time-series feature extraction sub-model based on a recurrent neural network is used to extract time-series variation features from internal structure scanning data.
[0075] A sub-model for classifying mechanical and physiological characteristics was built based on support vector machines, which was used to classify mechanical property data and physiological index data.
[0076] A feature fusion layer is constructed, and the features output by the three sub-models are fused using a weighted summation method to obtain a comprehensive feature vector;
[0077] A fully connected layer and a Softmax classifier are connected after the feature fusion layer to construct a complete composite detection model;
[0078] The training dataset is input into the composite detection model for training. The learning rate and regularization coefficient of the model are adjusted using the validation dataset. The model performance is verified using the test dataset. The model construction is completed when the detection accuracy of the model reaches more than 95%.
[0079] In the above embodiments, the composite detection model integrates the advantages of deep learning and traditional machine learning, achieving efficient collaboration between feature extraction and classification. Convolutional neural networks excel at extracting deep features from image data and are suitable for uncovering texture and color anomalies in appearance images. Recurrent neural networks can capture the changing patterns of time-series data and accurately extract the time-series features of internal structure scan data. Support vector machines perform well in classifying small-sample, high-dimensional data and can effectively handle classification tasks of mechanical and physiological features. The feature fusion layer integrates the outputs of each sub-model through weighted summation, giving full play to the synergistic effect of different features. The fully connected layer and the Softmax classifier achieve accurate mapping from comprehensive features to damage recognition results. During model training, parameter adjustment and performance verification ensure the high accuracy of the model, providing strong algorithmic support for the rapid and accurate identification of latent damage.
[0080] The process of extracting feature information includes:
[0081] Texture features, color features, and infrared thermal imaging temperature distribution features of the surface of the pear were extracted from the appearance image data. The texture features include gray-level co-occurrence matrix and local binary mode. The color features include RGB color space components and HSV color space components. Texture anomaly features and temperature anomaly features related to latent damage were screened out.
[0082] Structural density features are extracted from internal structural scanning data, including pulp density distribution parameters, intercellular space size, and the presence or absence of cavities or browning areas.
[0083] Extract mechanical anomaly features from mechanical property data, including peak values and variation curves of elastic modulus, compressive strength, and puncture force;
[0084] Physiological metabolic features are extracted from physiological index data, including the values and rates of change of sugar content, water content, and polyphenol content.
[0085] By integrating texture features, color features, infrared thermal imaging temperature distribution features, structural density features, mechanical anomaly features, and physiological metabolic features, complete latent damage-related feature information is obtained.
[0086] In the above embodiments, the feature information extraction process follows the principles of comprehensiveness, relevance, and correlation, accurately selecting key features highly correlated with latent damage. Multiple types of texture, color, and temperature features are extracted from appearance image data to comprehensively capture the visual and thermal signals of surface latent damage; structural density-related features are focused on from internal structure scan data to directly reflect the damage status of internal tissues; strength, elasticity-related parameters, and change curve features are extracted from mechanical property data to objectively reflect the impact of damage on physical properties; and component content and change rate are monitored from physiological indicator data to sensitively capture metabolic abnormalities caused by damage. Through the systematic extraction and integration of various features, a complete latent damage feature system is constructed, covering damage signals of different dimensions, highlighting the core role of key features, and effectively improving the correlation between features and latent damage.
[0087] Identifying whether Korla fragrant pears have latent damage based on the aforementioned feature information includes:
[0088] The damage impact weights corresponding to the relevant feature information of each latent injury are calculated. The damage impact weights are determined by combining the analytic hierarchy process with the entropy weight method, taking into account both expert experience and the information entropy of the data itself.
[0089] The relevant feature information of each latent injury is weighted and fused according to the injury impact weight to obtain a comprehensive feature value; the comprehensive feature value is compared with a preset injury judgment threshold, which is determined to be the optimal critical value through the receiver operating characteristic curve;
[0090] If the comprehensive feature value is greater than or equal to the preset damage judgment threshold, then the Korla fragrant pear is judged to have latent damage; otherwise, the Korla fragrant pear is judged to have no latent damage.
[0091] The process of determining the weight of damage impact includes:
[0092] A feature hierarchy structure is constructed, with the target layer representing the accuracy of latent damage detection, the criterion layer representing appearance features, structural features, mechanical features, and physiological features, and the scheme layer representing the feature information related to each specific latent damage.
[0093] Invite 5-8 experts in the field of agricultural testing to conduct pairwise comparisons of features at each feature level, construct a judgment matrix, calculate the maximum eigenvalue and corresponding eigenvector of the judgment matrix, and conduct a consistency test.
