A high-throughput classification and quantitative determination method and system for rice alkali content based on multi-source characteristics

Abstract

The application belongs to the technical field of agricultural product quality detection, and discloses a rice alkali digestion value high-throughput classification and quantitative determination method and system based on multi-source characteristics, which comprises the following steps: acquiring rice sample data, extracting depth characteristics, texture characteristics and morphological measurement characteristics; performing three-level combined optimization on the depth characteristics, texture characteristics and morphological measurement characteristics, and determining the optimal characteristic combination through quantitative evaluation; performing characteristic screening on the optimal characteristic combination based on the constructed alkali digestion value characteristic neural network classification system, and acquiring rice alkali digestion value key characteristics; training multiple regression models based on the rice alkali digestion value key characteristics and verifying the regression models, determining the optimal regression model, constructing a GCR-FS-RF prediction model, and realizing rice alkali digestion value characteristic quantification; and verifying the grade differentiation performance of the GCR-FS-RF prediction model by identifying adjacent alkali digestion value grade transition zones. The application realizes rice alkali digestion value high-throughput classification and quantification by utilizing multi-modal characteristic fusion and deep learning technology.

Description

Technical Field

[0001] This invention belongs to the field of agricultural product quality testing technology, specifically relating to a high-throughput classification and quantitative determination method and system for rice alkali digestibility based on multi-source characteristics. Background Technology

[0002] Alkali spreading value is a key indicator for measuring the gelatinization characteristics of rice starch, directly reflecting the cooking quality and taste of rice, and is an important basis for rice breeding and quality evaluation. Traditional alkali spreading value determination relies on manual visual grading, which has problems such as strong subjectivity, low efficiency, and unclear definition of transition grades. Existing instrumental methods (such as differential scanning calorimetry and rapid viscosity analysis) are relatively objective, but they are complicated to operate, time-consuming, and require destruction of samples, making it difficult to meet the needs of high-throughput breeding screening.

[0003] In recent years, machine vision and deep learning technologies have been applied in the quality inspection of agricultural products, but existing methods have the following limitations: feature extraction is singular, relying mainly on texture or morphological measurement features, making it difficult to comprehensively characterize the alkali spreading value diffusion pattern; classification models have weak generalization ability and low accuracy in identifying transitional samples between adjacent grades; and there is a lack of a complete technical chain from qualitative grading to quantitative measurement, which cannot meet the needs of refined quality evaluation.

[0004] Therefore, developing a method for classifying and quantifying alkali digestion value that integrates multi-source features and balances high throughput and high precision has become an urgent technical problem to be solved. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a high-throughput classification and quantification method and system for rice alkali spreading value based on multi-source features. It achieves high-throughput classification and quantification of rice alkali spreading value (ASV) by utilizing multimodal feature fusion and deep learning technology.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A high-throughput classification and quantification method for rice alkali degradation value based on multi-source characteristics, the method comprising:

[0008] Acquire rice sample data and extract depth features, texture features, and morphometric features;

[0009] A three-level combination optimization was performed on depth features, texture features, and morphometric features, and the optimal feature combination was determined through quantitative evaluation.

[0010] Based on the constructed alkali digestion value feature neural network classification system, the optimal feature combination is selected to obtain the key features of rice alkali digestion value.

[0011] Based on the key characteristics of rice alkali loss, multiple regression models were trained and validated to determine the optimal regression model. A GCR-FS-RF prediction model was then constructed to quantify the characteristics of rice alkali loss.

[0012] The performance of the GCR-FS-RF prediction model in distinguishing between adjacent alkali digestion value levels was verified by identifying the transition zone between them.

[0013] Preferably, the deep feature is a 62-dimensional vector, containing 12 statistical features and the first 50 original feature values;

[0014] Texture features are used to quantify surface heterogeneity;

[0015] Morphometric characteristics include boundary fractal dimension, diffusion coefficient, compactness, and rice-specific morphology index (SCI), totaling nine dimensions.

[0016] Preferably, the method for feature selection of optimal feature combinations based on the constructed alkali digestibility feature neural network classification system includes:

[0017] Twelve statistical features were evaluated, and the top three feature sets were selected based on F1 scores for integration. Then, Pearson correlation analysis was used to optimize the features and select the best-performing feature subset.

[0018] Preferably, the method for training and validating multiple regression models based on key characteristics of rice alkali loss, determining the optimal regression model, and constructing a GCR-FS-RF prediction model to quantify the characteristics of rice alkali loss includes:

[0019] Multimodal features are input into the regression model, including random forest, gradient boosting, linear regression, and ridge regression. The regression model selects and retains key features that meet the preset requirements through feature selection-embedded ASV-NNCM and hyperparameter optimization. Among them, the hyperparameter optimization uses the Optuna framework to find the optimal parameters in the set search space to maximize the R² score of the validation set.

[0020] During training, the samples were divided into training, validation, and test sets in a 7:2:1 ratio. The input features underwent automated cleaning and standardization preprocessing. 10-fold cross-validation was used to evaluate the model's stability. The quantitative evaluation metrics included the coefficient of determination (R²), root mean square error (RMS), and mean absolute error (MAE).

[0021] Preferably, the method for identifying the transition zone between adjacent alkali digestion value levels includes: using three methods in combination: Overlapping Method (OM) determines the transition zone based on the overlap of the range of regression predicted values; Density Intersection Method (DIM) fits the distribution curve of the predicted values ​​of each level through kernel density estimation (KDE), and takes the predicted value corresponding to the intersection of the curves of adjacent levels as the boundary of the transition zone; Statistical Method (SM) determines the transition zone based on the "±2 standard deviation confidence interval"; finally, a structured classification system of "original level + transition zone" is constructed based on the statistical method, expanding the original 5 levels into 9 classification units.

[0022] The present invention also provides a high-throughput classification and quantification system for rice alkali loss value based on multi-source features. The system is used to implement the aforementioned method and includes: a feature extraction module, a combinatorial optimization module, a feature screening module, a model building module, and a verification module.

[0023] The feature extraction module is used to acquire rice sample data and extract depth features, texture features and morphometric features;

[0024] The combined optimization module is used to perform three-level combined optimization of depth features, texture features and morphometric features, and determine the optimal feature combination through quantitative evaluation.

[0025] The feature selection module is used to select the optimal feature combination based on the constructed alkali digestion value feature neural network classification system to obtain the key features of rice alkali digestion value.

[0026] The model building module is used to train and validate multiple regression models based on the key features of rice alkali disappearance, determine the optimal regression model, construct the GCR-FS-RF prediction model, and realize the quantification of rice alkali disappearance features.

[0027] The verification module is used to verify the class differentiation performance of the GCR-FS-RF prediction model by identifying the transition zone between adjacent alkali digestion value classes.

[0028] Preferably, the deep feature is a 62-dimensional vector, containing 12 statistical features and the first 50 original feature values;

[0029] Texture features are used to quantify surface heterogeneity;

[0030] Morphometric characteristics include boundary fractal dimension, diffusion coefficient, compactness, and rice-specific morphology index (SCI), totaling nine dimensions.

[0031] Preferably, the process of feature selection for the optimal feature combination based on the constructed alkali digestibility feature neural network classification system includes:

[0032] Twelve statistical features were evaluated, and the top three feature sets were selected based on F1 scores for integration. Then, Pearson correlation analysis was used to optimize the features and select the best-performing feature subset.

[0033] Preferably, the process of training and validating multiple regression models based on key characteristics of rice alkali removal value, determining the optimal regression model, and constructing a GCR-FS-RF prediction model to quantify the characteristics of rice alkali removal value includes:

[0034] Multimodal features are input into the regression model, including random forest, gradient boosting, linear regression, and ridge regression. The regression model selects and retains key features that meet the preset requirements through feature selection-embedded ASV-NNCM and hyperparameter optimization. Among them, the hyperparameter optimization uses the Optuna framework to find the optimal parameters in the set search space to maximize the R² score of the validation set.

[0035] During training, the samples were divided into training, validation, and test sets in a 7:2:1 ratio. The input features underwent automated cleaning and standardization preprocessing. 10-fold cross-validation was used to evaluate the model's stability. The quantitative evaluation metrics included the coefficient of determination (R²), root mean square error (RMS), and mean absolute error (MAE).

[0036] Preferably, the process of identifying the transition zone between adjacent alkali digestion value levels includes: using three methods for joint identification: Overlapping Method (OM) determines the transition zone based on the overlap of regression predicted values; Density Intersection Method (DIM) fits the distribution curve of predicted values ​​for each level through kernel density estimation (KDE), and takes the predicted value corresponding to the intersection of curves of adjacent levels as the boundary of the transition zone; Statistical Method (SM) determines the transition zone based on the "±2 standard deviation confidence interval"; finally, a structured classification system of "original level + transition zone" is constructed based on the statistical method, expanding the original 5 levels into 9 classification units.

[0037] Compared with existing technologies, the beneficial effects of this invention are as follows: It solves the problems of strong subjectivity, low efficiency, and ambiguous transition level judgment in traditional manual ASV assessment. Through "standardized image acquisition + YOLOv7 preprocessing + multimodal feature fusion + dual-output prediction optimization," it achieves a leap from "manual qualitative" to "automated quantitative," eliminating human interference and preventing cross-level misjudgments, with efficiency adapted to high-throughput requirements. It overcomes the limitations of methods such as Differential Scanning Calorimetry (DSC) and Rapid Visco Analyzer (RVA), which suffer from residual subjectivity, poor reproducibility, and long detection times. It employs alkali digestion value determination, which is fast and low-cost, combined with an optimized model for rapid assessment. Reproducibility is ensured through 10-fold cross-validation and validation with 200 manual samples, meeting the needs of high-throughput breeding and commercial grading. It overcomes the bottlenecks of low accuracy in existing machine vision technology and poor generalization of some deep learning methods by integrating multimodal features (ResNet18 / ConvNeXt deep features, LBP / HOG). The model utilizes texture features and SCI morphometric features, achieving a total grading accuracy of 0.94 (without cross-grade error) and a regression prediction R² of 0.97 (RMSE=0.26) through ASV-NNCM and feature selection. Based on a dataset of 669 varieties and 19316 samples, it exhibits superior generalization. Filling the gap in existing technologies' inability to handle the continuous transition characteristics of ASV, a structured system of "original grade + transition zone" is constructed using three transition zone identification methods. The statistical method (SM) achieves an average confusion rate of 0.19 and a grade retention rate of 0.68, resolving the misjudgment problem of transitional samples. Through feature engineering of "four-layer feature combination + Pearson correlation analysis + dimensionality reduction," core features are selected from a subset of 3176 feature combinations, shortening model training time (e.g., reducing gradient boosting model training time from 26.51 seconds to 18.32 seconds) and reducing memory usage while maintaining accuracy, thus balancing accuracy and practicality.

