A Method and System for Intelligent Grading of Fresh Tea Leaves Based on Image Recognition

Abstract

This invention discloses an intelligent grading method and system for fresh tea leaves based on image recognition, belonging to the field of image recognition technology. The method includes: acquiring a forward image of the tea leaf area to be graded; dividing the forward image into multiple functional regions; performing image feature recognition to construct a multi-dimensional feature fingerprint set; judging the integrity of the visible area, predicting the feature confidence probability of the invisible area, and generating angle positioning annotation information for missing acquisition angles; obtaining a compensated acquisition image; performing compensated feature recognition based on the compensated acquisition image; and using the compensated recognition features, the multi-dimensional feature fingerprint set, and the feature confidence probability of the invisible area for fusion analysis and evaluation to obtain the grading result of the fresh tea leaves. This solves the technical problems of existing technologies that rely on manual experience for fresh tea leaf grading, resulting in low efficiency and inconsistent grading results, thus achieving the technical effect of improving the efficiency and accuracy of fresh tea leaf grading.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and specifically to a method and system for intelligent grading of fresh tea leaves based on image recognition. Background Technology

[0002] The grading of fresh tea leaves is a crucial step in the tea processing process, decisively impacting the final quality and market value of the tea. Traditional methods of grading fresh tea leaves rely primarily on manual experience, requiring graders to meticulously observe and evaluate the appearance, color, and shape of the leaves using their accumulated sensory judgment. However, this manual grading method has several significant drawbacks. Firstly, manual grading is extremely inefficient. Facing large-scale demands for grading fresh tea leaves, it often requires substantial manpower and time, leading to prolonged grading cycles, increased processing time costs, and potentially hindering the timely entry of tea into subsequent processing stages, thus affecting its freshness and quality. Secondly, the accuracy of manual grading is difficult to guarantee. Differences in the experience levels and subjective judgment standards of different graders can easily result in inconsistent grading outcomes, making it impossible to achieve standardization and normalization in fresh tea leaf grading and failing to meet the market's demand for a stable supply of high-quality tea. Summary of the Invention

[0003] This application provides a method and system for intelligent grading of fresh tea leaves based on image recognition, which solves the technical problems of low efficiency and inconsistent grading results caused by reliance on manual experience in the grading process of fresh tea leaves in the prior art.

[0004] The first aspect of this application provides a method for intelligent grading of fresh tea leaves based on image recognition, the method comprising:

[0005] A forward-facing image is acquired from the tea leaf area to be graded using a connected image acquisition device. Based on the structural characteristics of the fresh tea leaves, the forward-facing image is divided into multiple functional regions, including the leaf tip, leaf edge, midrib, and leaf base. Image feature recognition is performed within these functional regions to extract significant features such as color, texture, shape, and edge regions, constructing a multi-dimensional feature fingerprint set. The integrity of the visible area is assessed based on the multi-dimensional feature fingerprint set, predicting the feature confidence probability of the invisible area. Based on the feature confidence probability, the missing acquisition angle is determined, generating angle positioning annotation information. The image acquisition device is then controlled to perform compensated acquisition based on the angle positioning annotation information, obtaining a compensated acquisition image. Compensated feature recognition is performed on the compensated acquisition image, and the compensated recognition features are fused and analyzed with the multi-dimensional feature fingerprint set and the feature confidence probability of the invisible area to obtain the grading result of the fresh tea leaves.

[0006] A second aspect of this application provides an intelligent grading system for fresh tea leaves based on image recognition, the system comprising:

[0007] Image Acquisition Module: Connects to an image acquisition device to acquire a forward image of the tea leaf area to be graded, obtaining a forward image; Region Division Module: Divides the forward image into multiple functional regions based on the structural characteristics of the fresh tea leaves, including the leaf tip region, leaf edge region, midrib region, and leaf base region; Feature Extraction Module: Performs image feature recognition in the functional regions, extracting significant features of color, texture, shape, and edge regions to construct a multi-dimensional feature fingerprint set; Integrity Judgment Module: Judges the integrity of visible areas based on the multi-dimensional feature fingerprint set, predicts the feature confidence probability of invisible areas, and generates angle positioning annotation information based on the feature confidence probability of missing areas; Compensation Acquisition Module: Controls the image acquisition device to perform compensation acquisition based on the angle positioning annotation information, obtaining a compensated acquisition image; Grading Evaluation Module: Performs compensation feature recognition based on the compensated acquisition image, and uses the compensated recognition features, multi-dimensional feature fingerprint set, and feature confidence probability of invisible areas for fusion analysis and evaluation to obtain the grading result of the fresh tea leaves.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] First, an image acquisition device is connected to acquire a forward-facing image of the tea leaf area to be graded. Next, based on the structural characteristics of the fresh tea leaves, the forward-facing image is divided into multiple functional regions, including the leaf tip, leaf margin, midrib, and leaf base. Further, image feature recognition is performed within these functional regions to extract significant features such as color, texture, shape, and edge regions, constructing a multi-dimensional feature fingerprint set. Then, based on the multi-dimensional feature fingerprint set, the integrity of the visible area is assessed, the feature confidence probability of the invisible area is predicted, and the missing acquisition angle is determined based on the feature confidence probability, generating angle positioning annotation information. Next, the image acquisition device is controlled to perform compensated acquisition based on the angle positioning annotation information, obtaining compensated acquisition images. Finally, compensated feature recognition is performed on the compensated acquisition images, and the compensated recognition features are fused and analyzed with the multi-dimensional feature fingerprint set and the feature confidence probability of the invisible area to obtain the grading result of the fresh tea leaves. This solves the technical problems of existing technologies that rely on manual experience for fresh tea leaf grading, resulting in low efficiency and inconsistent grading results. It achieves the technical effect of automating and intelligently completing fresh tea leaf grading through image recognition technology, improving the efficiency and accuracy of fresh tea leaf grading. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0011] Figure 1 A schematic diagram of the intelligent grading method for fresh tea leaves based on image recognition provided in an embodiment of this application;

[0012] Figure 2 A schematic diagram of the structure of an image recognition-based intelligent grading system for fresh tea leaves provided in an embodiment of this application.

[0013] Figure labeling: Image acquisition module 11, Region segmentation module 12, Feature extraction module 13, Integrity judgment module 14, Compensation acquisition module 15, Grading evaluation module 16. Detailed Implementation

[0014] This application provides an intelligent grading method and system for fresh tea leaves based on image recognition, which solves the technical problems of low efficiency and inconsistent grading results in the existing tea leaf grading process due to reliance on manual experience.

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0016] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0017] Example 1, as Figure 1 As shown, this application provides an intelligent grading method for fresh tea leaves based on image recognition, wherein the method includes:

[0018] Connect the image acquisition device to acquire a forward image of the tea area to be graded.

[0019] By connecting the image acquisition device (such as an industrial camera, high-definition camera, or imaging device with automatic focus and exposure adjustment functions) to the control system via wired or wireless means, it can perform standardized and stable image acquisition operations on the target area where the fresh tea leaves to be graded are located under preset acquisition parameters, thereby obtaining a positive image.

[0020] Based on the structural characteristics of fresh tea leaves, the forward image is divided into regions to obtain multiple functional regions, including the leaf tip region, leaf margin region, midrib region, and leaf base region.

[0021] Based on the acquired forward images and combined with the morphological and structural features of fresh tea leaves in their natural state, image analysis algorithms are used to accurately identify and segment different parts, resulting in functional region divisions with clear structures and boundaries. Specifically, edge detection, contour extraction, and morphological processing methods are used to identify the boundary contours of the entire fresh tea leaf in the image; based on the geometric morphological features and symmetry axis of the leaf in the image, the direction of the main vein of the leaf is determined as the reference axis for functional region division; based on the reference axis, according to the natural growth structure of the leaf, the area near the top of the leaf is identified as the leaf tip area, the band-like area around the entire leaf boundary is identified as the leaf margin area, the linear structure distributed along the central axis of the leaf is identified as the main vein area, and the lower part near the petiole is identified as the leaf base area; through region labeling, image identification information for four types of functional regions is formed.

[0022] Image feature recognition is performed in the functional area to extract significant features of color, texture, shape, and edge regions, and to construct a multi-dimensional feature fingerprint set.

[0023] For multiple functional areas such as the leaf tip, leaf margin, midrib, and leaf base, various image processing and analysis techniques are used to extract color features, texture features, shape features, and edge salient features within the areas.

[0024] Color feature extraction employs a color space conversion method, transforming the RGB color space into HSV or Lab color space, and calculating the color histogram, mean, and variance of each functional region to reflect the color variations of the leaf. Texture feature extraction utilizes algorithms such as Gray-Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP) to obtain indicators such as contrast, uniformity, entropy, and texture directionality, reflecting the subtle structural differences on the leaf surface. Shape feature extraction extracts the leaf morphology and contour features of each functional region through contour analysis and edge detection. Edge region saliency feature recognition combines the Sobel operator and Canny edge detection algorithm to identify the edge clarity and morphological saliency of each region of the leaf. These multi-dimensional color, texture, shape, and edge feature data are integrated to construct a multi-dimensional feature fingerprint set representing each functional region of the fresh tea leaf, used to describe the visual information of different regions of the leaf.

[0025] The integrity of the visible region is determined based on the multi-dimensional feature fingerprint set, the feature confidence probability of the invisible region is predicted, and the missing acquisition angle is determined based on the feature confidence probability to generate angle positioning annotation information.

[0026] Based on a multi-dimensional feature fingerprint set, the integrity of visible areas in tea leaf images is assessed. Specifically, this involves analyzing the integrity indices of color, texture, shape, and edge features extracted from functional regions, combined with a preset integrity threshold, to evaluate whether feature loss or occlusion exists in each functional region of the currently acquired image. For detected incomplete or missing features in the visible region, a trained functional region prediction model is used. Based on existing feature association rules and historical sample data, the model predicts the feature representation of the invisible region and calculates the corresponding feature confidence probability, reflecting the reliability of the predicted features. Based on the feature confidence probability, for invisible regions with confidence scores below a preset threshold, further analysis of the missing acquisition angle is performed. By analyzing the spatial structure of the tea leaves and the viewing angle limitations of the acquisition device, the optimal compensation acquisition angle and corresponding positioning information are determined. This angle positioning information is used to generate angle positioning labels to guide the image acquisition device in adjusting its acquisition posture, thereby enabling supplementary acquisition of invisible regions.

[0027] Furthermore, based on the multi-dimensional feature fingerprint set, the integrity of the visible region is determined, the feature confidence probability of the invisible region is predicted, and the missing acquisition angle is determined based on the feature confidence probability to generate angle positioning annotation information, including:

[0028] Based on the multi-dimensional feature fingerprint set, the functional region structure ratio of fresh tea leaves is analyzed to obtain the visible region ratio. Region integrity is determined according to the image features of the functional regions and the visible region ratio to obtain integrity judgment results. For fresh tea leaves whose integrity judgment results do not meet the threshold, feature association prediction is performed on the invisible regions based on the visible region structure ratio and feature fingerprint to obtain predicted features and feature confidence probabilities. For invisible regions whose feature confidence probabilities do not reach the confidence threshold, missing acquisition angle analysis is performed to obtain positioning information and corresponding acquisition angle information, generating the angle positioning annotation information.

[0029] Specifically, based on the boundary contour and spatial location of each functional area, image segmentation algorithms (such as GrabCut, semantic segmentation, or edge-enhanced region growing algorithms) are used to accurately extract the pixel range of each functional area in the image; the number of pixels in each functional area is counted and the ratio is calculated with the total number of pixels in the whole image or the whole tea leaf to obtain the proportion of the visible area of ​​each functional area in the current image.

[0030] Based on the extracted image features of each functional region (including color mean and distribution, texture direction and density, edge continuity, shape closure, etc.) and the previously calculated visible area proportion of the functional region, a region integrity determination is performed for each functional region. Specifically, the image features of each region are matched and analyzed with the corresponding standard complete feature template in the sample library. Combined with its area proportion in the image, a functional region integrity score is calculated. The integrity score reflects whether the region has occlusion, cropping, distortion, folding, or missing features. The integrity score is compared with a preset integrity threshold. When the score is higher than or equal to the threshold, the region is considered complete; when the score is lower than the threshold, the region is considered to have missing integrity. The integrity determination results of each functional region are summarized to form the integrity determination result of the currently acquired image, which is used to evaluate whether the image has the feature completeness for hierarchical analysis.

[0031] Optionally, for each functional region, a standard complete feature template corresponding to that region is retrieved from a pre-established sample library. This template contains multi-dimensional feature indicators such as typical color distribution, texture features, shape contours, and edge information. The multi-dimensional feature fingerprint extracted from the functional region in the currently acquired image is matched with the standard template, and the feature matching degree is calculated using a similarity measurement method (such as cosine similarity, Euclidean distance, or correlation coefficient). Combined with the actual visible area ratio of the region in the image, a weighted fusion strategy is used to calculate the integrity score of the functional region, where the area ratio is used as a weighting factor to reflect the influence of the visibility of the region on the integrity score. The numerical range of the integrity score is usually normalized to between 0 and 1. The closer the value is to 1, the higher the consistency between the functional region and the standard complete features, and the better the integrity. Conversely, it indicates that there are missing or abnormal features.

[0032] For tea leaves whose integrity assessment results do not reach the threshold, based on the structural proportion of their visible areas and the corresponding multi-dimensional feature fingerprints, a functional region prediction model is used to predict the feature association of invisible areas, inferring the possible features of the invisible areas and calculating the confidence probability of the predicted features. For invisible areas where the confidence probability of the predicted features is lower than the confidence threshold, missing acquisition angle analysis is performed. Combining the spatial structural features of the tea leaves and the angle adjustment range of the image acquisition equipment, the optimal acquisition angle and corresponding spatial positioning information of the missing feature areas are determined, and accurate angle positioning annotation information is generated to guide the image acquisition equipment to perform compensated acquisition.

[0033] Furthermore, based on the structural proportion and feature fingerprint of the visible region, feature association prediction is performed on the invisible region to obtain predicted features and feature confidence probabilities, including:

[0034] Based on the sample set collected from functional areas, feature association analysis of the sample set is performed according to multi-dimensional features to establish a prediction sub-model for each functional area; the growth and spread relationship of each functional area is analyzed, association weights are configured, and the prediction sub-models of each functional area are connected based on the association weights to construct a feature prediction model; based on the feature fingerprint of the currently identified visible area, the feature prediction model is used to predict the features of the invisible area to obtain the predicted features, and the confidence probability analysis of the predicted features is performed; the confidence probability of the predicted features is corrected according to the structural proportion of the visible area to determine the predicted features and the feature confidence probability.

[0035] Based on a regional sample set collected from functional areas, feature association analysis is performed between samples using multi-dimensional image features (including color, texture, shape, etc.) to establish a prediction sub-model for each functional area. Secondly, the growth and spread relationships between functional areas are analyzed. Combining the topological structure and spatial distribution of fresh tea leaves during natural growth, association weights are configured between functional areas. These weights are calculated based on the similarity of image features between areas, spatial proximity, and growth logic. These weights are then used to connect the prediction sub-models of each functional area, constructing an overall feature prediction model. The feature fingerprints of the visible areas identified in the currently collected images are input into the feature prediction model to deduce and predict the feature performance of invisible areas, obtaining the corresponding predicted features. Simultaneously, confidence probability analysis is performed on the prediction results to quantify the reliability of the prediction. Based on the structural proportion of the visible areas, the confidence probability of the predicted features is adjusted. Considering the impact of the integrity of the visible areas on the prediction of invisible areas, the final predicted features and their confidence probabilities are determined.

[0036] Furthermore, based on the regional sample sets collected from functional areas, feature correlation analysis of the regional sample sets is performed according to multi-dimensional characteristics to establish prediction sub-models for each functional area, including:

[0037] For the leaf tip, leaf margin, midrib, and leaf base areas, regional sample sets are collected for each functional region, including sample features and corresponding complete tea leaf structures. The regional sample sets include positive and negative samples. Positive samples are undamaged tea leaves, and negative samples are tea leaves with damage characteristics. For each functional region, multi-dimensional image features such as color, texture, midrib distribution, and edge morphology are extracted from the positive and negative samples. The correspondence between multi-dimensional image features and complete structures, as well as the feature point distribution, feature transition patterns, and abnormal evolution trends within each functional region, are combined to construct a feature evolution path chain and generate a training dataset. Machine learning modeling is performed using the training dataset, and the model parameters are converged and verified to obtain prediction sub-models for each functional region, which are used to predict the complete feature state or abnormal feature manifestations within the functional region point by point.

[0038] For the leaf tip, leaf margin, midrib, and leaf base areas, regional sample sets were collected for each functional region. These regional sample sets contained multidimensional image features of the region and corresponding annotations of the complete structure of fresh tea leaves. The sample sets were divided into positive and negative samples, where positive samples were undamaged and structurally intact fresh tea leaves, and negative samples were fresh tea leaves exhibiting different types of damage. For each functional region, multidimensional image features such as color, texture, midrib distribution, and edge morphology were extracted from both positive and negative samples. Simultaneously, the correspondence between these multidimensional image features and the complete structure of the tea leaf was analyzed to determine the spatial distribution patterns, feature transition modes, and abnormal evolution trends of feature points within each functional region. An evolutionary path chain reflecting the feature change path was constructed, forming a sample dataset for training. Based on this training dataset, machine learning methods (such as support vector machines, random forests, or deep neural networks) were used for model training, and the model parameters were validated through convergence to obtain predictive sub-models for each functional region. These predictive sub-models can accurately predict and extrapolate the complete feature state or abnormal damage features within a functional region point by point.

[0039] In establishing the predictive sub-models for each functional region, supervised learning-based machine learning algorithms are employed, with Support Vector Machine (SVM), Random Forest, and Deep Neural Network (DNN) as the main modeling methods. Specifically, firstly, the aforementioned multi-dimensional image features (including color features, texture features, main vein distribution features, and edge morphology features) and their corresponding positive and negative examples are input into the model as training data. Subsequently, the training and validation sets are divided according to a cross-validation strategy, and the model parameters are optimized using gradient descent (for DNN) or heuristic search (for SVM and Random Forest). Iterative training continues until the loss function converges or reaches a preset threshold. Finally, a predictive sub-model is obtained that can be used to infer the complete feature state and abnormal behavior of functional regions point by point, accurately determining the health and damage status within a region based on the input feature fingerprint.

[0040] Furthermore, the growth and expansion relationships of each functional area are analyzed, and associated weights are configured, including:

[0041] Based on the natural structure of fresh tea leaves, functional areas are constructed as nodes in a structural topology graph. The spatial location, growth direction, leaf morphological symmetry, and main vein extension relationship between each functional area are analyzed to establish node connection edges. Each node connection edge is assigned an association weight, which is calculated based on the image feature similarity, spatial distance, and structural connection relationship between regions, and is used to represent the degree of propagation and influence of tea feature information between different functional areas.

[0042] Based on the natural structural characteristics of fresh tea leaves, each functional region is abstracted as a node in a structural topology diagram, specifically including nodes such as the leaf tip region, leaf margin region, midrib region, and leaf base region. By analyzing the spatial relationships, growth directions, leaf morphological symmetry, and midrib extension relationships among the functional regions, connection edges are established between the functional region nodes to reflect the growth and spread paths and structural dependencies between the regions.

[0043] Based on the natural growth structure of fresh tea leaves, the functional regions (leaf tip region, leaf margin region, midrib region, and leaf base region) are abstracted into nodes in the structural topology diagram, with each node representing a functional region. By analyzing the spatial relationship, growth direction, leaf morphology symmetry, and midrib extension relationship between the functional regions, connection edges between functional region nodes are established to reflect the growth and spread paths and structural dependencies between the regions.

[0044] For each node connection edge in the structural topology graph, the calculation of the association weight comprehensively considers the image feature similarity, spatial distance, and structural connectivity between regions. Specifically, firstly, the multidimensional image feature similarity connecting two functional regions is calculated, using indicators such as cosine similarity or correlation coefficient to measure the degree of matching of color, texture, and shape features. Secondly, based on the spatial coordinates of the two regions in the image, Euclidean distance or Manhattan distance is calculated; the closer the distance, the higher the association weight, and vice versa. Finally, considering the structural connectivity of fresh tea leaves, such as the continuity of the main vein extension, the connectivity of the leaf vein network, and morphological symmetry, the association weight is structurally corrected and weighted. Taking all these factors into account, the final association weight is calculated using a weighted summation or product method. This weight value reflects the strength and influence of feature propagation between functional regions.

[0045] Furthermore, determining the predicted features and the confidence probabilities of the features also includes:

[0046] Based on the association weights of adjacent functional areas in the structural topology graph, the confidence level of association prediction is verified. When the similarity of the prediction results of multiple adjacent functional areas reaches a similarity threshold, the confidence probability of the feature is increased; when the difference of the prediction results of adjacent functional areas reaches a difference threshold, the confidence probability of the feature is decreased.

[0047] In determining the predicted features and their confidence probabilities for invisible regions, the association weights between adjacent functional regions in the structural topology graph are used to verify the confidence of the association predictions. Specifically, the similarity of the predicted feature results of adjacent functional regions is calculated, and cosine similarity, correlation coefficient, or other similarity metrics are used to measure the consistency between the prediction results of multiple adjacent regions. When the similarity of the prediction results of adjacent functional regions reaches a preset similarity threshold, it indicates that the predicted features of each region are highly consistent, thus increasing the confidence probability of the predicted features of that region and enhancing the reliability of the prediction. Conversely, when the difference between the prediction results of adjacent functional regions (e.g., 1 minus similarity) reaches or exceeds a preset difference threshold, it indicates that there is a significant inconsistency in the prediction results, and the feature confidence probability of the corresponding region needs to be reduced, reflecting the uncertainty and potential error of the prediction results.

[0048] Based on the angle positioning annotation information, the image acquisition device is controlled to perform compensated acquisition to obtain a compensated acquired image.

[0049] Based on the angle positioning annotation information, the system sends control commands to the image acquisition device to adjust the shooting angle and position of the acquisition device, thereby achieving compensatory acquisition of invisible areas of fresh tea leaves. Specifically, according to the preset missing acquisition angle and corresponding spatial position information in the angle positioning annotation, the system controls the camera's rotation angle, tilt angle, and position relative to the tea leaves, enabling the acquisition device to acquire images of the target fresh tea leaves from a new perspective. During the acquisition process, the image acquisition device provides real-time feedback on the acquisition status, ensuring that the compensated acquisition image covers the previously missing or occluded areas. The final compensated acquisition image contains supplementary leaf information, effectively compensating for the invisible or missing feature areas in the previous forward acquisition.

[0050] Compensation feature recognition is performed on the compensated images, and the compensated recognition features are fused with multi-dimensional feature fingerprint sets and feature confidence probabilities of invisible areas for evaluation to obtain the grading results of fresh tea leaves.

[0051] After the compensated image is acquired, the system performs image feature recognition operations on the functional areas contained in the compensated image to extract multi-dimensional image features such as color, texture, edge morphology, and main vein structure to form compensation recognition features.

[0052] After acquiring the compensated image and completing the compensation feature recognition, the system fuses the extracted compensation recognition features with the multi-dimensional feature fingerprint set constructed during the initial acquisition. During the fusion process, weights are assigned according to the feature dimension type of each functional region (including color uniformity, main vein integrity, edge integrity, texture fineness, etc.), and the confidence probability of the features in the invisible region is combined to perform a confidence-weighted correction on the compensation features.

[0053] After completing the fusion feature construction of all functional areas, the following indicators are statistically analyzed: color consistency (calculated using RGB mean and standard deviation), texture clarity (contrast and energy values ​​obtained based on image gray-level co-occurrence matrix analysis), edge integrity (edge ​​closure rate obtained through Canny edge detection), and shape symmetry (difference ratio calculated through leaf principal axis symmetry). Based on the manual rules set for the above multiple indicators, each functional area is assigned a grade score (e.g., A, B, C). The scores of each functional area are weighted and summarized according to their respective weights (e.g., 40% for the main vein area, 30% for the leaf edge area, 20% for the leaf tip area, and 10% for the leaf base area) to form the final score of the fresh tea leaves. Based on the score range, the fresh tea leaves are divided into four grades: excellent (first-class), good (second-class), qualified (third-class), and unqualified (inferior), and the final grading results are output.

[0054] Furthermore, based on the compensated acquired images, compensation feature recognition is performed, which then includes:

[0055] The location features are stitched together using the compensated recognition features and the multi-dimensional feature fingerprint set to reset the feature fingerprint set; the feature integrity and main recognition features are evaluated using the reset feature fingerprint set to establish a fingerprint evaluation list; based on the fingerprint evaluation list, combined with the forward image acquisition angle and the compensated acquisition angle, acquisition angle analysis is performed to determine whether there are feature missing regions; if there are residual feature missing regions, acquisition position and angle analysis is performed, and the image acquisition device is controlled to perform compensated acquisition again.

[0056] Preferably, the localization features are stitched together using the compensated recognition features and the existing multi-dimensional feature fingerprint set. By matching the functional area coordinate index and the main feature distribution position, the newly added or corrected key image features in the compensated image are fused to generate a reset feature fingerprint set. Subsequently, based on the reset feature fingerprint set, the feature integrity assessment of the functional area and the expression quality evaluation of the main recognition features are performed to construct a fingerprint evaluation list including key features, weight scores, and stability labels for each functional area. Based on the fingerprint evaluation list, combined with the direction and pitch information of the forward image acquisition angle and the compensated acquisition angle, an acquisition angle analysis operation is performed to determine whether there are still residual feature missing phenomena in the functional area. If it is detected during the analysis that one or more functional areas still cannot complete the complete feature coverage under the existing angle combination, the optimized acquisition position and angle configuration of the image acquisition device are calculated based on the image missing direction of the residual area, the area occlusion mode, and the position information in the reset fingerprint set. A compensated acquisition command for device driving is generated to control the image acquisition device to automatically execute the next round of compensated image acquisition to achieve complete coverage of the target area features of fresh tea leaves.

[0057] Furthermore, if residual feature loss areas exist, the acquisition position and angle are analyzed, and the image acquisition device is controlled to perform re-compensation acquisition, including:

[0058] Obtain the angle adjustment range of the image acquisition device; using the angle adjustment range as a constraint, perform angle compensation search on the residual feature missing area; when the compensation relationship is satisfied, determine the acquisition position and angle to generate control commands to control the image acquisition device to perform compensation acquisition again.

[0059] When the system detects a region with missing residual features, it acquires the angle adjustment range of the image acquisition device. This range includes the adjustable limits of the device's rotation angle, pitch angle, and displacement space, ensuring that subsequent acquisition operations are performed within the physical and mechanical limits allowed by the device.

[0060] Based on the angle adjustment range, the system performs angle compensation search for regions with missing residual features. Specifically, it calculates the optimal acquisition angle that can cover the missing region by traversing preset angle combinations or using optimization algorithms (such as greedy search or genetic algorithms). It then determines whether the selected angle satisfies the compensation relationship, meaning that the angle can effectively supplement missing features and improve the integrity of regional features. When the compensation relationship is satisfied, the system determines the corresponding acquisition position and angle parameters, generates precise control commands, sends these commands to the image acquisition device, drives the device to adjust to the specified position and angle, and performs another compensation acquisition operation.

[0061] Furthermore, the process includes performing an angle-compensated search on the region with missing residual features, followed by:

[0062] When the compensation relationship is not satisfied, the confidence probability analysis of the current feature evaluation is performed based on the reset feature fingerprint set to obtain the graded confidence level; when the graded confidence level reaches a preset threshold, the current feature fingerprint is determined; when the graded confidence level does not reach the preset threshold, the distribution area of ​​fresh tea leaves is located, and a sorting and isolation instruction is sent to control the sorting equipment to perform sorting and isolation processing according to the location area.

[0063] After performing angle compensation search on the regions with missing residual features, if no acquisition angle that satisfies the compensation relationship is found, meaning that the compensation acquisition cannot effectively compensate for the missing features, the system performs an evaluation confidence probability analysis of the current features based on the reset feature fingerprint set, calculating the overall grading confidence level. This confidence level reflects the reliability of the current image features in determining the grading of fresh tea leaves. If the grading confidence level reaches a preset confidence threshold, the system confirms that the current feature fingerprint is sufficient to support effective grading and outputs the corresponding fresh tea leaf grading result. If the grading confidence level does not reach the preset threshold, it indicates that the feature information is insufficient to guarantee the accuracy of grading. The system further performs a location analysis of the distribution area of ​​fresh tea leaves to identify specific leaf areas that may have abnormalities or defects. Based on the location results, the system generates sorting and isolation instructions, controlling the sorting equipment to sort and isolate the fresh tea leaves according to the location areas.

[0064] In summary, the embodiments of this application have at least the following technical effects:

[0065] First, an image acquisition device is connected to acquire a forward-facing image of the tea leaf area to be graded. Next, based on the structural characteristics of the fresh tea leaves, the forward-facing image is divided into multiple functional regions, including the leaf tip, leaf margin, midrib, and leaf base. Further, image feature recognition is performed within these functional regions to extract significant features such as color, texture, shape, and edge regions, constructing a multi-dimensional feature fingerprint set. Then, based on the multi-dimensional feature fingerprint set, the integrity of the visible area is assessed, the feature confidence probability of the invisible area is predicted, and the missing acquisition angle is determined based on the feature confidence probability, generating angle positioning annotation information. Next, the image acquisition device is controlled to perform compensated acquisition based on the angle positioning annotation information, obtaining compensated acquisition images. Finally, compensated feature recognition is performed on the compensated acquisition images, and the compensated recognition features are fused and analyzed with the multi-dimensional feature fingerprint set and the feature confidence probability of the invisible area to obtain the grading result of the fresh tea leaves. This solves the technical problems of existing technologies that rely on manual experience for fresh tea leaf grading, resulting in low efficiency and inconsistent grading results. It achieves the technical effect of automating and intelligently completing fresh tea leaf grading through image recognition technology, improving the efficiency and accuracy of fresh tea leaf grading.

[0066] Example 2, based on the same inventive concept as the image recognition-based intelligent grading method for fresh tea leaves in the aforementioned examples, such as... Figure 2 As shown, this application provides an intelligent grading system for fresh tea leaves based on image recognition, wherein the system includes:

[0067] Image acquisition module 11: Connects to the image acquisition device to acquire a forward image of the tea leaf area to be graded, obtaining a forward image; Region division module 12: Divides the forward image into multiple functional regions based on the structural characteristics of the fresh tea leaves, including the leaf tip region, leaf edge region, midrib region, and leaf base region; Feature extraction module 13: Performs image feature recognition in the functional regions, extracts significant features of color, texture, shape, and edge regions, and constructs a multi-dimensional feature fingerprint set; Integrity judgment module 14: Makes a visible region integrity judgment based on the multi-dimensional feature fingerprint set, predicts the feature confidence probability of the invisible region, and generates angle positioning annotation information based on the feature confidence probability; Compensation acquisition module 15: Controls the image acquisition device to perform compensation acquisition based on the angle positioning annotation information, obtaining a compensated acquisition image; Grading evaluation module 16: Performs compensation feature recognition based on the compensated acquisition image, and uses the compensation recognition features, multi-dimensional feature fingerprint set, and feature confidence probability of the invisible region for fusion analysis and evaluation to obtain the grading result of the fresh tea leaves.

[0068] Furthermore, the grading evaluation module 16 is used to perform the following method:

[0069] The location features are stitched together using the compensated recognition features and the multi-dimensional feature fingerprint set to reset the feature fingerprint set; the feature integrity and main recognition features are evaluated using the reset feature fingerprint set to establish a fingerprint evaluation list; based on the fingerprint evaluation list, combined with the forward image acquisition angle and the compensated acquisition angle, acquisition angle analysis is performed to determine whether there are feature missing regions; if there are residual feature missing regions, acquisition position and angle analysis is performed, and the image acquisition device is controlled to perform compensated acquisition again.

[0070] Furthermore, the grading evaluation module 16 is used to perform the following method:

[0071] Obtain the angle adjustment range of the image acquisition device; using the angle adjustment range as a constraint, perform angle compensation search on the residual feature missing area; when the compensation relationship is satisfied, determine the acquisition position and angle to generate control commands to control the image acquisition device to perform compensation acquisition again.

[0072] Furthermore, the grading evaluation module 16 is used to perform the following method:

[0073] When the compensation relationship is not satisfied, the confidence probability analysis of the current feature evaluation is performed based on the reset feature fingerprint set to obtain the graded confidence level; when the graded confidence level reaches a preset threshold, the current feature fingerprint is determined; when the graded confidence level does not reach the preset threshold, the distribution area of ​​fresh tea leaves is located, and a sorting and isolation instruction is sent to control the sorting equipment to perform sorting and isolation processing according to the location area.

[0074] Furthermore, the integrity determination module 14 is used to perform the following method:

[0075] Based on the multi-dimensional feature fingerprint set, the functional region structure ratio of fresh tea leaves is analyzed to obtain the visible region ratio. Region integrity is determined according to the image features of the functional regions and the visible region ratio to obtain integrity judgment results. For fresh tea leaves whose integrity judgment results do not meet the threshold, feature association prediction is performed on the invisible regions based on the visible region structure ratio and feature fingerprint to obtain predicted features and feature confidence probabilities. For invisible regions whose feature confidence probabilities do not reach the confidence threshold, missing acquisition angle analysis is performed to obtain positioning information and corresponding acquisition angle information, generating the angle positioning annotation information.

[0076] