Multidimensional digital evaluation method and system for minnan oolong tea

Abstract

The present application relates to the technical field of tea quality evaluation, and particularly relates to a Minnan oolong tea multi-dimensional digital evaluation method and system. In view of the fact that the existing Minnan oolong tea quality evaluation mainly relies on artificial sensory evaluation, and the problems of strong subjectivity, poor stability, difficulty in realizing standardized and large-scale application, etc., a multi-source sensor collaborative digital representation and comprehensive evaluation system is constructed around key quality factors such as dry tea appearance, leaf appearance, tea soup color, tea soup taste and tea leaf bulk density, combined with machine learning and deep learning models and multi-factor discrimination logic. First, the dry tea appearance and leaf appearance are discriminated, and if the type cannot be determined, the tea soup color, tea soup taste and tea leaf bulk density are further introduced for multi-factor discrimination, so as to realize objective, stable and reproducible digital evaluation of Minnan oolong tea quality, and provide technical support for industrialized production and quality control of tea.

Description

Technical Field

[0001] This invention relates to the field of tea quality evaluation technology, specifically to a multi-dimensional digital evaluation method and system for Minnan Oolong tea. Background Technology

[0002] Oolong tea, as one of my country's important traditional tea categories, holds a significant position in tea production and consumption due to its unique processing characteristics and flavor. The quality grade of Minnan oolong tea is typically determined by a combination of sensory indicators, including the appearance of the dry tea leaves, the color of the brewed tea liquor, the taste, and the infused leaves. In actual production, even with the same raw material source and initial grade, different processing techniques and parameters can lead to variations in the quality of the finished tea. Therefore, after production, tea leaves usually require sensory evaluation by tea appraisers to determine their quality grade.

[0003] However, traditional manual sensory evaluation methods rely heavily on the experience of the evaluators, and the evaluation results are easily affected by individual differences and environmental factors. They are highly subjective and difficult to standardize and scale up, making it difficult to meet the needs of the modern tea industry for consistent quality and online evaluation.

[0004] With the development of multi-source sensing and intelligent analysis technologies, non-destructive testing methods such as machine vision, near-infrared spectroscopy, and visible / near-infrared spectroscopy are gradually being applied to the field of tea quality evaluation. These methods can objectively characterize the appearance, internal quality, and tea liquor properties of tea without requiring complex sample pretreatment. Based on this, a digital multi-dimensional evaluation method for Minnan Oolong tea using the X factor is constructed to achieve the fusion analysis and comprehensive judgment of multi-factor information. This is of great significance for improving the objectivity, stability, and reproducibility of tea quality evaluation. Summary of the Invention

[0005] To address the issues that existing quality evaluation methods for Southern Fujian Oolong tea primarily rely on manual sensory assessment, which suffers from strong subjectivity, poor stability, and difficulty in standardization and large-scale application, this invention proposes a multi-dimensional digital evaluation method and system for Southern Fujian Oolong tea. This method utilizes multi-source sensing techniques to digitally characterize the dry tea's external quality, leaf residue's external quality, tea liquor characteristics, and physical features. By combining machine learning and deep learning models with multi-factor discrimination logic, it achieves an objective, stable, and reproducible evaluation of Oolong tea's quality grade.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] The multi-dimensional digital evaluation method for Minnan Oolong tea includes the following steps:

[0008] S1. Collect images of the dry tea surface quality of the tea sample. Based on the dry tea surface quality digital discrimination model corresponding to the dry tea surface quality, perform dry tea surface quality factor discrimination. If the quality type is clear in this step, the process will terminate. If the quality type is in doubt in this step, proceed to the next step.

[0009] S2. Collect images of the tea leaves after brewing. Based on the digital discrimination model of the tea leaves, identify the tea leaf exfoliation factors. If the quality type is clear in this step, the process will terminate. If the quality type is in doubt in this step, proceed to the next step.

[0010] S3. The external quality of the dry tea in step S1 and the external quality of the leaf bottom in step S2 are comprehensively judged by the first fuzzy membership function. If the comprehensive membership degree reaches the discrimination threshold at a certain level, the quality type is clear and the process ends. If the comprehensive membership degree does not reach the discrimination threshold at any level, the quality type is unclear and the next step is executed.

[0011] S4. Based on the external quality of dry tea in step S1 and the external quality of tea leaves in step S2, the color of tea soup, the taste of tea soup and the bulk density of tea leaves are introduced and digitally represented and graded respectively. The relationship between each evaluation factor and different quality types is described by the second fuzzy membership function. The multi-factor information is fused and calculated. Finally, the grade is determined based on the largest comprehensive membership degree R value of the grade.

[0012] Optionally, the digital discrimination model for the external quality of dried tea includes a multi-classification sub-model and a binary classification sub-model for the external quality of dried tea. Step S1 specifically comprises:

[0013] The classification probability distribution of each dry tea surface quality image under different quality grades was obtained by using a multi-class sub-model of dry tea surface quality.

[0014] Statistical analysis of the multiple classification results was used to determine the discrimination results of the tea samples on the external quality factors of dry tea, and the number of samples judged to each grade was counted.