[0038] This invention utilizes multimodal feature fusion and deep learning techniques to achieve high-throughput classification and quantification of rice alkali spreading value (ASV). Attached Figure Description

[0039] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the architecture of the Alkali Dissipation Value Recognition System (ASVRS) according to an embodiment of the present invention, wherein (a) is a schematic diagram of hardware device network construction and dataset acquisition; (b) is a schematic diagram of multimodal feature extraction; (c) is a schematic diagram of feature engineering (embedded ASV-NNCM); and (d) is a schematic diagram of constructing a classification system through an alkali dissipation value neural network and realizing feature quantization through various regression methods.

[0041] Figure 2 This is a schematic diagram of the Alkali Value Neural Network Classification System (ASV-NNCM) according to an embodiment of the present invention; wherein, (a) is a feature acquisition module, used to extract depth features, texture features and morphometric features from the alkali value image; (b) is a schematic diagram of feature optimization through Pearson correlation analysis (threshold = 0.5) to select the best-performing statistical features; (c) is a core alkali value classification head, composed of a multilayer perceptron (MLP) with hidden layers, used for feature verification and category prediction; (d) is an optimization module, integrating techniques such as mixed precision training, gradient pruning, feature noise enhancement, early stopping, ensemble voting and category weighting to improve the final verification performance.

[0042] Figure 3 This is a schematic diagram illustrating the acquisition of images via a scanner according to an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of mathematical coarse segmentation + YOLOv7 fine segmentation in an embodiment of the present invention;

[0044] Figure 5 This is a schematic diagram of the specific results after quantification in an embodiment of the present invention; wherein, 3-7—alkali digestion value level;

[0045] Figure 6 This is a schematic flowchart of a high-throughput classification and quantitative determination method for rice alkali degradation value based on multi-source characteristics, according to an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] Example 1

[0049] like Figure 6 As shown, this invention provides a high-throughput classification and quantification method for rice alkali degradation value based on multi-source characteristics, the method comprising:

[0050] Acquire rice sample data and extract depth features, texture features, and morphometric features;

[0051] A three-level combination optimization was performed on depth features, texture features, and morphometric features, and the optimal feature combination was determined through quantitative evaluation.

[0052] Based on the constructed alkali digestion value feature neural network classification system, the optimal feature combination is selected to obtain the key features of rice alkali digestion value.

[0053] Based on the key characteristics of rice alkali loss, multiple regression models were trained and validated to determine the optimal regression model. A GCR-FS-RF prediction model was then constructed to quantify the characteristics of rice alkali loss.

[0054] The performance of the GCR-FS-RF prediction model in distinguishing between adjacent alkali digestion value levels was verified by identifying the transition zone between them.

[0055] In this embodiment, the depth feature is a 62-dimensional vector, containing 12 statistical features and the first 50 original feature values;

[0056] Texture features are used to quantify surface heterogeneity;

[0057] Morphometric characteristics include boundary fractal dimension, diffusion coefficient, compactness, and rice-specific morphology index (SCI), totaling nine dimensions.

[0058] In this embodiment, the method for feature selection of the optimal feature combination based on the constructed alkali digestibility feature neural network classification system includes:

[0059] Twelve statistical features were evaluated, and the top three feature sets were selected based on F1 scores for integration. Then, Pearson correlation analysis was used to optimize the features and select the best-performing feature subset.

[0060] The formulas for feature combination and model evaluation are as follows:

[0061] Coefficient of determination ( ): ; =True ASV value, =Predicted value, =True value mean, sample size ;

[0062] Root mean square error (RMSE) ): ;same The symbol definition.

[0063] The formula for feature selection is:

[0064] Pearson correlation coefficient ( ): ; =Eigenvalue, = Feature mean, others are the same , , For variables.

[0065] In this embodiment, the method for training and validating multiple regression models based on key features of rice alkali removal value, determining the optimal regression model, and constructing a GCR-FS-RF prediction model to quantify the features of rice alkali removal value includes:

[0066] Multimodal features are input into the regression model, including random forest, gradient boosting, linear regression, and ridge regression. The regression model selects and retains key features that meet the preset requirements through feature selection-embedded ASV-NNCM and hyperparameter optimization. Among them, the hyperparameter optimization uses the Optuna framework to find the optimal parameters in the set search space to maximize the R² score of the validation set.

[0067] During training, the samples were divided into training, validation, and test sets in a 7:2:1 ratio. The input features underwent automated cleaning and standardization preprocessing. 10-fold cross-validation was used to evaluate the model's stability. The quantitative evaluation metrics included the coefficient of determination (R²), root mean square error (RMS), and mean absolute error (MAE).

[0068] The formula for calculating the statistical properties of the predicted value distribution is as follows:

[0069] mean ( ): ; =Number of samples at a certain level =Predicted value within this level;

[0070] Standard deviation ( ): The same as the symbol definition for the mean;

[0071] scope( Range ): The same as the symbol definition for the mean;

[0072] Interquartile range ( IQR ): ; =25th percentile, =75th percentile.

[0073] The core regression model calculation formula is as follows:

[0074] Random Forest ( RF ): ; (Number of decision trees) =Predicted value for a single tree To score points.

[0075] In this embodiment, the method for identifying the transition zone between adjacent alkali degradation value levels includes: using three methods in combination: Overlap Method (OM) determines the transition zone based on the overlap of regression predicted values; Density Intersection Method (DIM) fits the distribution curves of predicted values ​​for each level through kernel density estimation (KDE), and takes the predicted value corresponding to the intersection of curves of adjacent levels as the boundary of the transition zone; Statistical Method (SM) determines the transition zone based on the "±2 standard deviation confidence interval"; finally, a structured classification system of "original level + transition zone" is constructed based on the statistical method, expanding the original 5 levels into 9 classification units.

[0076] The calculation formula for transition region identification is as follows:

[0077] Overlap Method (OM): ; =No. Maximum / minimum predicted grade values;

[0078] Statistical Method (SM) Confidence Interval: ; , =No. Mean / standard deviation of predicted grades;

[0079] Density overlap method (DIM) kernel density: ; =Kernel function, =Bandwidth parameter.

[0080] Example 2

[0081] The present invention also provides a high-throughput classification and quantification system for rice alkali loss value based on multi-source features. The system is used to implement the method described in Example 1. The system includes: a feature extraction module, a combination optimization module, a feature screening module, a model construction module, and a verification module.

[0082] The feature extraction module is used to acquire rice sample data and extract depth features, texture features, and morphometric features.

[0083] The combined optimization module is used to perform three-level combined optimization of depth features, texture features and morphometric features, and determine the optimal feature combination through quantitative evaluation.

[0084] The feature selection module is used to select the optimal feature combination based on the constructed alkali digestion value feature neural network classification system to obtain the key features of rice alkali digestion value.

[0085] The model building module is used to train and validate various regression models based on the key features of rice alkali loss, determine the optimal regression model, construct the GCR-FS-RF prediction model, and realize the quantification of rice alkali loss features.

[0086] The validation module is used to verify the class differentiation performance of the GCR-FS-RF prediction model by identifying the transition zone between adjacent alkali digestion value classes.

[0087] In this embodiment, the depth feature is a 62-dimensional vector, containing 12 statistical features and the first 50 original feature values;

[0088] Texture features are used to quantify surface heterogeneity;

[0089] Morphometric characteristics include boundary fractal dimension, diffusion coefficient, compactness, and rice-specific morphology index (SCI), totaling nine dimensions.

[0090] In this embodiment, the process of selecting the optimal feature combination based on the constructed alkali digestibility feature neural network classification system includes:

[0091] Twelve statistical features were evaluated, and the top three feature sets were selected based on F1 scores for integration. Then, Pearson correlation analysis was used to optimize the features and select the best-performing feature subset.

[0092] In this embodiment, the process of training and validating multiple regression models based on key features of rice alkali removal value, determining the optimal regression model, and constructing a GCR-FS-RF prediction model to quantify the rice alkali removal value features includes:

[0093] Multimodal features are input into the regression model, including random forest, gradient boosting, linear regression, and ridge regression. The regression model selects and retains key features that meet the preset requirements through feature selection-embedded ASV-NNCM and hyperparameter optimization. Among them, the hyperparameter optimization uses the Optuna framework to find the optimal parameters in the set search space to maximize the R² score of the validation set.

[0094] During training, the samples were divided into training, validation, and test sets in a 7:2:1 ratio. The input features underwent automated cleaning and standardization preprocessing. 10-fold cross-validation was used to evaluate the model's stability. The quantitative evaluation metrics included the coefficient of determination (R²), root mean square error (RMS), and mean absolute error (MAE).

[0095] In this embodiment, the process of identifying the transition zone between adjacent alkali degradation value levels includes: using three methods for joint identification: the overlap method (OM) determines the transition zone based on the overlap of the range of regression predicted values; the density intersection method (DIM) fits the distribution curve of the predicted values ​​of each level through kernel density estimation (KDE), and takes the predicted value corresponding to the intersection of the curves of adjacent levels as the boundary of the transition zone; the statistical method (SM) determines the transition zone based on the "±2 standard deviation confidence interval"; finally, a structured classification system of "original level + transition zone" is constructed based on the statistical method, expanding the original 5 levels into 9 classification units.

[0096] The specific implementation process is as follows:

[0097] The Alkali Spreading Value Recognition System (ASVRS) is an automated framework for identifying alkali spread values ​​(ASV) that integrates multimodal features and optimized deep learning. Figure 1The process consists of four parts, including (a) standardized image acquisition. A customized 49-well reaction tray, PID-controlled incubator, and flatbed scanner were used. The hulled rice was placed in the reaction tray, and 1.7% potassium hydroxide solution was added. After incubation at (30±2)°C for 23 hours, the images were scanned. Image preprocessing included adaptive threshold coarse extraction, YOLOv7 fine segmentation, morphological operation boundary optimization, and generation of standardized 224×224 pixel grayscale image blocks. Specifically, the images were loaded into a specified folder, and then preprocessed by grayscale conversion, Gaussian filtering, Sobel gradient calculation, and binarization to highlight the edges of straight lines. Then, probabilistic Hough transform was used to detect horizontal and vertical straight lines and filter similar lines according to angle (5°) and distance (20 pixels) thresholds. Next, the effective intersection points (screened within the image range) were obtained by solving the linear equations of the two lines. After normalization (divided by the maximum image size) and Z-score standardization of the intersection points, hierarchical clustering (Euclidean distance, complete linkage) was used to cluster them into No_line×No_line. The model calculates cluster centers and sorts them in ascending order by x. Each No_line cluster is then grouped into a single group, and within each group, sorted in ascending order by y. Based on the sorted centers, a rectangular cropping region with an indentation of ex_value pixels is calculated. After checking for boundaries (within the image) and aspect ratio (0.5-2), the valid cropped image blocks are saved to a specified folder. Based on 300 manually annotated images (avoiding borders), the trained YOLOv7 model weights are used. After determining the input image size, the test image is preprocessed (scaled proportionally to the set size, BGR to RGB, pixel normalization, converted to Tensor, and batch dimension added). The model then performs inference, filtering low-confidence prediction boxes using a set confidence threshold, and removing overlapping redundant boxes using NMS thresholding to obtain accurate detection boxes containing target category, confidence level, and coordinates. The detection results are then saved in txt format (labels are not displayed if necessary). Finally, the corresponding target region is cropped from the original image based on the detection box coordinates, completing the YOLOv7 target recognition and cropping process.