[0094] The subjective weights of each latent injury-related feature information are calculated based on the effective judgment matrix; the information entropy of each latent injury-related feature information is calculated, and the objective weights are determined based on the information entropy.
[0095] By combining subjective and objective weights using a weighted average method, the damage impact weights of each latent damage-related feature information are obtained.
[0096] In the above embodiments, the analytic hierarchy process (AHP) combined with the entropy weight method is used to determine the weight of damage impact, achieving an organic combination of subjective experience and objective data. This makes the weight allocation more scientific and reasonable. The AHP constructs a clear feature hierarchy structure and leverages the professional experience of experts to compare the importance of each feature pairwise, fully utilizing the knowledge advantages of domain experts. The entropy weight method determines objective weights based on the information entropy of the data itself, reflecting the inherent laws and differences of the data and avoiding the one-sidedness of subjective assumptions. The comprehensive feature value after weighted fusion can comprehensively and reasonably reflect the contribution of each feature to the detection of latent damage. The use of the receiver operating characteristic (ROC) curve to determine the preset damage judgment threshold ensures the optimality of the threshold and improves the accuracy and reliability of damage judgment.
[0097] This embodiment also includes:
[0098] When it is determined that there is latent damage in Korla fragrant pears, the specific situation of the latent damage is further analyzed based on the extracted feature information, including the damage type, damage location and damage degree. The damage type includes mechanical damage, physiological damage and disease damage. The damage location includes shallow pulp and deep pulp. The damage degree includes mild, moderate and severe.
[0099] A test report is generated based on the type, location, and extent of damage. The test report includes basic information about the pear, damage assessment results, and corresponding treatment suggestions (e.g., mild damage can be stored at room temperature for a short period, moderate damage requires refrigeration, and severe damage is recommended for processing and utilization). The basic information includes the test number, test time, and maturity level.
[0100] The test report is stored in a cloud database and a traceable QR code is generated for easy subsequent query and quality traceability;
[0101] Determining the extent of latent damage includes:
[0102] A damage assessment system was constructed based on the percentage of damaged area, the degree of structural damage, and the degree of abnormality in physiological indicators. Mild damage was defined as ≤5% of the damaged area, structural damage involving only the shallow pulp, and an abnormality rate of physiological indicators ≤10%; moderate damage was defined as 5%-15% of the damaged area, structural damage involving the middle pulp, and an abnormality rate of physiological indicators 10%-20%; severe damage was defined as ≥15% of the damaged area, structural damage involving the deep pulp or the presence of cavities, and an abnormality rate of physiological indicators ≥20%.
[0103] Extract the damage feature parameters corresponding to latent damage, including the damage area ratio, the decrease in pulp density, the decrease in elastic modulus ratio, and the abnormality rate of physiological indicators;
[0104] Input the damage characteristic parameters into the damage degree evaluation system, and calculate the quantified value D of the damage degree by using the weighted summation method;
[0105] Determine the damage level according to the quantified value D of the damage degree. When D ≤ 0.05, it is a minor damage; when 0.05 < D ≤ 0.15, it is a moderate damage; when D > 0.15, it is a severe damage.
[0106] In the above embodiment, when it is determined that there is a hidden damage, further analyze the damage type, location and degree, realizing the in-depth detection from whether there is damage to how the damage is, significantly improving the practicality and application value of the detection result. By clarifying the damage type, it can provide a basis for tracing the cause of the damage; accurately positioning the damage location is convenient for taking targeted treatment measures; dividing the damage degree level provides scientific guidance for subsequent storage, processing and utilization, etc. The generated detection report contains comprehensive basic information, detailed damage assessment results and targeted treatment suggestions, making the detection result more readable and operable, facilitating relevant personnel to quickly master the quality status of Korla fragrant pears and take corresponding measures. Store the detection report in the cloud database and generate a traceable QR code, realizing the systematic management and full-process traceability of the detection data, which is not only convenient for subsequent query and statistical analysis, but also can improve the transparency and credibility of product quality control, providing strong support for the quality grading, market circulation and brand protection of Korla fragrant pears.
[0107] In the above embodiment, the core of constructing the damage degree evaluation system lies in comprehensively considering three key factors: damage area, structural damage and abnormal physiological indexes, realizing the quantitative evaluation of the hidden damage degree and avoiding the limitations of a single evaluation dimension. By clarifying the division criteria for different damage degrees, the determination of the damage level is more objective and unified, ensuring the consistency and comparability of the evaluation results. The extraction of damage characteristic parameters is highly targeted, directly reflecting the severity of the damage. The weighted summation method is used to calculate the quantified value of the damage degree, reasonably allocating the weights of each parameter, comprehensively reflecting the influence of different factors on the damage degree, and determining the damage level based on the quantified value, making the determination of the damage degree more scientific and accurate, and being able to provide clear and reliable basis for subsequent storage, processing and utilization.
[0108] Preprocess the comprehensive detection data set, including:
[0109] Detect the abnormal data in the data set, mark the data exceeding ±3 times the standard deviation as abnormal data and剔除 it; perform standardization processing on the data set after剔除 the abnormal data, convert different types of data to the [0,1] interval, and eliminate the dimension difference; perform data augmentation on the standardized data set by using random flipping, rotation, noise addition and data interpolation to expand the scale of the data set and generate a standardized detection data set.
[0110] This embodiment also includes updates and optimizations to the composite detection model:
[0111] New data on Korla fragrant pears and corresponding results of hidden damage verification are collected regularly. The verification results are confirmed through manual dissection and professional instrument testing.
[0112] The newly collected detection data and validation results were added to the training dataset to retrain the composite detection model. Based on the performance metrics of the retrained model, the model structure and parameters were adjusted to optimize the model's detection performance.
[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for detecting latent damage in Korla fragrant pears, characterized in that, Includes the following steps: Korla fragrant pears with uniform maturity and no obvious appearance damage were selected as the test objects. The appearance image data, internal structure scan data, mechanical property data and physiological index data of the fragrant pears were collected by multi-source detection equipment to form a comprehensive test dataset. The comprehensive detection dataset is preprocessed to obtain a standardized detection dataset; A composite detection model integrating deep learning and traditional machine learning is constructed. A standardized detection dataset is input into the composite detection model to extract feature information related to latent damage in pears. The feature information includes texture anomaly features, structural density features, elastic modulus features, and physiological metabolic features. Based on the aforementioned feature information, the probability value of latent damage is calculated. Combined with a preset damage judgment threshold, the system identifies whether there is latent damage in Korla pears and outputs the detection results.
2. The method for detecting latent damage in Korla fragrant pears as described in claim 1, characterized in that, The multi-source detection device includes: A high-resolution industrial camera is equipped with an infrared thermal imaging module to acquire RGB images and infrared thermal images of the surface of the pear, thereby obtaining appearance image data. The RGB images are used to extract microscopic texture anomalies and color distribution features, while the infrared thermal images are used to capture differences in heat conduction in damaged areas, thus helping to identify hidden damage to the skin and shallow flesh. X-ray tomography equipment is used for non-destructive scanning of pears to obtain internal structure scanning data, including pulp density distribution and cell structure integrity. A universal testing machine is used to perform compression and puncture tests on pears and collect mechanical property data, including elastic modulus, compressive strength and puncture force. A near-infrared spectroscopy analyzer is used to detect physiological indicators of pears, including sugar content, moisture content, and polyphenol content.
3. The method for detecting latent damage in Korla fragrant pears as described in claim 1, characterized in that, The construction process of the composite detection model includes: Samples of Korla fragrant pears with known latent damage were obtained, and corresponding comprehensive detection data were collected to construct training datasets, validation datasets, and test datasets, with the ratio of training datasets, validation datasets, and test datasets being 7:2:
1. A sub-model for appearance feature extraction based on convolutional neural network is built to extract texture anomaly features and color distribution features from appearance image data. A time-series feature extraction sub-model based on a recurrent neural network is used to extract time-series variation features from internal structure scanning data. A sub-model for classifying mechanical and physiological characteristics was built based on support vector machines, which was used to classify mechanical property data and physiological index data. A feature fusion layer is constructed, and the features output by the three sub-models are fused using a weighted summation method to obtain a comprehensive feature vector; A fully connected layer and a Softmax classifier are connected after the feature fusion layer to construct a complete composite detection model; The training dataset is input into the composite detection model for training. The learning rate and regularization coefficient of the model are adjusted using the validation dataset. The model performance is verified using the test dataset. The model construction is completed when the detection accuracy of the model reaches more than 95%.
4. The method for detecting latent damage in Korla fragrant pears as described in claim 1, characterized in that, The process of extracting the feature information includes: Texture features, color features, and infrared thermal imaging temperature distribution features of the surface of fragrant pears are extracted from appearance image data. The texture features include gray-level co-occurrence matrix and local binary mode. The color features include RGB color space components and HSV color space components. Texture anomaly features and temperature anomaly features related to latent damage are screened out. Structural density features are extracted from internal structural scanning data, including pulp density distribution parameters, intercellular space size, and the presence or absence of cavities or browning areas. Extract mechanical anomaly features from mechanical property data, including peak values and variation curves of elastic modulus, compressive strength, and puncture force; Physiological metabolic features are extracted from physiological index data, including the values and rates of change of sugar content, water content, and polyphenol content. By integrating texture features, color features, infrared thermal imaging temperature distribution features, structural density features, mechanical anomaly features, and physiological metabolic features, complete latent damage-related feature information is obtained.
5. The method for detecting latent damage in Korla fragrant pears as described in claim 1, characterized in that, Identifying whether Korla fragrant pears have latent damage based on the aforementioned feature information includes: Calculate the damage impact weight corresponding to each latent damage-related feature information, and the damage impact weight is determined by combining the analytic hierarchy process (AHP) with the entropy weight method. The relevant feature information of each latent injury is weighted and fused according to the injury impact weight to obtain a comprehensive feature value; the comprehensive feature value is compared with a preset injury judgment threshold, which is determined to be the optimal critical value through the receiver operating characteristic curve; If the comprehensive feature value is greater than or equal to the preset damage judgment threshold, then the Korla fragrant pear is judged to have latent damage; otherwise, the Korla fragrant pear is judged to have no latent damage.
6. The method for detecting latent damage in Korla fragrant pears as described in claim 5, characterized in that, The process of determining the weight of the damage impact includes: A feature hierarchy structure is constructed, with the target layer representing the accuracy of latent damage detection, the criterion layer representing appearance features, structural features, mechanical features, and physiological features, and the scheme layer representing the feature information related to each specific latent damage. Invite 5-8 experts in the field of agricultural testing to conduct pairwise comparisons of features at each feature level, construct a judgment matrix, calculate the maximum eigenvalue and corresponding eigenvector of the judgment matrix, and conduct a consistency test. The subjective weights of each latent injury-related feature information are calculated based on the effective judgment matrix; the information entropy of each latent injury-related feature information is calculated, and the objective weights are determined based on the information entropy. By combining subjective and objective weights using a weighted average method, the damage impact weights of each latent damage-related feature information are obtained.
7. The method for detecting latent damage in Korla fragrant pears as described in claim 1, characterized in that, Also includes: When it is determined that there is latent damage in Korla fragrant pears, the specific damage situation is further analyzed based on the extracted feature information. The specific damage situation includes damage type, damage location and damage degree. The damage type includes mechanical damage, physiological damage and disease damage. The damage location includes shallow pulp and deep pulp. The damage degree includes mild, moderate and severe. A test report is generated based on the type, location and extent of damage. The test report includes basic information about the pear, damage assessment results and corresponding treatment suggestions. The basic information includes the test number, test time and maturity level. Store the detection report in the cloud database and generate a traceable QR code at the same time.
8. The method for detecting latent damage in Korla fragrant pears as described in claim 7, characterized in that, Determine the degree of hidden damage, including: Construct an evaluation system for the degree of damage based on the proportion of damaged area, the degree of structural damage, and the degree of physiological index abnormality. Among them, for mild damage, the proportion of damaged area ≤ 5%, the structural damage only involves the superficial pulp, and the abnormal rate of physiological indexes ≤ 10%; for moderate damage, the proportion of damaged area is 5% - 15%, the structural damage involves the middle pulp, and the abnormal rate of physiological indexes is 10% - 20%; for severe damage, the proportion of damaged area ≥ 15%, the structural damage involves the deep pulp or there is a cavity, and the abnormal rate of physiological indexes ≥ 20%. Extract the damage characteristic parameters corresponding to the hidden damage. The damage characteristic parameters include the proportion of damaged area, the decrease amplitude of pulp density, the reduction ratio of elastic modulus, and the abnormal rate of physiological indexes. Input the damage characteristic parameters into the damage degree evaluation system and calculate the quantitative value D of the damage degree by using the weighted summation method. Determine the damage level according to the quantitative value D of the damage degree. When D ≤ 0.05, it is mild damage; when 0.05 < D ≤ 0.15, it is moderate damage; when D > 0.15, it is severe damage.
9. The method for detecting latent damage in Korla fragrant pears as described in claim 1, characterized in that, The preprocessing of the comprehensive detection data set includes: Detect the abnormal data in the detection data set, mark the data exceeding ±3 times the standard deviation as abnormal data and剔除 it; perform standardization processing on the data set after剔除 abnormal data, convert different types of data to the [0,1] interval, and eliminate the dimension difference; perform data augmentation on the standardized data set by using random flipping, rotation, noise addition, and data interpolation to expand the scale of the data set and generate a standardized detection data set.
10. The method for detecting latent damage in Korla fragrant pears as described in claim 1, characterized in that, It also includes the update and optimization of the composite detection model: Regularly collect new detection data of Korla fragrant pears and the corresponding verification results of hidden damage. The verification results are confirmed by manual dissection and professional instrument detection. Add the newly collected detection data and verification results to the training data set and retrain the composite detection model. Adjust the model structure and parameters based on the performance indicators of the retrained model to optimize the detection performance of the model.