[0015] If a certain level of discrimination rate meets the grading threshold, the quality type is determined and the process is terminated.

[0016] If no grade discrimination rate meets the grading threshold, but the sum of the discrimination rates of two grades meets the binary classification threshold, then the binary sub-model of the dry tea extrinsic quality corresponding to the two grades is called for further discrimination: if the discrimination rate of a grade meets the grading threshold, the quality type is clear and the process terminates; otherwise, the quality type remains questionable.

[0017] Optionally, the grading threshold in the multiple classification and binary classification of dry tea external quality is 85%, and the binary classification threshold in the binary classification of dry tea external quality is 90%.

[0018] Optionally, the digital discrimination model for leaf ectoplasm includes a multi-classification sub-model and a binary classification sub-model for leaf ectoplasm, and step S2 specifically involves:

[0019] The classification probability distribution of each leaf basal ectoplasm image under different quality grades was obtained using a multi-class sub-model of leaf basal ectoplasm.

[0020] By statistically analyzing the results of multiple classification, the discrimination results of tea samples on the extrinsic factors of leaf base were determined, and the number of samples judged to each grade was counted.

[0021] If a certain level of discrimination rate meets the grading threshold, the quality type is determined and the process is terminated.

[0022] If no level's discrimination rate meets the grading threshold, but the sum of the discrimination rates of two levels meets the binary classification call threshold, then the leaf basal ectoma binary classification sub-model corresponding to those two levels is called for further discrimination: if a level's discrimination rate meets the grading threshold, the quality type is clear and the process terminates; otherwise, the quality type remains uncertain.

[0023] Optionally, the grading threshold in leaf effusion multiple classification and leaf effusion binary classification is 85%, and the binary classification call threshold in leaf effusion binary classification is 90%.

[0024] Optionally, after obtaining the results of the dry tea quality assessment and the leaf residue quality assessment, the two evaluation factors are comprehensively judged using the first fuzzy membership function. The first fuzzy membership function describes the relationship between each evaluation factor and different quality types. The information of the two evaluation factors is fused and calculated using the following formula:

[0025] In the formula, Rj is the comprehensive membership degree of the tea sample to grade j; n=2, which is the number of evaluation factors currently involved in the judgment; ω i μ is the combined weight of the i-th factor; ij (x i ) represents the membership degree of the i-th factor to level j; xi represents the measured value of the i-th factor;

[0026] In step S3, if the overall membership degree of several tea extrins and leaf residues reaches the discrimination threshold, the quality type is clear and the process terminates; if the overall membership degree of several tea extrins and leaf residues does not reach the discrimination threshold, the quality type is unclear.

[0027] Optionally, the discrimination threshold in step S3 is 80%.

[0028] Optionally, in step S4, each evaluation factor is assigned a corresponding weight based on its contribution to the quality evaluation. The relationship between each evaluation factor and different quality types is described by a second fuzzy membership function. The multi-factor information is then fused and calculated using the following formula:

[0029] In the formula, Rj is the comprehensive membership degree of the tea sample to the grade; n is the number of evaluation factors, and the number of n is 5; ω i μ is the combined weight of the i-th factor; ij (x i ) represents the membership degree of the i-th factor to level j; xi represents the measured value of the i-th factor.

[0030] This invention also provides a multi-dimensional digital evaluation system for Minnan Oolong tea that performs the above-described evaluation method, comprising:

[0031] Image acquisition unit is used to acquire information about the appearance of dry tea leaves and tea leaves;

[0032] The spectral acquisition unit is used to acquire visible / near-infrared spectral information of dry tea and tea liquor;

[0033] The physical feature acquisition unit is used to acquire physical parameters such as the bulk density of tea leaves;

[0034] The data processing and model calculation unit is used to realize multi-factor feature extraction, model inference and result output.

[0035] Optionally, it also includes a digital twin platform for mapping the physical testing process, evaluation model, and judgment results to the digital space to achieve data synchronization, status updates, and online analysis.

[0036] Compared with the prior art, the present invention has the following beneficial effects: This invention is based on the formation mechanism of oolong tea quality. It constructs a multi-source sensor-based, collaborative digital representation and comprehensive evaluation system focusing on key quality factors such as the external quality of dry tea leaves, the appearance of the infused leaves, the internal quality of dry tea leaves, the color and taste of the tea liquor, and the bulk density of the tea leaves. By constructing an X-factor digital multi-dimensional evaluation method, it achieves a comprehensive and objective representation of oolong tea quality information. Through the integrated application of deep learning and machine learning models, it improves the accuracy and stability of tea quality evaluation. Finally, through multi-stage, multi-factor digital discrimination logic, it enables the step-by-step screening and comprehensive judgment of oolong teas of different quality grades.

[0037] The method utilizes an industrial line scan camera and feature extraction algorithm to digitally detect the external quality information of dry tea leaves and leaf bases; it uses a 400-1050 nm transmission visible / near-infrared spectral system to digitally detect the color of the tea liquor through color space conversion and spectral analysis models; it collects spectral information of dry tea leaves based on a 900-1700 nm near-infrared spectrum and combines pattern recognition and physicochemical index modeling to digitally evaluate the flavor of the tea liquor; and it obtains dynamic and static bulk density information of oolong tea through a bulk density detection mechanism that combines vibration and non-vibration modes.

[0038] Furthermore, this invention constructs a multi-factor digital comprehensive evaluation model integrating deep learning and machine learning. It employs a ConvNeXt network for extrinsic image discrimination and combines models such as D-PLS, KNN, and Si-CARSPLS to achieve refined discrimination of tea liquor color and taste. Through a multi-stage decision-making logic of "probability distribution + multi-classification model + discrimination result analysis," it performs step-by-step screening and comprehensive judgment of oolong tea of ​​different quality levels. Based on this, a multi-factor digital evaluation integration system and digital twin integrating biomimetic variables such as visual and taste are constructed to achieve collaborative perception, online monitoring, and standardized quality evaluation of multi-source sensor information, environmental information, and equipment status.

[0039] This invention enables objective, stable, and reproducible digital evaluation of oolong tea quality, providing technical support for industrialized tea production and quality control. Detailed Implementation

[0040] This embodiment takes samples of oolong tea from different batches in the Minnan production area as the research object. The samples are from no less than 13 production batches. Under the condition of similar raw material grade, they are formed by different processing technology and have distinguishable quality differences of Minnan oolong tea.

[0041] (1) Data collection and labeling of dry tea surface quality and leaf bottom surface quality.

[0042] For the oolong tea samples, images of the dry tea leaves and the infused leaves were collected. A unified acquisition process and annotation standards were established during image acquisition, and each batch of samples was numbered and managed to form a standardized image dataset for subsequent model training and validation.

[0043] The dry tea images are mainly used to characterize the appearance features of tea leaves, such as leaf shape, color distribution, and uniformity. The leaf bottom images are used to characterize the degree of leaf unfolding, color consistency, and processing status. Images of tea leaves after steeping in tea liquor are captured using a camera and taken after the leaves have cooled to a suitable temperature. Through this process, a sample library of 7500 dry tea and leaf bottom images covering 13 batches and 5 quality grades is constructed.

[0044] (2) Construction of a deep learning-based digital discrimination model for the external quality of dried tea.

[0045] In the construction of the dry tea surface quality discrimination model, ConvNeXt is used as the core deep learning model for classification training of dry tea surface quality images. The ConvNeXt model integrates the design concepts of convolutional neural networks and Transformers, which can extract color, texture, and structural features from images while ensuring computational efficiency. Compared with traditional convolutional neural network structures, the ConvNeXt network maintains the advantages of convolutional structures while improving the ability to represent subtle differences in the texture and color of tea leaves, making it more suitable for the task of discriminating fine-grained differences in the surface quality of Minnan Oolong tea. A multi-classification sub-model and a binary classification sub-model for dry tea surface quality features are established. Both sub-models are based on the ConvNeXt network. The multi-classification sub-model outputs a classification level label based on the classified type. For two similar grades of Oolong tea surface quality, a binary classification function is used to distinguish them, and the classification level label between the two grades is output as the discrimination result of the sample in the dry tea surface quality dimension.

[0046] The digital characterization method for the external quality of dried tea includes: acquiring appearance images of dried tea samples under controlled illumination; preprocessing the images and extracting regions of interest; extracting multidimensional visual feature parameters reflecting the color, texture, shape and purity of the tea based on the regions of interest; and realizing a digital comprehensive evaluation of the external quality of Minnan Oolong tea through the analysis of the external quality characteristics of dried tea.

[0047] A digital discrimination model for the external quality of dried tea was established by training on labeled samples, enabling automatic identification of the external quality of different batches of oolong tea samples. In this embodiment, the model achieved an accuracy rate of 95.70% in identifying the external quality of dried tea.

[0048] (3) Construction of a deep learning-based digital discrimination model for leaf ectophysis.

[0049] In the process of constructing the digital discrimination model for leaf ectoplasm, the ConvNeXt model is used as the core deep learning model to classify and train leaf ectoplasm images. The acquisition of leaf ectoplasm mainly focuses on the color, brightness, and consistency of different leaves from the same batch of leaf ectoplasm.

[0050] In the process of model building, the leaf effluent quality also adopts a combination of multi-class sub-model and binary sub-model. The binary sub-model is also designed for similar tea grades, and multiple corresponding binary sub-models are established.

[0051] The digital characterization method for the external quality characteristics of tea leaves includes: acquiring images of the appearance of the tea leaves after brewing using an industrial line scan camera under controlled illumination; preprocessing the images and extracting regions of interest; extracting multi-dimensional visual feature parameters reflecting the morphology, color uniformity, and unfolding state of the tea leaves after brewing based on the regions of interest; and realizing the digital characterization of the external quality of Minnan Oolong tea leaves by analyzing the external quality characteristics of the tea leaves.

[0052] A digital discrimination model for leaf residue quality was established by training on labeled samples to automatically identify the quality of leaf residue from different batches of oolong tea samples. In this embodiment, the model achieved an accuracy rate of 95.56% in identifying leaf residue quality.

[0053] The discrimination of leaf residue images and dry tea residue images are parallel processes. Both employ a deep learning network structure centered on ConvNeXt for feature learning and classification training. While the model architecture is the same, the extracted features differ: dry tea residue images primarily extract appearance information such as tea leaf texture, particle size, and strip morphology; leaf residue images primarily extract visual features such as the color, size, and uniformity of the brewed leaves. Dry tea residue and leaf residue residue images each form independent discrimination model systems.

[0054] (4) Visible / near-infrared spectral acquisition and digital discrimination modeling of tea soup color.

[0055] A digital characterization method for tea liquor color is developed by acquiring spectral information of tea liquor in the 400–1050 nm wavelength range through transmission visible / near-infrared spectroscopy. The color is processed based on the CIE 1931 XYZ colorimetric system recommended by the International Commission on Illumination (CIE) and converted to [the appropriate colorimetric system] under selected standard illuminant conditions. A uniform color space is used to obtain the brightness and color parameters of tea infusion. The core method can be summarized as "based on spectral reflectance / transmittance → CIE XYZ tristimulus values ​​→..." "Color space conversion" establishes the relationship between the color space values ​​of tea soup and the grade of tea through an MLR model, establishes the distribution range of tea soup color parameters for different quality grades of Minnan Oolong tea, and thus obtains tea grade information.

[0056] (5) Construction of a digital discrimination model for tea flavor.

[0057] Visible / near-infrared spectral data of tea infusions in the 900-1700 nm wavelength range were collected from various batches of oolong tea samples under uniform brewing conditions. The collected spectral data underwent preprocessing, including SG smoothing and MSC correction, and key spectral variables were extracted using a variable screening method. Based on the processed spectral data, the Si-CARS-PLS modeling method was used to model and analyze flavor-related indicators such as tea polyphenols, amino acids, water extracts, and caffeine. The alcohol content and body of the tea were calculated and transformed accordingly to characterize the differences in tea infusion flavor.

[0058]

[0059]

[0060] Among them, the tea soup mellowness, the tea soup richness, the tea soup amino acid content, the tea soup polyphenol content, and the tea soup water extract content are all measured.

[0061] (6) Tea bulk density.

[0062] The method for detecting the bulk density of oolong tea includes: obtaining the static bulk density parameters of tea samples under non-vibration conditions using a bulk density detection unit, and obtaining the dynamic bulk density change characteristics under controlled vibration conditions; wherein, the static bulk density is used to characterize the degree of compaction of tea in a natural stacking state, and the dynamic bulk density is used to characterize the compaction characteristics of tea under external force disturbance; the bulk density information of oolong tea is closely related to the degree of rolling, compaction, and roasting process of tea. The structural differences of tea formed under different processing conditions exhibit different bulk density change characteristics during vibration compaction. Therefore, the bulk density parameter can serve as an effective physical characterization index reflecting the processing characteristics of Minnan oolong tea.

[0063] The static and dynamic bulk density parameters are used as important auxiliary factors reflecting the processing state and structural characteristics of Minnan Oolong tea, and are used in the multi-factor comprehensive evaluation of the quality of Minnan Oolong tea.

[0064] In the digital evaluation of the quality of Minnan Oolong tea, dry tea bulk density is introduced as an auxiliary discriminant factor to characterize the structural differences in the volume and mass dimensions of tea leaves. The bulk density characteristics of dry tea samples under non-vibration and vibration conditions are obtained through a bulk density detection unit, yielding static and dynamic bulk density parameters.

[0065] This method integrates multi-dimensional digital information such as the external quality of dry tea leaves, the external quality of tea leaves, the color and taste of tea soup, and the bulk density of tea leaves. It assigns corresponding weights to each evaluation factor based on its contribution to the quality evaluation, describes the relationship between each evaluation factor and different quality grades through fuzzy membership functions, and integrates and calculates multi-factor information to finally output the comprehensive quality evaluation result of Minnan Oolong tea, realizing the transformation from single-factor discrimination to multi-factor collaborative decision-making.

[0066] The weights of each evaluation factor were determined using a combined weighting method of the Analytic Hierarchy Process (AHP) and the entropy weighting method. This approach integrates subjective and objective weighting, taking into account both professional experience in quality evaluation and the inherent characteristics of the data. The weight calculation formula is as follows: (1)

[0067] In the formula, ω represents the combined weight of the evaluation factors, and α is the adjustment coefficient for the weights in the analytic hierarchy process (0 < α < 1). AHP The subjective weights determined by the analytic hierarchy process, ω entropy The objective weights are determined by the entropy weight method. This formula can effectively calculate the contribution of each evaluation factor to the quality. An example of the combined weight allocation for dry tea surface quality, leaf residue surface quality, tea liquor color, tea liquor taste, and tea bulk density is 0.34, 0.24, 0.20, 0.16, and 0.06, respectively. The fuzzy relationship between each evaluation factor and different quality types is described by a triangular membership function, with the formula: (2)

[0068] In the formula, μ A (x) represents the membership degree of factor characteristic value x to quality type A, and a and b are the interval boundary values ​​of factor characteristic values ​​corresponding to quality type A. This function realizes the fuzzy calculation of information between each evaluation factor and different quality levels.

[0069] Based on the above factor weights and single-factor membership calculation results, the comprehensive membership degree of the tea sample to each quality grade is further integrated to achieve a fuzzy evaluation of multi-factor information. The comprehensive calculation formula is as follows: (3)

[0070] In this formula, R j ω represents the overall membership degree of the tea sample to the grade; n is the number of evaluation factors (n=5 in this study); i μ is the combined weight of the i-th factor; ij (x i ) represents the membership degree of the i-th factor to level j; x iLet be the measured value of the i-th factor. This formula organically integrates the single-level membership results of multiple factors through weighted summation, realizing the transformation from independent single-factor discrimination to multi-factor collaborative comprehensive evaluation, and ultimately completing the quantitative determination of the quality grade of Minnan Oolong tea.

[0071] This invention also provides a multi-dimensional digital evaluation system for Minnan Oolong tea, comprising: an image acquisition unit for acquiring information on the appearance of dry tea and tea leaves; a spectral acquisition unit for acquiring visible / near-infrared spectral information of dry tea and tea liquor; a physical feature acquisition unit for acquiring physical parameters such as tea bulk density; a data processing and model calculation unit for realizing multi-factor feature extraction, model inference, and result output; and a digital twin platform for mapping the physical detection process, evaluation model, and judgment results to a digital space to achieve data synchronization, status updates, and online analysis.

[0072] Example 1

[0073] S1. Preliminary judgment based on the external quality of dry tea leaves.

[0074] First, the external quality of dry tea is used as the first-level discriminant factor for quality evaluation. Based on the appearance characteristics of dry tea, such as shape, color, uniformity and cleanliness, an external quality discrimination strategy of "probability distribution representation + multi-classification modeling + hierarchical decision analysis" is constructed to realize end-to-end digital mapping from the visual characteristics of dry tea external quality to quality grade.

[0075] Specifically, 10 images of the external properties of dried tea leaves are used as a set of input samples, let X = {x1, x2, x3, x4, x...} 5…… x 10 Let C = {XNX200, XNX500, SZZ1000, SZZ2000, AXLC} be the sample set, and let C = {XNX200, XNX500, SZZ1000, SZZ2000, AXLC} be the 5 quality grade labels (hereinafter referred to as c). j (Representing 5 quality grade labels). The x-value of each image is obtained using the multi-classification sub-model of dry tea external quality in the digital discrimination model for dry tea external quality. k Classification probability distribution under different quality levels: P k =[p k1 ,p k2 ,p k3 ,p k4 ,p k5 ]. Where p kj =P(Level=c) j |x k ) represents the image x k Belongs to level C j The probability satisfies Statistical analysis of the multiclassification results was used to determine the discrimination results of the samples in the ectophysiological dimension. The specific discrimination method is as follows: (4)

[0076] Count the number of images classified into each level out of 10 images. If a certain level c exists... j If the discrimination rate meets the grading threshold (85%), then the sample is determined to be grade c. j Otherwise, the sample is marked as a "suspected sample" and proceeds to further evaluation. (Level c) j The discrimination rate refers to the rate at which the model classifies a class as belonging to level c. j The percentage of the total number of dry tea exogenous images in the sample set.

[0077] If none of the classification results for the 10 images meet the grading threshold, meaning the votes for each grade are dispersed (i.e., "suspected samples"), it indicates grade confusion in the five-category classification of the dry tea's external quality. In this case, statistical analysis of the confusion is performed to extract the two grades with the highest probability distribution. and Furthermore, the sum of the discrimination rates of the two levels meets the binary classification call threshold (90%), and the call is pre-selected for that level. A specially trained binary sub-model of dry tea exogenous matter is used for fine discrimination.

[0078] The binary classification sub-model for dry tea extrinsic quality uses the same ConvNeXt deep learning network structure as the multi-classification sub-model. It performs binary classification on two similar grades of oolong tea extrinsic quality, outputting the sample's classification status. and The classification level label between the samples is used as the final discrimination result of the sample in the dimension of dry tea external quality. After the dry tea external quality binary classification model outputs the discrimination level label of dry tea external quality, it enters the leaf bottom image discrimination process.

[0079] S2. Digital discrimination based on leaf basal ectoplasmic images.

[0080] The specific process for judging the external quality of tea leaves is exactly the same as that for judging the external quality of dried tea leaves: 10 images of tea leaves after brewing are used as a set of input samples. The digital discrimination model of tea leaf external quality (ConvNeXt, with 5 output categories) is used to perform five-class classification probability output, and the number of samples in the 10 images is counted. If the discrimination rate of a certain level meets the classification threshold (85%), the label of that level is output; if the votes for each level are scattered, two dominant levels are extracted, and a special binary classification model of tea leaf external quality (ConvNeXt, with 2 output categories) trained for that level is called to perform fine discrimination, and the label of the sample between the two levels is output. The discrimination formula is the same as that of equation (4).

[0081] It should be noted that the discrimination of the external quality of dried tea leaves and the external quality of tea leaves both adopt a hierarchical discrimination strategy of "five-category coarse classification + two-category fine classification". The five-category model is used for preliminary discrimination of five quality grades; the two-category model is used to finely distinguish between two similar grades when there is confusion in the five-category results, so as to output a clear grade classification. In this embodiment, both the five-category model and the two-category model use ConvNeXt as the basic architecture, with the only difference being that the number of output categories in the two-category model is adjusted from 5 to 2. In practical applications, the two-category model is not limited to the ConvNeXt structure; it can also use deep learning network structures such as ResNet and EfficientNet, or traditional machine learning methods such as SVM and MLP. Simply adjusting the number of output categories to 2 is sufficient to achieve fine discrimination between two similar grades. The hierarchical strategy of "five-category coarse classification + two-category fine classification" can effectively improve the accuracy of external quality discrimination while reducing the modeling difficulty of directly performing five-category fine discrimination.

[0082] S3. A secondary discrimination method based on "dry tea surface quality + leaf residue surface quality".

[0083] The results of the dry tea external quality discrimination and the leaf bottom external quality discrimination have been obtained. The two are used as two independent evaluation factors and are substituted into the fuzzy membership function shown in equation (3) for comprehensive judgment. At this time, the number of evaluation factors n=2, and the weights of dry tea external quality and leaf bottom external quality are both 0.5.

[0084] If the overall membership degree reaches the discrimination threshold (80%), the quality type is clear and the discrimination is terminated; if the overall membership degree does not reach the discrimination threshold (80%), the quality type is unclear, and the characteristics of tea soup color, tea soup taste and tea bulk density are introduced to carry out multi-factor comprehensive discrimination. At this time, n=5, and the weight of each evaluation factor is redistributed according to formula (1).

[0085] S4. Multi-factor comprehensive judgment is made by introducing tea soup color, tea soup taste and tea leaf bulk density.

[0086] When the information on the external quality of dry tea and the external quality of tea leaves is insufficient to determine the quality grade, the color of tea soup, the taste of tea soup, and the bulk density of tea leaves are introduced for multi-factor comprehensive judgment.

[0087] Regarding the color of tea soup, digital characterization parameters of tea soup in terms of brightness, hue, and color saturation are obtained through visible / near-infrared spectral analysis and color space conversion. The quality type of tea sample in terms of tea soup color factors is then determined using a digital discrimination model of tea soup color.

[0088] Regarding the taste of tea soup, we introduce the taste of tea soup as a core factor at a higher level of discrimination. Based on the digital discrimination model of tea soup taste, we digitally predict the sensory attributes of tea soup such as mellowness, richness and astringency, and output the quality type of tea sample in terms of tea soup taste factors.

[0089] The taste information reflects the comprehensive characteristics of soluble endogenous components in the tea infusion and is highly correlated with the tea quality grade. By comparing and analyzing the taste prediction results with the aforementioned external quality and infusion color discrimination results, the quality grade determination of the samples can be further verified or corrected, thereby improving the reliability of the overall discrimination results.

[0090] Regarding tea bulk density, this invention innovatively introduces tea bulk density information as an auxiliary factor for evaluating the quality of Minnan Oolong tea. Through a bulk density detection unit, the dynamic and static bulk density characteristics of tea leaves are obtained under vibrating and non-vibrating conditions, respectively, to characterize the particle structure, compactness, and packing properties of the tea leaves.

[0091] Using tea bulk density as supplementary information, along with factors such as appearance, liquor color, and taste, in quality assessment is particularly effective in distinguishing between samples with similar appearance and taste. In most cases, high-quality tea within the same grade has a greater bulk density than low-quality tea. In the comprehensive evaluation process, dry tea bulk density is not used as the sole criterion but rather as an auxiliary factor to correct the evaluation results of appearance, internal quality, and tea liquor, thereby improving the stability and consistency of quality assessment across different batches of Minnan Oolong tea samples.

[0092] If the quality type remains unclear after step S3, a feature set of Minnan Oolong tea X factors is constructed by integrating the characteristics of dry tea leaves, infused leaf appearance, tea liquor color, tea liquor taste, and tea bulk density. A second fuzzy membership function describes the relationship between each evaluation factor and different quality types. Multi-factor information is fused and calculated, and the grade is ultimately determined based on the largest R-value of the corresponding discrimination level. This comprehensive discrimination logic can improve the discrimination accuracy by 2-5% and avoid random errors during sample collection.

[0093] In this embodiment, the quality type determination results for each step are as follows:

[0094] Table 1: Results of the five-category classification of the external quality of dried tea in step S1

[0095] As shown in Table 1, samples ①, ②, and ⑤ reached the grading threshold of 85% in the multiple classification of dry tea appearance, indicating a clear quality type. Samples ③ and ④ did not reach the threshold of 85%, but the sum of the discrimination rates of the two grades, XNX500 and SZZ1000, exceeded 90%. Therefore, a two-class classification of dry tea appearance was used for discrimination, with the two quality grades being XNX500 and SZZ1000.

[0096] Table 2: Results of the binary classification of the external quality of dried tea in step S1

[0098] As shown in Table 2, sample ③ reached the grading threshold of 85% after the binary classification of dry tea surface quality, while sample ④ did not reach the 85% grading threshold. Therefore, the surface quality of the tea leaves was introduced for the classification in step S2.

[0099] It should be noted that the discrimination of samples ①, ②, and ⑤ can be terminated after the multiple classification of the external quality of dry tea, and sample ③ can be terminated after the binary classification of the external quality of dry tea. The following steps can be skipped for these four samples, and the following results are only used as auxiliary verification.

[0100] Table 3: Results of the five-category classification of leaf basal exfoliants in step S2

[0101] As shown in Table 3, samples ①, ②, and ⑤ reached the grading threshold of 85% in the leaf apocryphae five-class classification, while samples ③ and ④ did not reach the threshold of 85%. However, the sum of the discrimination rates of the two grades XNX500 and SZZ1000 in the two samples exceeded 90%, so leaf apocryphae binary classification was used for discrimination.

[0102] Table 4: Results of binary classification of leaf basal excipients in step S2

[0103] As shown in Table 4, sample ③ reached the grading threshold of 85% after the leaf effusion binary classification, while sample ④ did not reach the 85% grading threshold after the leaf effusion binary classification.

[0104] At this point, the external quality of the dry tea and the external quality of the infused leaves have been determined. Treating the external quality of the dry tea and the external quality of the infused leaves as two independent evaluation factors, and substituting them into the first fuzzy membership function for comprehensive judgment, calculations show that the comprehensive membership degree of sample ④ did not reach the 80% judgment threshold, and the quality type remains unclear. Therefore, tea liquor color, tea liquor taste, and tea leaf bulk density are introduced for multi-factor comprehensive judgment in step S4.

[0105] Table 5 shows the results of tea liquor color determination in step S4.

[0106] In the existing model, a ≈ 3.2 ~ 3.8, b ≈ 8 ~ 10, the correspondence between the predicted tea soup color and the quality grade is as follows: predicted value less than -0.5 corresponds to AXLC, -0.5~0 corresponds to XNX200, 0~0.5 corresponds to XNX500, 0.5~1.0 corresponds to SZZ1000, and greater than 1.0 corresponds to SZZ2000.

[0107] Table 6 shows the results of the tea flavor discrimination in step S4.

[0108] The values ​​of each indicator are the average values ​​of multiple measurements of samples of the same grade. The overall taste judgment is derived based on the matching degree between the measured values ​​of each indicator and the grade standard library.

[0109] Table 7: Results of tea bulk density determination in step S4

[0110] The static bulk density was measured in a 500mL container. The reference ranges for bulk density of each grade are as follows: AXLC 153~160, XNX200 121~124, XNX500 146~150, SZZ1000 136~139, SZZ2000 170~173.

[0111] The fuzzy membership function was used to fuse and calculate the multi-factor information. The conclusion is that, considering the discrimination results of five dimensions, namely the external quality of dry tea, leaf residue, tea soup color, tea soup taste, and dry tea bulk density, the following samples were identified: ① Anxi Old Bush (AXLC), ② Small Strong Aroma 200 (XNX200), ③ Small Strong Aroma 500 (XNX500), ④ Sai Zhenzhu 1000 (SZZ1000), and ⑤ Sai Zhenzhu 2000 (SZZ2000). Based on the manual labeling, the final result was correctly identified.

[0112] Evaluation results show that the method of the present invention can effectively distinguish between different batches of Minnan Oolong tea samples, and can automatically classify and identify new batches of tea samples that were not included in the modeling. The classification results have good stability and consistency.

[0113] The foregoing has shown and described the basic principles, main features and advantages of this invention. Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are only illustrative of the principles of this invention. Various changes and modifications can be made to this invention without departing from the spirit and scope of this invention. All such changes and modifications fall within the scope of this invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A multi-dimensional digital evaluation method for Minnan Oolong tea, characterized by: Includes the following steps: S1. Collect images of the dry tea surface quality of the tea sample. Based on the dry tea surface quality digital discrimination model corresponding to the dry tea surface quality, perform dry tea surface quality factor discrimination. If the quality type is clear in this step, the process will terminate. If the quality type is in doubt in this step, proceed to the next step. S2. Collect images of the tea leaves after brewing. Based on the digital discrimination model of the tea leaves, identify the tea leaf exfoliation factors. If the quality type is clear in this step, the process will terminate. If the quality type is in doubt in this step, proceed to the next step. S3. The external quality of the dry tea in step S1 and the external quality of the leaf bottom in step S2 are comprehensively judged by the first fuzzy membership function. If the comprehensive membership degree reaches the discrimination threshold at a certain level, the quality type is clear and the process ends. If the comprehensive membership degree does not reach the discrimination threshold at any level, the quality type is unclear and the next step is executed. S4. Based on the external quality of dry tea in step S1 and the external quality of tea leaves in step S2, the color of tea soup, the taste of tea soup and the bulk density of tea leaves are introduced and digitally represented and graded respectively. The relationship between each evaluation factor and different quality types is described by the second fuzzy membership function. The multi-factor information is fused and calculated. Finally, the grade is determined based on the largest comprehensive membership degree R value of the grade.

2. The multi-dimensional digital evaluation method for Minnan Oolong tea according to claim 1, characterized in that: The digital discrimination model for the external quality of dried tea includes a multi-classification sub-model and a binary classification sub-model for the external quality of dried tea. Step S1 specifically involves: The classification probability distribution of each dry tea surface quality image under different quality grades was obtained by using a multi-class sub-model of dry tea surface quality. Statistical analysis of the multiple classification results was used to determine the discrimination results of the tea samples on the external quality factors of dry tea, and the number of samples judged to each grade was counted. If a certain level of discrimination rate meets the grading threshold, the quality type is determined and the process is terminated. If no grade discrimination rate meets the grading threshold, but the sum of the discrimination rates of two grades meets the binary classification threshold, then the binary sub-model of the dry tea extrinsic quality corresponding to the two grades is called for further discrimination: if the discrimination rate of a grade meets the grading threshold, the quality type is clear and the process terminates; otherwise, the quality type remains questionable.

3. The multi-dimensional digital evaluation method for Minnan Oolong tea according to claim 2, characterized in that: The grading threshold in the multiple classification and binary classification of dry tea external quality is 85%, and the binary classification threshold in the binary classification of dry tea external quality is 90%.

4. The multi-dimensional digital evaluation method for Minnan Oolong tea according to claim 1, characterized in that: The digital discrimination model for leaf ectoplasm includes a multi-classification sub-model and a binary classification sub-model for leaf ectoplasm. Step S2 specifically involves: The classification probability distribution of each leaf basal ectoplasm image under different quality grades was obtained using a multi-class sub-model of leaf basal ectoplasm. By statistically analyzing the results of multiple classification, the discrimination results of tea samples on the extrinsic factors of leaf base were determined, and the number of samples judged to each grade was counted. If a certain level of discrimination rate meets the grading threshold, the quality type is determined and the process is terminated. If no level's discrimination rate meets the grading threshold, but the sum of the discrimination rates of two levels meets the binary classification call threshold, then the leaf basal ectoma binary classification sub-model corresponding to those two levels is called for further discrimination: if a level's discrimination rate meets the grading threshold, the quality type is clear and the process terminates; otherwise, the quality type remains uncertain.

5. The multi-dimensional digital evaluation method for Minnan Oolong tea according to claim 4, characterized in that: The grading threshold in leaf effusion multiple classification and leaf effusion binary classification is 85%, and the binary classification call threshold in leaf effusion binary classification is 90%.

6. The multi-dimensional digital evaluation method for Minnan Oolong tea according to claim 1, characterized in that: After obtaining the results of the dry tea quality assessment and the leaf quality assessment, the two evaluation factors are comprehensively judged using the first fuzzy membership function. The first fuzzy membership function describes the relationship between each evaluation factor and different quality types. The information of the two evaluation factors is fused and calculated using the following formula: ； In the formula, R j ω represents the overall membership degree of the tea sample to grade j; n=2, which is the number of evaluation factors currently involved in the judgment; i μ is the combined weight of the i-th factor; ij (x i ) represents the membership degree of the i-th factor to level j; x i Let be the measured value of the i-th factor; In step S3, if the overall membership degree of several tea extrins and leaf residues reaches the discrimination threshold, the quality type is clear and the process terminates; if the overall membership degree of several tea extrins and leaf residues does not reach the discrimination threshold, the quality type is unclear.

7. The multi-dimensional digital evaluation method for Minnan Oolong tea according to claim 6, characterized in that: The discrimination threshold in step S3 is 80%.

8. The multi-dimensional digital evaluation method for Minnan Oolong tea according to claim 1, characterized in that: In step S4, each evaluation factor is assigned a corresponding weight based on its contribution to the quality evaluation. The relationship between each evaluation factor and different quality types is described by the second fuzzy membership function. The multi-factor information is then fused and calculated using the following formula: ； In the formula, R j ω represents the overall membership degree of the tea sample to the grade; n represents the number of evaluation factors, and the number of n factors is 5; i μ is the combined weight of the i-th factor; ij (x i ) represents the membership degree of the i-th factor to level j; x i Let be the measured value of the i-th factor.

9. A multi-dimensional digital evaluation system for Minnan Oolong tea that implements the evaluation method as described in any one of claims 1 to 8, characterized in that, include: Image acquisition unit is used to acquire information about the appearance of dry tea leaves and tea leaves; The spectral acquisition unit is used to acquire visible / near-infrared spectral information of dry tea and tea liquor; The physical feature acquisition unit is used to acquire physical parameters such as the bulk density of tea leaves; The data processing and model calculation unit is used to realize multi-factor feature extraction, model inference and result output.

10. The multi-dimensional digital evaluation system for Minnan Oolong tea according to claim 9, characterized in that: It also includes a digital twin platform, which maps the physical testing process, evaluation model and judgment results to the digital space to achieve data synchronization, status updates and online analysis.

Images