An intelligent tea screening method and system

Abstract

The present invention relates to the technical field of intelligent tea screening, and discloses a tea intelligent screening method and system, including: collecting the morphological characteristics of tea leaves, calculating the texture information of tea leaves, fusing multi-modal data through a CNN-SVM hybrid model, establishing a non-linear mapping model of tea leaf morphology-texture, and generating unified screening features. Calculating the screening feature similarity of different varieties of tea leaves according to the unified screening features, constructing a feature transfer matrix, and converting the features of different tea varieties. Establishing a cross-variety adaptive screening tea model, and optimizing the screening discrimination boundary by using a domain adaptation loss function to improve the screening accuracy. The present invention enables the screening system to achieve adaptive migration of screening features between different varieties, reduces the cost of retraining, and improves the flexibility and scalability of the screening system. The screening model can maintain a high accuracy rate in different variety screening tasks, and can still perform efficient screening even when facing new tea varieties that have never been seen before.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent tea screening, and particularly to an intelligent tea screening method and system. Background Art

[0002] The screening of tea is crucial in the tea processing process, affecting the quality, uniformity and market value of tea. Traditional tea screening methods mainly rely on mechanical vibration screening, air separation and manual sorting. However, these methods usually have problems such as low screening accuracy, low efficiency and poor adaptability to tea varieties. In recent years, with the development of artificial intelligence, big data analysis and machine vision technology, intelligent screening has become a research hotspot in the tea processing field. For example, computer vision combined with image recognition technology can be used for the automatic recognition of tea morphology, and deep learning models can improve the accuracy of screening classification. In addition, multi-modal perception methods based on hyperspectral imaging and sensor technology have been applied to the field of food sorting to obtain more comprehensive physical property data. The progress of these technologies provides new possibilities for improving tea screening accuracy and adaptability. However, existing methods still have certain limitations when screening different varieties of tea, especially the weak generalization ability under small sample conditions, which is difficult to meet the industrialization requirements.

[0003] Although certain progress has been made in existing tea screening technologies, there are still many deficiencies. First of all, the screening method based on image processing mainly relies on the color and shape features of tea, lacking effective modeling of the internal texture features of tea, resulting in low adaptability to specific varieties. Secondly, the current intelligent screening system mainly uses a single data source, such as RGB images or near-infrared spectra, and fails to make full use of the complementarity of multi-modal data, affecting the screening accuracy. In addition, traditional machine learning methods (such as support vector machine SVM or random forest) are prone to feature redundancy when dealing with high-dimensional and multi-modal data, and it is difficult to efficiently fuse morphological and texture information. Furthermore, existing screening models are often trained based on large-scale labeled data, and there are many tea varieties and the data acquisition cost is relatively high, resulting in poor adaptability of the screening system to small sample data. When facing newly introduced tea varieties, traditional methods usually need to retrain the model, increasing the maintenance cost of the screening system. Finally, existing screening methods lack a cross-variety adaptation mechanism and fail to construct a unified screening feature space, making it difficult for the screening model to migrate between different varieties. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the existing tea screening methods have low screening accuracy, fail to effectively combine multi-modal data, are difficult to adapt to the screening requirements of different varieties, and the problems of how to improve the screening accuracy through multi-modal perception and transfer learning and maintain high-efficiency screening ability in a small sample environment.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: An intelligent tea screening method, including: collecting tea morphological characteristics, calculating tea texture information, fusing multi-modal data through a CNN-SVM hybrid model, establishing a non-linear mapping model of tea morphology-texture, and generating unified screening features;

[0007] Calculating the screening feature similarity of different varieties of tea according to the unified screening features, constructing a feature transfer matrix, and converting the features of different tea varieties;

[0008] Establishing a cross-variety adaptive screening tea model, and optimizing the screening discrimination boundary by using a domain adaptive loss function to improve the screening accuracy.

[0009] As a preferred scheme of the intelligent tea screening method described in the present invention, wherein: the collecting of tea morphological characteristics and calculating of tea texture information include:

[0010] Using hyperspectral imaging to obtain the morphological characteristics of tea, including the color, shape, and texture of tea;

[0011] Using a tactile pressure sensor to measure the deformation of tea under different pressures, and calculating the texture information of tea through a non-linear elastic model, including the stiffness and deformation coefficient of tea particles.

[0012] As a preferred scheme of the intelligent tea screening method described in the present invention, wherein: the non-linear elastic model includes:

[0013] The tactile pressure data is represented by a pressure matrix The force-deformation relationship of each measurement point is calculated by a modified Mooney-Rivlin hyperelastic model to obtain the stiffness characteristics, expressed as:

[0014] ;

[0015] Among them, represents the texture deformation coefficient of the tea sample, indicating the non-linear deformation degree of the tea under force. represents the maximum applied pressure on the tea. 、 represent the parameters of the hyperelastic model, indicating the material properties. represents the relative deformation rate. represents the pressure data matrix collected by the tactile sensor. Represents the texture feature calculation function, integrating the deformation coefficient and pressure data. Represents the tea texture feature matrix.

[0016] As a preferred solution of the intelligent tea screening method described in the present invention, wherein: the establishment of the non-linear mapping model of tea shape-texture by fusing multi-modal data through the CNN-SVM hybrid model includes:

[0017] Using CNN to extract the deep features of hyperspectral data , and SVM is used for classification boundary optimization, expressed as:

[0018] ;

[0019] Wherein, Represents the hyperspectral shape features extracted by CNN. Represents the CNN feature extraction function, which obtains high-dimensional features through convolution operations. Represents the feature fusion function, which is used for non-linear mapping in the feature space. Represents the final shape-texture fusion feature. Represents the tea screening category output by SVM classification. Represents the SVM classification function, which classifies based on the fusion features; Represents the original hyperspectral image data;

[0020] Establishing a non-linear mapping model of tea shape-texture to generate unified screening features includes, after data fusion, using support vector regression + Laplacian kernel for non-linear modeling, and using the Laplacian kernel function for non-linear mapping to map the fusion features to the screening space, expressed as:

[0021] ;

[0022] Wherein, Represents the final unified screening feature. Represents the weight coefficient of the SVR model. Represents the Laplacian kernel parameter, which controls the smoothness of feature mapping. Represents the L1 norm distance metric, which is used to measure the similarity between data points. Represents the number of training data points. Represents the th fusion feature vector of the training sample.

[0023] As a preferred solution of the intelligent tea screening method described in the present invention, wherein: the calculation of the screening feature similarity of different varieties of tea according to the unified screening features and the construction of the feature transfer matrix include:

[0024] To measure the distribution differences in the morphology and texture of different tea varieties, calculate the screening feature similarity matrix, measure the feature similarity between tea varieties through Mahalanobis distance, and use kernel density estimation to calculate the probability distribution difference of screening features, expressed as:

[0025] ;

[0026] where, represents the Mahalanobis distance between variety i and variety j. represents the screening feature vector of variety i. represents the screening feature vector of variety j. represents the screening feature similarity between variety i and variety j. represents the KDE kernel bandwidth parameter, controlling the smoothness. represents the feature mean between variety i and variety j. represents the sample points of the feature distribution; represents the joint covariance matrix, measuring the correlation of features;

[0027] Based on the screening feature similarity , construct a feature transfer matrix to transform the screening features of the new tea variety into the screening feature space of the existing varieties, and use the weighted feature projection method to construct the feature transformation relationship from variety i to variety j, expressed as:

[0028] ;

[0029] where, represents the element of the feature transfer matrix from variety i to variety j. represents the feature similarity calculated in the first step. represents the number of existing tea varieties. represents the feature contribution weight of variety k, ensuring the maximum retention of information during feature transformation. represents the feature distribution adjustment coefficient, controlling the non - linear degree of feature projection. represents the Euclidean distance, measuring the screening feature difference between variety i and variety k;

[0030] The value range is 0 - 1, and a value close to 1 indicates that the features between variety i and variety j can be directly transferred.

[0031] As a preferred scheme of the intelligent screening method for tea described in the present invention, wherein: the conversion of the features of different tea varieties includes:

[0032] Use the Laplacian eigenmap method to transform the screening features of the new tea variety n into the screening space of the existing varieties, expressed as:

[0033] ;

[0034] Among them, represents the screening characteristics of the new tea variety n after conversion. represents the transformation coefficient from variety n to variety j in the feature migration matrix. represents the feature divergence measure between variety n and variety j. represents the L1 norm distance, ensuring local similarity during feature transformation.

[0035] As a preferred solution of the intelligent tea screening method described in the present invention, wherein: the establishment of a cross-variety adaptable tea screening model includes:

[0036] Using a deep neural network to extract screening features and reducing the feature distribution difference between different varieties through feature adversarial training, so that the new variety can seamlessly adapt to the existing screening model, expressed as:

[0037] ;

[0038] Among them, represents the output of the screening features of the new tea variety n, which is used for final discrimination after being processed by the DNN. and represent the trainable weight matrices of the DNN, which are respectively used for non-linear transformation. and represent the bias terms. represents the Sigmoid activation function, ensuring output normalization. represents the hyperbolic tangent activation function, improving the feature expression ability;

[0039] As a preferred solution of the intelligent tea screening method described in the present invention, wherein: the adoption of a domain adaptation loss function to optimize the screening discrimination boundary and improve the screening accuracy includes:

[0040] There are inter-domain biases in the feature distributions of different varieties. Introducing a domain adaptation loss function to optimize the screening decision boundary, so that the screening model can maintain a high screening accuracy between different varieties;

[0041] Combining the maximum mean discrepancy and the adversarial loss to optimize feature alignment and classification stability at the same time, expressed as:

[0042] ;

[0043] Among them, represents the domain adaptation loss function. represents the number of training samples. and Represents the final screening characteristics of variety i and variety j. Represents the Euclidean distance metric, which is used to calculate the MMD loss to align the feature distributions of different varieties. Represents the loss weight hyperparameter. Represents the class label of the training samples. and Represents the classifier weights and bias terms. Represents the adversarial loss term, which ensures that the features can be used for screening classification;

[0044] The value range is [0, +∞). The smaller the value, the closer the screening feature distributions of different varieties are, and the better the adaptation effect.

[0045] As a preferred embodiment of the intelligent tea screening method described in the present invention, wherein:

[0046] The optimized screening discriminant boundary includes:

[0047] Adopt the maximum margin optimization method to ensure the robustness of the screening model among different varieties;

[0048] Use Lagrangian duality optimization to optimize the discriminant boundary of the screening categories, so that different varieties maintain classification stability during the screening process, expressed as:

[0049] ;

[0050] Wherein, Represents the discriminant boundary optimization loss. and Represents the Lagrange multiplier. and Represents the class label of the training samples. Represents the kernel function, which measures the non-linear similarity between screening features;

[0051] The value range is from negative infinity to positive infinity. During the optimization process, the smaller the value, the clearer the discriminant boundary and the higher the screening accuracy;

[0052] The cross-variety screening model with optimized discriminant boundary is applicable to the intelligent screening tasks of different tea varieties.

[0053] An intelligent tea screening system, characterized in that it includes,

[0054] A multi-modal perception feature extraction module, which collects the morphological features of tea leaves, calculates the texture information of tea leaves, fuses multi-modal data through a CNN-SVM hybrid model, establishes a non-linear mapping model of tea leaf morphology-texture, and generates unified screening features;

[0055] Build a module for constructing the characteristic migration matrix of tea varieties, calculate the similarity of screening characteristics of different tea varieties according to the unified screening characteristics, construct the characteristic migration matrix, and transform the characteristics of different tea varieties;

[0056] A screening optimization module in a small-sample environment, establish a cross-variety adaptation screening tea model, and use a domain adaptation loss function to optimize the screening discrimination boundary to improve the screening accuracy.

[0057] The beneficial effects of the present invention:

[0058] Collect the morphological characteristics and texture information of tea leaves, establish a non-linear mapping model. Through this step, the morphological and texture information of tea leaves is accurately quantified, and after data fusion, a unified screening characteristic is formed, enhancing the adaptability of the screening system among varieties. Compared with the traditional single image recognition method, this step significantly improves the screening accuracy and reduces the screening misjudgment problem caused by tea variety differences, providing solid data support for subsequent feature migration;

[0059] Calculate the similarity of screening characteristics of different tea varieties, construct the characteristic migration matrix. Through this step, the screening system changes from "specific variety adaptation" to "cross-variety adaptation", can automatically learn the screening characteristics of new tea varieties, achieve rapid migration, and reduce the costs of data collection and model training. The implementation of this step enables the screening system to have the ability of adaptive learning, overcomes the drawbacks of traditional screening methods that cannot be universal among different varieties, and improves the scalability and practicality of intelligent screening;

[0060] Establish a cross-variety adaptation screening tea model and optimize the screening discrimination boundary. Through this step, the screening system not only has the ability of cross-variety adaptation, but also can maintain high accuracy in the screening process of different varieties, avoiding the decline of screening performance caused by variety feature drift. Compared with the traditional fixed screening standard, the screening model of the present invention continuously optimizes the screening decision boundary through adaptive learning, improves the stability of the model, and enables it to still have high screening accuracy and generalization ability in the case of small samples. Brief Description of the Drawings

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to these drawings. Among them:

[0062] Figure 1 It is the overall flowchart of a tea intelligent screening method and system provided by the first embodiment of the present invention. Detailed Embodiment

[0063] To make the above objects, features, and advantages of the present invention more apparent and understandable, the following provides a detailed description of the specific embodiments of the present invention in conjunction with the accompanying drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0064] Example 1, referring to Figure 1 , which is an embodiment of the present invention, provides an intelligent tea screening method, including:

[0065] S1: Collect the morphological characteristics of tea leaves, calculate the texture information of tea leaves, fuse multi-modal data through a CNN-SVM hybrid model, establish a non-linear mapping model of tea leaf morphology-texture, and generate unified screening features.

[0066] Use hyperspectral imaging to obtain the morphological characteristics of tea leaves, including the color, shape, and texture of tea leaves.

[0067] It should be noted that hyperspectral imaging (Hyperspectral Imaging, HSI) can obtain the reflection information of tea leaves in different spectral bands to extract their morphological characteristics, including color, shape, and texture. Represent tea leaf samples using a high-dimensional spectral data matrix, and screen high-value spectral bands through information entropy to remove redundant data and improve data quality.

[0068] Furthermore, assuming that the tea leaf reflectance data collected by the hyperspectral imaging system is a three-dimensional tensor R (wavelength × space × sampling point), first perform information entropy screening on the band data to reduce redundant data and improve classification efficiency, expressed as:

[0069] ;

[0070] where represents the information entropy value at wavelength , measuring the effective information carried by this wavelength. represents the normalized probability density of the i-th spectral band. represents the set of optimal wavelengths selected for subsequent feature extraction. represents the hyperspectral data matrix, containing the tea leaf reflectance information at different wavelengths. represents the hyperspectral feature extraction function, whose role is to select spectral bands with high information content. represents the hyperspectral morphological feature matrix after screening.

[0071] It should be noted that the deficiencies of the prior art include that RGB images are greatly affected by ambient light, and there are color deviations in the images collected under different lighting conditions, which affects the screening consistency. Low-dimensional image data is difficult to accurately distinguish tea varieties, and the internal quality of tea cannot be accurately reflected only by color and shape information, which easily leads to misclassification. The morphological feature extraction method is single. Traditional screening systems often classify based on artificially set color thresholds or morphological features and cannot adapt to complex varietal changes.

[0072] However, our invention introduces hyperspectral imaging, which extends to multiple spectral channels, enabling the screening system to capture more detailed information and improve the ability to distinguish different varieties. The spectral normalization method (such as Standard Normal Variate transformation SNV, Principal Component Analysis PCA) is used to reduce the impact of light changes on the screening accuracy and ensure feature stability. Combining with a deep learning model (CNN) for automatic feature extraction to avoid the limitations of traditional manual feature engineering and make the screening more intelligent.

[0073] Our invention improves the robustness of the screening system, which is not affected by ambient light and can maintain a consistent screening effect even under different acquisition conditions. It enhances the ability to distinguish varieties. Hyperspectral data provides richer spectral information, enabling the screening system to classify based on finer-grained features. Automated screening feature extraction, without the need for artificially set thresholds, improves adaptability, enabling the system to more efficiently handle the screening tasks of different varieties.

[0074] A tactile pressure sensor is used to measure the deformation of tea leaves under different pressures, and through a non-linear elastic model, the texture information of the tea leaves is calculated, including the stiffness and deformation coefficient of the tea particles.

[0075] Furthermore, the texture characteristics of tea leaves such as softness, elasticity, and granularity will affect the screening effect. We use a tactile pressure sensor to measure the deformation of tea leaves under different pressures. Through a non-linear elastic model, key parameters such as the stiffness and deformation coefficient of the tea particles are calculated and converted into numerical features that can be used for machine learning modeling.

[0076] The tactile pressure data is represented by a pressure matrix The force and deformation relationship at each measurement point is used to calculate the stiffness feature through a modified Mooney-Rivlin hyperelastic model, which is expressed as:

[0077] ;

[0078] where represents the texture deformation coefficient of the tea leaf sample, indicating the degree of non-linear deformation of the tea leaf under force. represents the maximum applied pressure on the tea leaf. 、 Parameters representing the hyperelastic model, representing material properties. Represents the relative deformation rate. Represents the pressure data matrix collected by the tactile sensor. Represents the texture feature calculation function, integrating the deformation coefficient and pressure data. Represents the tea texture feature matrix.

[0079] Furthermore, the deficiencies of the prior art include that a single visual feature cannot characterize the physical texture of tea, and the screening system only relies on morphological features, resulting in misclassification of some varieties due to similar morphology. Traditional texture measurement methods are inefficient. For example, manual kneading tests are highly subjective and have inconsistent standards, making it difficult to apply to automated screening systems. Existing sensing technologies lack the ability of non-linear modeling, and traditional hardness measurement methods (such as compression tests) can only provide simple mechanical parameters and cannot accurately describe the dynamic deformation characteristics of tea.

[0080] However, our invention uses a tactile pressure sensor to measure the deformation data of tea in real time after being stressed, and obtains physical characteristics such as stiffness and elasticity. Combining with the modified Mooney-Rivlin hyperelastic model to establish a non-linear relationship, enabling the system to accurately characterize the complex mechanical behavior of tea. Using data fusion technology, combining texture information with hyperspectral features to avoid the limitations of classifying relying only on a single data source and improve the screening accuracy.

[0081] Our invention significantly improves the accuracy of the screening model, avoids misjudgment caused by similar morphology, and has a stronger discrimination ability for high-end tea varieties in particular. Enhances the adaptability of the screening system, and can accurately classify through texture features regardless of the change in the size of tea particles. Automated measurement reduces manual intervention and greatly improves the efficiency of the screening system, making it more suitable for large-scale industrial applications.

[0082] It should be noted that CNN is used to extract the morphological features of hyperspectral images, SVM is used for classification, and hyperspectral and tactile features are combined in the fusion layer to achieve non-linear mapping. The key is to extract key morphological features through convolution operations and use the feature projection function to map high-dimensional data to a unified screening feature space.

[0083] Use CNN to extract the deep features of hyperspectral data , and SVM is used for classification boundary optimization, expressed as:

[0084] ;

[0085] Among them, Represents the hyperspectral morphological features extracted by CNN. Represents the CNN feature extraction function, which obtains high-dimensional features through convolution operations. denotes the feature fusion function, which is used for non - linear mapping in the feature space. denotes the final shape - texture fusion feature. denotes the tea sieve classification category output by the SVM classification. denotes the SVM classification function, which classifies based on the fusion feature. denotes the original data of the hyperspectral image.

[0086] It should be noted that through the above data fusion process, a mathematical model needs to be established to map the shape and texture features of tea leaves to the sieve feature space, which is ultimately used for subsequent transfer learning and variety adaptation. Support Vector Regression (SVR)+Laplacian kernel is used for non - linear modeling to ensure the smoothness and generalization ability of feature mapping.

[0087] Establish a non - linear mapping model for tea leaf shape - texture to generate unified sieve features. After data fusion, support vector regression + Laplacian kernel is used for non - linear modeling. The non - linear mapping uses the Laplacian kernel function to map the fusion feature to the sieve space, which is expressed as:

[0088] ;

[0089] where denotes the final unified sieve feature. denotes the weight coefficient of the SVR model. denotes the Laplacian kernel parameter, which controls the smoothness of feature mapping. denotes the L1 - norm distance metric, which is used to measure the similarity between data points. denotes the number of training data points. denotes the th fusion feature vector of the training sample.

[0090] This step finally outputs a unified sieve feature matrix , which represents the shape and texture features of tea leaves and is subsequently used to construct a feature transfer matrix to achieve cross - variety sieve adaptation.

[0091] It should be noted that the defects of the existing technology include that the multi - modal data fusion is difficult, and it is difficult for traditional methods to effectively combine visual features and mechanical features. The feature dimension is too high and the computational complexity is large, and it is difficult for existing machine learning methods to efficiently process high - dimensional and multi - modal data. The linear classification method performs poorly on high - dimensional data, and traditional sieve models (such as SVM or decision tree) are prone to classification errors when facing complex tea varieties.

[0092] In our invention, CNN is used to extract the hyperspectral morphological features, reducing the need for manual feature engineering and improving the degree of automation of classification. SVM is used for classification optimization to ensure that the screening system still has good generalization ability in the small-sample environment. Laplacian kernel regression is introduced to construct a non-linear mapping model, enabling the morphological features and texture features to be jointly analyzed in the same screening space and improving the classification accuracy.

[0093] Our invention conducts deep fusion of multi-modal data, improving the adaptability of the screening system to different varieties and enabling the system to classify based on richer information. The screening accuracy is improved, the discrimination ability for complex varieties is enhanced through non-linear mapping, and the screening misjudgment caused by single features is reduced. The computing efficiency is enhanced. The combination of CNN + SVM performs excellently in high-dimensional data processing, ensuring that the screening system can operate efficiently in practical applications.

[0094] Preferably, in this step, the morphological information (color, shape, texture, etc.) of the tea leaves is obtained through hyperspectral imaging, and the physical texture (stiffness, deformation coefficient, etc.) of the tea leaves is collected using tactile pressure sensors. Subsequently, a CNN-SVM hybrid model is used to fuse these heterogeneous data, construct a non-linear mapping model of tea leaf morphology-texture, and finally generate unified screening features. Hyperspectral imaging can provide much finer-grained spectral data than ordinary RGB cameras, enabling the screening system to capture the microscopic structure and quality characteristics of tea leaves and improving the classification accuracy. The tactile pressure sensor makes up for the defect that it is difficult to identify the internal texture information of tea leaves only relying on visual features, thereby improving the generalization ability of the screening model for different varieties. The CNN-SVM hybrid model realizes the deep fusion of hyperspectral data and tactile data, and establishes a unified feature space through non-linear mapping, laying a foundation for subsequent feature migration.

[0095] Furthermore, through this step, the morphological and texture information of the tea leaves is accurately quantified, and unified screening features are formed after data fusion, enhancing the adaptability of the screening system among varieties. Compared with traditional single image recognition methods, this step significantly improves the screening accuracy and reduces the screening misjudgment problem caused by tea leaf variety differences, providing solid data support for subsequent feature migration.

[0096] S2: Calculate the screening feature similarity of different varieties of tea leaves according to the unified screening features, construct a feature migration matrix, and transform the features of different tea leaf varieties.

[0097] To measure the distribution differences in morphology and texture among different varieties of tea leaves, calculate the screening feature similarity matrix, measure the feature similarity among tea leaf varieties through Mahalanobis distance, and calculate the probability distribution difference of the screening features using kernel density estimation, expressed as:

[0098] ;

[0099] Among them, represents the Mahalanobis distance between variety i and variety j. represents the screening feature vector of variety i. represents the screening feature vector of variety j. represents the screening feature similarity between variety i and variety j. represents the KDE kernel bandwidth parameter, which controls the smoothness. represents the feature mean between variety i and variety j. represents the sample points of the feature distribution. represents the joint covariance matrix, which measures the correlation of features.

[0100] It should be noted that the Mahalanobis distance is applicable to measuring the similarity of different categories of data, while kernel density estimation can reduce data noise and improve the robustness of feature matching.

[0101] Furthermore, the defects of the existing technologies include that traditional screening methods are based on fixed classification boundaries and cannot be adaptively adjusted according to the similarity between varieties, resulting in possible misclassification of newly introduced varieties. Traditional metric methods such as Euclidean distance are unstable in high-dimensional spaces and are difficult to effectively measure the similarity of high-dimensional feature data. Ignoring the influence of data distribution, existing methods only rely on point-to-point similarity calculation and do not fully consider the overall distribution pattern of tea leaf features, affecting the reliability of feature similarity.

[0102] However, our invention uses the Mahalanobis distance instead of the Euclidean distance, enabling the consideration of the correlation between features when calculating similarity and improving the stability of feature matching. Combining kernel density estimation (KDE), it models the distribution of screening features between varieties to avoid calculation errors caused by feature data distribution deviations. A feature similarity matrix is constructed to enable the quantitative representation of features of different varieties, providing an accurate calculation basis for subsequent feature migration.

[0103] The screening system of our invention can accurately measure the similarity between different tea varieties and improve the adaptability of new tea varieties. The calculation is more stable, avoiding the common distance failure problem in high-dimensional data and improving the generalization ability of the screening model. It provides a basis for feature migration, and subsequent steps can use the similarity matrix for feature transformation to enable new tea varieties to match existing screening models.

[0104] Based on the screening feature similarity , a feature migration matrix is constructed to transform the screening features of new tea varieties into the screening feature space of existing varieties. A feature weight adjustment strategy is adopted to ensure that the transformed features retain the maximum amount of information while reducing feature distribution deviations. Using the weighted feature projection method, the feature transformation relationship from variety i to variety j is constructed, expressed as:

[0105] ;

[0106] Among them, represents the element of the feature transfer matrix from variety i to variety j. represents the feature similarity calculated in the first step. represents the number of existing tea varieties. represents the feature contribution weight of variety k, ensuring the maximum retention of information during feature transformation. represents the feature distribution adjustment coefficient, controlling the non - linear degree of feature projection. represents the Euclidean distance, measuring the difference in screening features between variety i and variety k.

[0107] The value range is 0 - 1. A value close to 1 indicates that the features between variety i and variety j can be directly transferred.

[0108] Furthermore, based on the constructed feature transfer matrix, the feature transformation of new tea varieties is realized so that they can be adapted to the existing screening model. A feature mapping function based on Laplacian regularization is adopted to ensure that the transformed features maintain the screening discriminability.

[0109] It should be noted that the defects of the existing technology include that the existing screening methods lack the cross - variety adaptation ability. When introducing new tea varieties, the model must be retrained, resulting in too high maintenance costs for the screening system. Simple feature weighting methods are prone to information loss. Traditional weighting methods are likely to cause information redundancy or loss in high - dimensional feature spaces, affecting the screening effect. Feature projection is unstable, and some methods will cause feature distribution deviation during the mapping process, affecting the accuracy of subsequent screening.

[0110] Our invention constructs a weighted feature projection matrix to ensure that the information of the original features is retained to the greatest extent during the feature transformation of different varieties, improving the adaptability. Combining regularization optimization to prevent data distribution deviation during feature projection and ensuring the stability of screening features. Dynamically adjusting the transfer weight to ensure that the screening features of different varieties can effectively match the existing screening model and improve the adaptation degree.

[0111] Our invention improves the cross - variety adaptation ability of the screening system, enabling efficient feature conversion between different varieties, improving the screening accuracy of new tea varieties. Reducing the need for model retraining, lowering the maintenance cost of the screening system, and improving the flexibility of the system. Enhancing the generalization ability of the screening model to ensure stable operation in a multi - variety environment.

[0112] Using the Laplacian eigenmap method, the screening features of the new tea variety n are transformed into the screening space of the existing varieties, expressed as:

[0113] ;

[0114] Among them, represents the screening characteristics of the new tea variety n after conversion. represents the transformation coefficient from variety n to variety j in the feature migration matrix. represents the feature divergence measure between variety n and variety j. represents the L1 norm distance, ensuring local similarity during feature transformation.

[0115] Preferably, based on the unified screening characteristics extracted in S1, this step calculates the similarity of the screening characteristics of different tea varieties and constructs a feature migration matrix based on this for the transformation of the screening characteristics of different varieties, enabling the new tea variety to adapt to the existing screening model. This process uses the Mahalanobis Distance to calculate the similarity of the characteristics of different varieties and combines the weighted feature projection technology to establish the mapping relationship between varieties, thereby reducing the feature distribution deviation of cross-variety screening.

[0116] The calculation of feature similarity can effectively measure the screening feature deviation between different varieties, making the screening system no longer limited to a specific variety but able to generalize to multiple tea types.

[0117] The construction of the feature migration matrix enables the newly introduced tea variety to quickly adapt to the existing screening system in the case of small samples or even no samples, greatly reducing the cost of model retraining.

[0118] The weighted feature projection optimizes the accuracy of feature migration, ensures the minimum information loss during the screening process, and enhances the stability of cross-variety screening.

[0119] Furthermore, through this step, the screening system changes from "specific variety adaptation" to "cross-variety adaptation", can automatically learn the screening characteristics of new tea varieties, achieve rapid migration, and reduce the costs of data collection and model training. The implementation of this step enables the screening system to have the ability of adaptive learning, overcomes the drawbacks of traditional screening methods that cannot be generalized between different varieties, and improves the scalability and practicality of intelligent screening.

[0120] S3: Establish a cross-variety adaptation screening tea model, and use a domain adaptation loss function to optimize the screening discrimination boundary to improve the screening accuracy.

[0121] It should be noted that after the feature migration is completed in the second step, the transformed screening characteristics Sieving discrimination still needs to be carried out. The goal of this small step is to establish a cross-variety sieving model based on a deep neural network so that new tea varieties can be adapted to the existing sieving categories. We adopt a two-stream neural network structure, with one branch for processing known varieties and the other branch for processing newly introduced varieties, and finally fuse them at the shared layer to optimize the classification ability.

[0122] Use a deep neural network to extract sieving features and reduce the feature distribution differences between different varieties through feature adversarial training, so that new varieties can seamlessly adapt to the existing sieving model, expressed as:

[0123] ;

[0124] Among them, represents the sieving feature output of the new tea variety n, which is used for the final discrimination after being processed by the DNN. and represent the trainable weight matrices of the DNN, which are used for non-linear transformation respectively. and represent the bias terms. represents the Sigmoid activation function to ensure output normalization. represents the hyperbolic tangent activation function to improve the feature expression ability.

[0125] It should be noted that the defects of the existing technology include that the existing sieving model is limited to specific varieties and it is difficult to generalize and apply between different tea varieties. There is a lack of a feature fusion mechanism, and traditional sieving methods only rely on single features and cannot make full use of multi-modal information for adaptation. The risk of model overfitting is high. When facing new tea varieties, the traditional method is prone to the problem of decreased sieving accuracy.

[0126] However, our invention uses DNN for feature fusion, jointly models morphological features and texture features, and improves the generalization ability of the model. Combining the transfer learning strategy, optimizing the sieving rules of new varieties through the data of existing varieties, and improving the model adaptability. Using self-supervised learning methods, enabling the model to still maintain a high sieving accuracy under the condition of a small amount of data.

[0127] Our invention improves the generalization ability of the sieving system, enables it to be applicable to different tea varieties, and reduces the model update cost. Enhances the sieving accuracy, combines multi-modal data for discrimination, and improves the stability of the sieving effect. Reduces the dependence on large-scale training data, enabling the sieving system to still operate efficiently in a small-sample environment.

[0128] There are inter-domain biases in the feature distributions of different varieties. An inter-domain adaptive loss function is introduced to optimize the sieving decision boundary, so that the sieving model can maintain a high sieving accuracy between different varieties.

[0129] Combining the maximum mean discrepancy and adversarial loss, and optimizing feature alignment and classification stability simultaneously, which is expressed as:

[0130] ;

[0131] where, represents the domain adaptation loss function. represents the number of training samples. and represent the final screening features of variety i and variety j. represents the Euclidean distance metric, which is used to calculate the MMD loss to align the feature distributions of different varieties. represents the loss weight hyperparameter. represents the class label of the training samples. and represent the classifier weights and bias terms. represents the adversarial loss term, which ensures that the features can be used for screening classification.

[0132] The value range is [0, +∞), and the smaller the value, the closer the screening feature distributions of different varieties are, and the better the adaptation effect.

[0133] Adopt the maximum margin optimization method to ensure the robustness of the screening model among different varieties.

[0134] Using Lagrangian dual optimization to optimize the discriminant boundary of the screening categories, so that different varieties maintain classification stability during the screening process, which is expressed as:

[0135] ;

[0136] where, represents the discriminant boundary optimization loss. and represent the Lagrange multipliers. and represent the class labels of the training samples. represents the kernel function, which measures the non-linear similarity between screening features.

[0137] The value range is from negative infinity to positive infinity. During the optimization process, the smaller the value, the clearer the discriminant boundary and the higher the screening accuracy.

[0138] The cross-variety screening model optimized by the discriminant boundary is applicable to the intelligent screening tasks of different tea varieties.

[0139] Furthermore, the deficiencies of the prior art include that the screening boundary is greatly affected by data distribution, and different varieties may result in a decrease in screening accuracy due to distribution deviation. There is a lack of an adaptive adjustment mechanism, and it is difficult for existing methods to adaptively adjust the screening rules according to the characteristics of new varieties. The optimization method is too static, and traditional methods cannot dynamically optimize the screening decision boundary.

[0140] In contrast, our invention uses MMD to reduce the feature deviation between varieties, ensuring that the screening rules are consistent among different varieties. Combining adversarial training, the robustness of the model is improved by generating adversarial samples. The domain adaptation method is used to optimize the screening rules, ensuring that the model can dynamically adapt to new tea varieties.

[0141] Our invention improves the robustness of the screening system and reduces the classification error between varieties. It enhances the screening accuracy to ensure stable classification of different tea varieties. It strengthens the adaptive ability, enabling the system to adapt to changing tea varieties.

[0142] Preferably, with the support of the feature transfer matrix, this step establishes a cross-variety adaptable tea screening model and optimizes the screening discriminant boundary through the domain adaptation loss function to improve the screening accuracy. Specifically, this step uses a deep neural network (DNN) to extract tea screening features, and at the same time uses the maximum mean discrepancy (MMD) to calculate the feature distribution differences between different varieties, and combines adversarial loss function training, so that the screening model can maintain stable classification performance even when facing new tea varieties that have never been seen before.

[0143] The deep neural network (DNN) further optimizes the screening features, making them have stronger discriminative ability among different varieties.

[0144] The domain adaptation loss function reduces the feature distribution deviation between different varieties through the maximum mean discrepancy (MMD), enabling the screening system to adapt to new tea varieties without being affected by variety feature drift.

[0145] Adversarial loss training makes the discriminant boundary of the screening model more robust through adaptive learning, improving the screening accuracy in a small sample environment.

[0146] Through this step, the screening system not only has the cross-variety adaptation ability, but also can maintain high accuracy in the screening process of different varieties, avoiding the decline of screening performance caused by variety feature drift. Compared with the traditional fixed screening standard, the screening model of the present invention continuously optimizes the screening decision boundary through adaptive learning, improving the stability of the model and enabling it to still have high screening accuracy and generalization ability in small sample cases.

[0147] In the above embodiments, there is also an intelligent tea screening system, specifically:

[0148] The multi-modal perception feature extraction module collects the morphological features of tea leaves, calculates the texture information of tea leaves, fuses multi-modal data through a CNN-SVM hybrid model, establishes a non-linear mapping model of tea leaf morphology-texture, and generates unified screening features.

[0149] The module for constructing the tea variety feature transfer matrix calculates the screening feature similarity of different tea varieties according to the unified screening features, constructs a feature transfer matrix, and transforms the features of different tea varieties.

[0150] The screening optimization module in a small sample environment establishes a cross-variety adaptive screening model for tea leaves, uses a domain adaptation loss function to optimize the screening discrimination boundary, and improves the screening accuracy.

[0151] A computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method described above are implemented.

[0152] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method described above are implemented.

[0153] The computer device can be a server. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the data cluster data of the power monitoring system. The input / output interface of the computer device is used for the processor to exchange information with external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. When the computer program is executed by the processor, it implements an intelligent screening method for tea leaves.

[0154] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memories can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0155] Embodiment 2, an embodiment of the present invention, provides a method and system for intelligent screening of tea. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0156] The purpose of this experiment is to verify the effectiveness of the intelligent tea screening method proposed by the present invention in terms of screening accuracy, feature migration ability, and adaptability to different tea varieties. The experiment uses hyperspectral imaging and tactile sensors to obtain the morphological and texture characteristics of tea, and establishes a tea morphology-texture mapping model through a CNN-SVM hybrid model. Then, a feature migration matrix is used to optimize cross-variety screening, and finally, the screening discrimination boundary is optimized through a domain adaptation loss function to improve the screening accuracy.

[0157] To verify the effectiveness of the proposed intelligent tea screening method, the experiment conducted screening tests among multiple tea varieties, used hyperspectral imaging technology to extract the morphological characteristics of tea, and combined with tactile pressure sensors to measure the texture information of tea. Six different varieties of tea (Longjing, Biluochun, Dahongpao, Tieguanyin, Pu'er, Baihao Yinzhen) were selected for the experiment, and a CNN-SVM hybrid model was used to fuse hyperspectral data and tactile data to establish a tea morphology-texture mapping model.

[0158] The experimental equipment included: hyperspectral imaging equipment (collecting hyperspectral data in the wavelength range of 400nm - 1000nm), tactile pressure sensors (measuring the stiffness and deformation coefficient of tea), computer vision processing system (for CNN model training and classification calculation), and screening device (for screening different categories of tea).

[0159] The experimental steps are as follows:

[0160] Collect hyperspectral data of six kinds of tea, and analyze morphological characteristics such as color, shape, and texture. Measure the deformation of tea under different pressures through tactile sensors, and calculate its stiffness and deformation coefficient. Use the CNN network to extract the deep features of hyperspectral data and remove redundant information. Use the SVM classification model to optimize the classification of the extracted features and improve the recognition accuracy. Perform non-linear mapping on the morphological and texture data through the SVR model and Laplace kernel to generate unified screening features. Calculate the similarity of screening features of different varieties of tea and construct a feature transfer matrix. Calculate the similarity of screening features of different varieties of tea through Mahalanobis distance to reduce the classification error between different varieties. Optimize the screening decision through a deep neural network (DNN) to improve the classification adaptability between different varieties. Use a domain adaptation loss function to optimize the screening discrimination boundary and improve the classification accuracy. Use the maximum mean discrepancy (MMD) to optimize the cross-variety screening accuracy so that new tea varieties can adapt to the existing classification model.

[0161] The experiment used 600 tea samples for testing, with 100 samples for each variety, and recorded the experimental data parameters as shown in Table 1 and Table 2.

[0162] Table 1 Experimental Data

[0163]

[0164] Table 2 Calculation Data

[0165]

[0166] The experimental data shows that through the intelligent screening method, the characteristics of different varieties of tea in terms of color, shape, texture, and texture are accurately extracted and used for classification optimization.

[0167] The screening eigenvalue is used to represent the comprehensive characteristics of tea during the screening process. The higher the value, the more obvious the screening characteristics and the easier it is to classify. By observing the data, Dahongpao has the highest screening eigenvalue (0.88), indicating that its morphological and texture characteristics are more obvious during the screening process and are easier to distinguish. While Baihao Yinzhen has the lowest (0.80), suggesting that its characteristics are less obvious and may require a more refined classification strategy.

[0168] The feature similarity is used to measure the degree of similarity of characteristics between different varieties. The higher the value, the stronger the transferability between varieties. The highest value appears in Dahongpao (0.94), indicating that its screening characteristics have a higher similarity with other varieties. While Baihao Yinzhen has the lowest feature similarity (0.85), suggesting that its characteristics are relatively independent and are more difficult to transfer to other varieties.

[0169] The transfer matrix weight is used to measure the importance of different varieties of tea in cross-variety screening. The higher the weight, the more of its characteristics are retained during the transfer process. Dahongpao has the highest transfer matrix weight (0.91), indicating that its screening characteristics can be better transferred to other varieties. While Baihao Yinzhen has the lowest (0.81), suggesting that its screening characteristics are more difficult to transfer.

[0170] The classification boundary optimization loss measures how well the screening model optimizes the classification boundaries for different varieties. The smaller the value, the higher the classification stability of the model. Dahongpao (0.012) and Tieguanyin (0.014) have the lowest losses, indicating that their classification stabilities are better and the screening boundaries are clear. While Baihao Yinzhen has the largest loss (0.022), suggesting that it is more difficult to optimize its classification boundary.

[0171] The domain adaptation loss is used to measure the adaptability of the screening model to different varieties. The smaller the value, the more similar the characteristic distributions of different varieties are and the stronger the adaptability of the screening system. Dahongpao has the lowest domain adaptation loss (0.038), indicating that its screening characteristics have a higher matching degree with the model. While Baihao Yinzhen has the highest loss (0.050), suggesting that the screening characteristics of this variety are quite different from those of other varieties and further optimization of the adaptation strategy is needed.

[0172] The final classification accuracy of the method of the present invention is higher than 90% for all cases. Among them, Pu'er is the highest (95.1%), indicating that its screening characteristics are the most stable and it is easy to classify. While Baihao Yinzhen is the lowest (91.6%), suggesting that there are still certain challenges in the classification process of this variety.

[0173] Through the intelligent screening method of the present invention, the screening characteristics of tea leaves are accurately modeled, the differences in screening characteristics of different varieties are clear, and it can adapt to the screening requirements of different varieties. Compared with the traditional classification method based on color and shape, the present invention utilizes hyperspectral data, tactile sensor information, multi-modal fusion, and deep learning to achieve a higher classification accuracy, a more stable screening boundary, and a stronger cross-variety adaptation ability, providing a more intelligent and efficient screening solution for the tea processing industry.

[0174] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An intelligent screening method for tea leaves, characterized in that, Including: Collect the morphological characteristics of tea leaves, calculate the texture information of tea leaves, fuse multi-modal data through a CNN-SVM hybrid model, establish a non-linear mapping model of tea leaf morphology-texture, and generate unified screening features; Calculate the screening feature similarity of different tea varieties according to the unified screening features, construct a feature transfer matrix, and transform the features of different tea varieties; Establish a cross-variety adaptable tea leaf screening model, and use a domain adaptation loss function to optimize the screening discrimination boundary to improve the screening accuracy; The collection of tea leaf morphological characteristics and the calculation of tea leaf texture information include: Use hyperspectral imaging to obtain the morphological characteristics of tea leaves, including the color, shape, and texture of tea leaves; Use a tactile pressure sensor to measure the deformation of tea leaves under different pressures, and calculate the texture information of tea leaves through a non-linear elastic model, including the stiffness and deformation coefficient of tea leaf particles; The non-linear elastic model includes: The tactile pressure data is represented by a pressure matrix P, and the force-deformation relationship of each measurement point is calculated by a modified Mooney-Rivlin hyperelastic model to obtain the stiffness characteristics, expressed as: X touch = f t (D s , P) Among them, D s represents the texture deformation coefficient of the tea sample, indicating the degree of non-linear deformation of the tea under force; F max represents the maximum applied pressure on the tea; C1 and C2 represent the parameters of the hyperelastic model, indicating material properties; λ represents the relative deformation rate; P represents the pressure data matrix collected by the tactile sensor; f t (·) represents the texture feature calculation function, integrating the deformation coefficient and pressure data; X touch represents the tea texture feature matrix; The fusion of multi-modal data through a CNN-SVM hybrid model and the establishment of a non-linear mapping model of tea leaf morphology-texture include: Using CNN to extract hyperspectral morphological feature X cnn , and the SVM for classification boundary optimization is expressed as: X cnn = g c (X hsi ) X fusion = Φ(X cnn , X touch ) Y = f svm (X fusion ) Among them, X cnn represents the hyperspectral morphological features extracted by the CNN; X hsi represents the original data of the hyperspectral image; g c (·) represents the CNN feature extraction function, which obtains high-dimensional features through convolution operations; Φ(·) represents the feature fusion function, which is used for non-linear mapping in the feature space; X fusion represents the final morphological-textural fusion features; Y represents the tea sieve classification category output by the SVM classification; f svm (·) represents the SVM classification function, which performs classification based on the fusion features; Establish a non-linear mapping model of tea leaf morphology-texture to generate unified screening features. After data fusion, support vector regression + Laplace kernel is used for non-linear modeling, and the Laplace kernel function is used for non-linear mapping to map the fused features to the screening space, expressed as: Among them, X final represents the final unified screening feature; α i represents the weight coefficient of the SVR model; γ represents the Laplace kernel parameter, which controls the smoothness of the feature mapping; ||X fusion -X i ||1 represents the L1 norm distance metric, which is used to measure the similarity between data points; N represents the number of training data points; X i represents the fused feature vector of the i-th training sample.

2. The intelligent tea screening method according to claim 1, characterized in that: The calculation of the screening feature similarity of different tea varieties according to the unified screening features and the construction of a feature transfer matrix include: To measure the distribution differences of different tea varieties in terms of morphology and texture, calculate the screening feature similarity matrix, measure the feature similarity between tea varieties through the Mahalanobis distance, and calculate the probability distribution difference of the screening features using kernel density estimation, expressed as: Among them, D m (i, j) represents the Mahalanobis distance between variety i and variety j; represents the screening feature vector of variety i; represents the screening feature vector of variety j; Σ represents the joint covariance matrix, measuring the correlation of features; S ij represents the screening feature similarity between variety i and variety j; h represents the KDE kernel bandwidth parameter, controlling the smoothness; μ ij represents the feature mean between variety i and variety j; X represents the sample points of the feature distribution; Σ -1 (·) represents the negative matrix; Based on the screening feature similarity S ij , a feature transfer matrix is constructed to transform the screening features of the new tea variety into the screening feature space of the existing variety. Using the weighted feature projection method, the feature transformation relationship from variety i to variety j is constructed and expressed as: Among them, M ij represents the element of the feature migration matrix from variety i to variety j; S ij represents the feature similarity calculated in the first step; N t represents the number of existing tea varieties; α k represents the feature contribution weight of variety k to ensure the maximum information retention during feature transformation; ε represents the feature distribution adjustment coefficient to control the non-linearity of feature projection; represents the Euclidean distance to measure the screening feature difference between variety i and variety k.

3. The intelligent tea sieving method according to claim 2, characterized in that: The transformation of the features of different tea varieties includes: Use the Laplace feature mapping method to transform the screening features of the new tea variety n into the screening space of the existing varieties, expressed as: Among them, represents the screening characteristics of the new tea variety n after conversion; M nj represents the transformation coefficient from variety n to variety j in the feature migration matrix; Δ n j represents the feature divergence measure between variety n and variety j; represents the L1 norm distance, ensuring local similarity during feature conversion.

4. The intelligent tea sieving method according to claim 3, wherein: The establishment of a cross-variety adaptable tea leaf screening model includes: Use a deep neural network to extract screening features, and reduce the feature distribution differences between different varieties through feature adversarial training, so that the new variety can seamlessly adapt to the existing screening model, expressed as: Among them, represents the screening feature output of the new tea variety n, which is used for final discrimination after DNN processing; W1 and W2 represent the trainable weight matrices of the DNN, which are used for non-linear transformation respectively; b1 and b2 represent the bias terms; σ(·) represents the Sigmoid activation function to ensure output normalization; tanh(·) represents the hyperbolic tangent activation function to improve the feature expression ability.

5. The intelligent tea screening method according to claim 4, wherein: The use of a domain adaptation loss function to optimize the screening discrimination boundary and improve the screening accuracy includes: There are domain biases in the feature distributions of different varieties. Introduce a domain adaptation loss function to optimize the screening decision boundary, so that the screening model can maintain a high screening accuracy between different varieties; Combine the maximum mean discrepancy and the adversarial loss to optimize feature alignment and classification stability at the same time, expressed as: Among them, represents the domain adaptation loss function; N represents the number of training data points; and represent the final screening features of variety i and variety j; represents the Euclidean distance metric, which is used to calculate the MMD loss to align the feature distributions of different varieties; λ' represents the loss weight hyperparameter; y i represents the class label of the training sample; W c and b c represent the classifier weights and bias terms; represents the adversarial loss term, ensuring that the features can be used for screening classification; The value range is [0, +∞). The smaller the value, the closer the screening characteristics distributions of different varieties are, and the better the adaptation effect is.

6. The intelligent tea screening method according to claim 5, wherein: The optimization of the screening discrimination boundary includes: Adopt the maximum margin optimization method to ensure the robustness of the screening model between different varieties; Use Lagrangian duality optimization to optimize the discrimination boundary of the screening categories, so that different varieties maintain classification stability during the screening process, expressed as: Among them, represents the discriminant boundary optimization loss; α i and α j represent Lagrange multipliers; y i and y j represent the class labels of the training samples; represents the kernel function, which measures the non - linear similarity between screening features; The value range is from negative infinity to positive infinity. During the optimization process, the smaller the value, the clearer the discrimination boundary and the higher the screening accuracy. The cross-variety screening model optimized by the discrimination boundary is applicable to the intelligent screening tasks of different tea varieties.

7. An intelligent tea screening system using the method described in any one of claims 1-6, characterized in that: A multi-modal perception feature extraction module collects the morphological features of tea leaves, calculates the texture information of tea leaves, fuses multi-modal data through a CNN-SVM hybrid model, establishes a non-linear mapping model of tea leaf morphology-texture, and generates unified screening features; A module for constructing a feature transfer matrix of tea varieties calculates the similarity of screening features of different tea varieties according to the unified screening features, constructs a feature transfer matrix, and transforms the features of different tea varieties; A screening optimization module in a small sample environment establishes a cross-variety adaptive screening tea model, and uses a domain adaptive loss function to optimize the screening discrimination boundary to improve the screening accuracy.

Images