[0098] (b) For multimodal feature extraction, the CPU / GPU is automatically selected and the CPU utilization is controlled to not exceed 85%. Custom training models such as ResNet18 / 50 and EfficientNet (to handle key name matching problems) and PyTorch pre-trained models are loaded and converted into feature extractors. After preprocessing the input image (resize to 224x224, convert to Tensor, normalize), depth features (the first 50 values ​​of the original features and statistical features such as mean, standard deviation, and skewness) are extracted. SCI feature extraction: the extractor is initialized and the core feature weights are set (α=0.42, β=0.35, γ=0.23). When batch processing images, they are first preprocessed (converted to grayscale, Gaussian blur, adaptive threshold binarization, morphological operations) to obtain rice grain masks. Then, the core region is obtained through erosion and the boundary pixels are obtained by dilation and mask reduction. The core compactness (circumference ratio 4π × area / Perimeter²), boundary roughness (fractal dimension using box counting), and dispersion (coefficient of variation of distance from boundary pixels to the core center) are weighted and combined to obtain the SCI value. Simultaneously, morphological measurement features such as area and perimeter are extracted. Image texture feature extraction script: Initializes input and output paths and creates an Excel and visualization subfolder; starts CPU monitoring and control to ensure usage does not exceed 85%; after acquiring the input image, converts it to grayscale; extracts LBP (mean, entropy, etc. + the first 10 histogram bins), GLCM (if successful import, calculates contrast and other statistical values; otherwise, uses placeholder values), Gabor (statistical analysis of different frequency angle responses + overall features), HOG (mean, entropy, etc. + the first 20 descriptors), and basic statistics and gradient features.

[0099] (c) The feature engineering process involves a customized ASV neural network classification subsystem (which directly evaluates the classification performance of the dataset on the ASV neural network through F1 scores, and obtains the optimal feature set combination through different combination regressions) evaluating the feature dataset (features are directly concatenated after flattening). Different combinations are then performed according to a 4-level decreasing sequence (1-2-3-4), and based on R... 2 The best-performing feature combination is selected, and then Pearson correlation analysis (threshold ≥ 0.5) is performed to retain strongly correlated features to eliminate redundancy. Finally, the features are merged into a unified feature vector through channel splicing.

[0100] (d) is the integrated validation framework, strictly following NY / T 83-2017, specifying ASV levels as 3-7. The core framework comprises three parts: feature classification validation, feature regression quantization, and the construction of a structured system for the transition region. The specific methods are as follows: Classification validation uses a 3-layer deep neural network with an attention layer (ASV-NNCM core branch). The input is the core features selected through feature engineering, and the output is the classification probability of 5 ASV levels. The 3 hidden layers have 64, 32, and 16 neurons respectively, employing a structure of "linear layer + Batch Normalization + ReLU activation + Dropout (probability 0.3) + residual connection". The output layer outputs the level probability through the Softmax activation function, and the result is determined by the "maximum probability criterion". During training and validation, 19316 samples are divided into training, validation, and test sets in a 7:2:1 ratio. After Z-score standardization, the data is balanced using a class-weighted loss function, employing Adam... The optimizer (initial learning rate 0.001) combines cosine annealing learning rate scheduling (minimum 1e-6), along with label smoothing (factor 0.1), mixed precision training, and an early stopping mechanism (stopping if the validation set loss does not decrease after 25 consecutive epochs). Accuracy, recall, and F1 score are used as evaluation metrics. Regression quantization aims to transform ASV from discrete levels to continuous values ​​ranging from 3.0 to 7.0. Four models were selected: Gradient Boosting (GB, used to capture nonlinear interactions of features), Random Forest (RF, used to reduce single-tree noise through ensemble of multiple decision trees), Ridge Regression (RR, used to suppress multicollinearity through L2 regularization), and Linear Regression (LR, used as the baseline model). The optimal GCR feature combination (Gradient + ConvNeXt + ResNet18) was first selected through comparison using 12 statistical features (5 deep features, 6 texture features, and 1 SCI morphometric feature). Then, features with a Pearson correlation coefficient ≥ 0.5 were selected to retain the 23-dimensional core features. Hyperparameter optimization was performed using Optuna with 150 trials (600-second timeout), adjusting parameters such as the number of decision trees, maximum tree depth, and minimum number of samples for node splits in the RF model. Validation employed 10-fold cross-validation (randomly dividing the training set into 10 parts, 9 for training and 1 for validation, repeated 10 times). The evaluation was conducted using 200 manually measured samples (representative samples from 669 varieties were selected, and the ASV values ​​were manually measured by 3 senior engineers according to NY / T 83-2017 as the benchmark), with R², RMSE, and MAE as the evaluation indicators.To address the ambiguity of transitional state samples in traditional grading, the transition zone structured system employs three identification methods: the Overlap Method (OM), which determines the transition zone based on the overlap of regression predicted values ​​(taking the overlapping interval of predicted values ​​from adjacent levels); the Density Intersection Method (DIM), which fits the distribution curves of predicted values ​​for each level using kernel density estimation (KDE) and takes the predicted value corresponding to the intersection of curves from adjacent levels as the boundary of the transition zone; and the Statistical Method (SM), which determines the transition zone based on a "±2 standard deviation confidence interval" (first calculating the mean and standard deviation of predicted values ​​for each level, and taking the overlapping part of the confidence intervals from adjacent levels). Finally, based on the statistical method, a structured classification system of "original level + transition zone" is constructed, expanding the original 5 levels to 9 classification units (3, 3-4, 4, 4-5, 5, 5-6, 6, 6-7, 7), to retain the original NY / T 83-2017 grading framework and solve the misjudgment problem of "either / or" in traditional qualitative grading.

[0101] The high-throughput classification and quantification system for rice alkali degradation value (ASVRS) described in this invention consists of four core modules (hardware system construction, multivariate feature extraction, feature engineering, qualitative classification and quantitative regression). The experimental process uses the following materials and equipment ( Figure 1 ):

[0102] Hardware equipment: Customized 49-well reaction tray (144×144×52 mm, optical transparency 80%±2%), LEDR-600 PID control incubator (temperature stability ±0.3℃), Epson Expression 12000XL flatbed scanner (600 dpi resolution, 48-bit color depth), NVIDIA GeForce RTX 3060 GPU workstation (12GB video memory).

[0103] Software environment: Python 3.8, PyTorch 1.12, Scikit-learn 1.0, Optuna 3.1.

[0104] Experimental sample: 3,000-5,000 images of rice grains, from 669 rice varieties in field trials in the middle and lower reaches of the Yangtze River from 2022 to 2024, covering alkali spreading value levels 3-7 (of which level 3 accounted for 15.5%, level 4 for 23.2%, level 5 for 17.1%, level 6 for 17.0%, and level 7 for 26.9%).

[0105] Core module implementation steps:

[0106] (1) Standardized data acquisition and image preprocessing

[0107] Sample preparation: Select mature and plump brown rice, dehull and mill it into polished rice, remove damaged grains and impurities, and retain uniform samples.

[0108] Reaction treatment: Place individual grains of polished rice into the individual compartments of a 49-well reaction tray, and add a fixed volume of 1.7% KOH solution to each well, ensuring that the liquid level submerges the grains but does not exceed the height of the compartment (to prevent drifting during movement). After covering, place the tray in an incubator and incubate at (30±2)℃ for 23 hours.

[0109] Image acquisition: After the culture is complete, place the reaction tray in the scanner and acquire 48-bit color images (600 dpi resolution).

[0110] Preprocessing flow:

[0111] The location of reaction wells was determined using an adaptive threshold method, and single-grain rice regions were segmented using a YOLOv7 model (trained with 300 labeled grains, mAP@0.5=0.92).

[0112] After obtaining the images, we invited professional alkali spreading value testers from the Rice Quality Invention Center of the China National Rice Research Institute to rate the degree of digestion of each individual rice grain and construct the original dataset.

[0113] (2) Multimodal feature extraction

[0114] This invention extracts three types of features from preprocessed images, as follows:

[0115] Deep features:

[0116] Based on pre-trained models such as ConvNeXt-Base and RegNetY-32GF (ImageNet initialization), a 62-dimensional feature vector is extracted, containing 12 statistical features (mean, standard deviation, maximum / minimum, median, quantile, skewness, kurtosis, energy, root mean square, entropy) and the first 50 original feature values.

[0117] Texture features:

[0118] Local binary pattern (LBP, radius 3, neighborhood 8): 15-dimensional;

[0119] Histogram of Oriented Gradients (HOG, 9 orientation bins, 16×16 pixel units): 26-dimensional;

[0120] Multi-scale Gabor filter (4 directions, 3 scales): 51 dimensions;

[0121] Gradient features (amplitude, direction, intensity): 7 dimensions;

[0122] Statistical indicators (central tendency, dispersion measure, etc.): 10 dimensions;

[0123] Total of 110 dimensions.

[0124] Morphological characteristics:

[0125] The system comprises nine dimensions: boundary fractal dimension, diffusion coefficient, compactness, and rice-specific morphology index (SCI). The SCI calculation formula is as follows:

[0126] ;

[0127]

[0128] ;

[0129] ;

[0130] in, A core It refers to the area of ​​the core area. P core It is the perimeter of the core area. D i It is the distance from each discrete boundary pixel to the center of the core. N This represents the number of discrete boundary pixels. Here, α , β and γ It is a weighting coefficient that adjusts the relative importance of each feature.

[0131] (3) Feature engineering and model training

[0132] Traditional assessment of alkali spreading value (ASV) relies on manual observation of starch diffusion phenotypes in rice grains after alkali treatment, classifying them into 3-7 levels. This method suffers from technical problems such as strong subjectivity (large differences in judgment of transition zones), low efficiency (inability to conduct high-throughput screening), and insufficient precision (difficulty in quantifying subtle differences). The core objective of ASV-NNCM is to solve these problems through automated and intelligent methods, achieving high-precision, high-throughput, and objective classification and quantification of ASV levels. The implementation process is as follows: First, standardized images (224×224 pixels) are acquired using a customized experimental setup (49-well reaction tray, constant temperature incubator, high-precision scanner) to ensure consistent alkali treatment conditions and reduce noise. Then, multimodal features are extracted (depth features are captured from diffusion abstraction patterns using networks such as ConvNeXt, texture features such as LBP and HOG quantify microscopic details, morphometric features are calculated based on SCI to determine macroscopic morphology, and possible color features, etc.). After Z-score standardization, these features are fused into a high-dimensional feature vector through channel splicing. Next, feature engineering is performed (selecting the top 3 features from a set of 12 features based on F1 scores, and then integrating them using Pearson analysis). Correlation analysis (threshold ≥ 0.5) retains strong correlation features (simplified from 186 dimensions to 23 dimensions). Then, the optimized features are input into a custom three-layer deep neural network (ASV-NNCM) for qualitative classification (including an attention layer, using batch normalization, ReLU activation, dropout, and other operations, combined with optimization strategies such as feature noise enhancement, early stopping, and gradient pruning). At the same time, models such as random forests are used for quantitative regression. After verification through 10-fold cross-validation, a high classification accuracy is finally achieved.

[0133] Figure 2 Schematic diagram of the Alkali Value Feature Neural Network Classification System (ASV-NNCM). (a) Feature acquisition module, used to extract depth features, texture features, and morphometric features from alkali value images; (b) Feature optimization through Pearson correlation analysis (threshold=0.5) to select the best-performing feature subset; (c) Core alkali value classification head, composed of a multilayer perceptron (MLP) with hidden layers, used for feature verification and class prediction; (d) Optimization module, integrating techniques such as feature upsampling, mixed precision training, gradient pruning, feature noise enhancement, dropout, ensemble voting, and class weighting to improve the final verification performance.

[0134] Quantization model construction:

[0135] In ASVRS, the regression-based quantization process revolves around transforming ASVs from qualitative hierarchical values ​​into quantitative continuous values, as follows: First, multimodal features are input into a regression model (including random forest, gradient boosting, linear regression, ridge regression, etc.). Random forest and similar models improve performance through feature selection-embedded ASV-NNCM (FS) and hyperparameter optimization (HT). Feature selection preserves key features, and hyperparameter optimization uses the Optuna framework to find optimal parameters within a set search space (150 trials, 600-second timeout limit) to maximize the R² score of the validation set. During training, samples are divided into training, validation, and test sets in a 7:2:1 ratio. Input features undergo automated cleaning (filling NaN values, truncating extreme values, removing near-zero variance features) and standardization preprocessing. 10-fold cross-validation is used to evaluate model stability. Quantitative evaluation metrics include the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE).

[0136] Example 3

[0137] Using samples with alkali digestibility values ​​of grades 3-7 as test samples, the identification accuracy of this method was verified, and the results are as follows:

[0138] Images are obtained through a scanner, such as Figure 3 As shown.

[0139] Mathematical coarse segmentation combined with YOLOv7 fine segmentation yields individual image cells, eliminating border interference, such as... Figure 4 As shown.

[0140] The alkali digestion value of each rice grain was determined using the ASV-NNCM model and quantified using a random forest model with feature selection and hyperparameter tuning.

[0141] Please have a professional determine the alkali digestibility level for data verification.

[0142] The final results are shown in Tables 1 and 2:

[0143] Table 1 Comparison results of different classification models

[0144]

[0145] Table 2 Ablation experiments with different regression models

[0146]

[0147] Note: FS represents feature selection result, and HT represents hyperparameter optimization.

[0148] The specific results after quantification:

[0149] The Alkali Spreading Value (ASVRS) system established in this invention provides a comprehensive intelligent solution for rice ASV assessment through the systematic integration of its technical architecture. This system eliminates environmental interference through standardized imaging and employs multimodal feature fusion optimized by hierarchical machine learning (combining deep features from convolutional neural networks / residual networks with texture and morphometric features), forming a complete process from phenotypic imaging to feature analysis. Addressing the boundary ambiguity issue in the 5-level classification, the independently developed ASV-NNCM model, embedding variety-specific mechanistic knowledge, achieved an accuracy of 94%, significantly outperforming manual assessment (70%-90%), providing strong technical support for replacing manual methods. Furthermore, its quantitative predictive ability (coefficient of determination R²=0.92, root mean square error RMSE=0.23) accurately captures continuous phenotypic changes in the transition stages between grades, overcoming the limitations of traditional discrete grading in characterizing continuous reactions and enabling refined assessment of rice starch quality. Figure 5 As shown.

[0150] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A high-throughput classification and quantitative determination method for rice alkali digestibility based on multi-source characteristics, characterized in that, The method includes: Rice sample data was obtained, and depth features, texture features and morphometric features were extracted. The depth features were 62-dimensional vectors containing 12 statistical features and the first 50 original feature values. Optimization is performed on depth features, texture features, and morphometric features, and the optimal feature combination is determined through quantitative evaluation. Based on the constructed alkali digestion value feature neural network classification system, feature selection is performed on the optimal feature combination to obtain key features of rice alkali digestion value; the method for feature selection on the optimal feature combination based on the constructed alkali digestion value feature neural network classification system includes: Based on the F1 score, 12 statistical features were evaluated, and the top three performing features were selected and integrated to obtain a feature set. Then, Pearson correlation analysis was used to optimize the features and select the best feature subset. Multiple regression models were trained and validated based on the key features of rice alkali removal value, the optimal regression model was determined, and based on the optimal regression model, the best feature combination-key feature screening-regression model GCR-FS-RF prediction model was constructed to realize the quantification of rice alkali removal value features. The performance of the GCR-FS-RF prediction model in distinguishing between adjacent alkali digestion value levels was verified by identifying the transition zone between adjacent alkali digestion value levels. The method for identifying the transition zone between adjacent alkali digestion value levels includes: using three methods in combination: Overlap Method (OM) determines the transition zone based on the overlap of regression predicted values; Density Intersection Method (DIM) fits the distribution curves of predicted values ​​for each level through kernel density estimation (KDE), and takes the predicted value corresponding to the intersection of curves of adjacent levels as the boundary of the transition zone; Statistical Method (SM) determines the transition zone based on the "±2 standard deviation confidence interval"; finally, a structured classification system of "original level + transition zone" is constructed based on the statistical method, expanding the original 5 levels into 9 classification units.

2. The method according to claim 1, characterized in that, Texture features are used to quantify surface heterogeneity; Morphometric characteristics include boundary fractal dimension, diffusion coefficient, compactness, and rice-specific morphology index (SCI), totaling nine dimensions.

3. The method according to claim 1, characterized in that, Based on the key characteristics of rice alkali disappearance, various regression models were trained and validated to determine the optimal regression model. A GCR-FS-RF prediction model was then constructed to quantify the characteristics of rice alkali disappearance. The methods include: Deep features, texture features, and morphometric features are input into the regression model, including random forest, gradient boosting, linear regression, and ridge regression. The regression model performs feature selection and hyperparameter optimization through an embedded alkaline elimination neural network classification model ASV-NNCM. Key features that meet the preset requirements are selected and retained. The hyperparameter optimization uses the Optuna framework to find the optimal parameters in the set search space to maximize the R² score of the validation set. During training, the samples were divided into training, validation, and test sets in a 7:2:1 ratio. The input features underwent automated cleaning and standardization preprocessing. 10-fold cross-validation was used to evaluate the model's stability. The quantitative evaluation metrics included the coefficient of determination (R²), root mean square error (RMS), and mean absolute error (MAE).

4. A high-throughput classification and quantification system for rice alkali degradation value based on multi-source characteristics, the system being used to implement the method described in any one of claims 1-3, characterized in that, The system includes: a feature extraction module, a combinatorial optimization module, a feature filtering module, a model building module, and a verification module; The feature extraction module is used to acquire rice sample data and extract depth features, texture features and morphometric features. The depth features are 62-dimensional vectors containing 12 statistical features and the first 50 original feature values. The combined optimization module is used to optimize depth features, texture features and morphometric features, and determine the optimal feature combination through quantitative evaluation. The feature selection module is used to select the optimal feature combination based on the constructed alkali digestibility feature neural network classification system to obtain key features of rice alkali digestibility; wherein, the method for selecting the optimal feature combination based on the constructed alkali digestibility feature neural network classification system includes: Based on the F1 score, 12 statistical features were evaluated, and the top three performing features were selected and integrated to obtain a feature set. Then, Pearson correlation analysis was used to optimize the features and select the best feature subset. The model building module is used to train and validate multiple regression models based on the key features of rice alkali removal, determine the optimal regression model, and construct the best feature combination-key feature screening-regression model GCR-FS-RF prediction model based on the optimal regression model to realize the quantification of rice alkali removal features. The verification module is used to verify the class differentiation performance of the GCR-FS-RF prediction model by identifying the transition zone between adjacent alkali digestion value classes.

5. The system according to claim 4, characterized in that, Texture features are used to quantify surface heterogeneity; Morphometric characteristics include boundary fractal dimension, diffusion coefficient, compactness, and rice-specific morphology index (SCI), totaling nine dimensions.

6. The system according to claim 4, characterized in that, The process of training and validating multiple regression models based on key characteristics of rice alkali disappearance, determining the optimal regression model, and constructing a GCR-FS-RF prediction model to quantify the characteristics of rice alkali disappearance includes: Deep features, texture features, and morphometric features are input into the regression model, including random forest, gradient boosting, linear regression, and ridge regression. The regression model performs feature selection and hyperparameter optimization through an embedded alkaline elimination neural network classification model ASV-NNCM. Key features that meet the preset requirements are selected and retained. The hyperparameter optimization uses the Optuna framework to find the optimal parameters in the set search space to maximize the R² score of the validation set. During training, the samples were divided into training, validation, and test sets in a 7:2:1 ratio. The input features underwent automated cleaning and standardization preprocessing. 10-fold cross-validation was used to evaluate the model's stability. The quantitative evaluation metrics included the coefficient of determination (R²), root mean square error (RMS), and mean absolute error (MAE).

7. The system according to claim 6, characterized in that, The process of identifying the transition zone between adjacent alkali digestion value levels includes: using three methods in combination: Overlap Method (OM) determines the transition zone based on the overlap of regression predicted values; Density Intersection Method (DIM) fits the distribution curves of predicted values ​​for each level through kernel density estimation (KDE), and takes the predicted value corresponding to the intersection of curves of adjacent levels as the boundary of the transition zone; Statistical Method (SM) determines the transition zone based on the "±2 standard deviation confidence interval"; finally, a structured classification system of "original level + transition zone" is constructed based on the statistical method, expanding the original 5 levels into 9 classification units.

Images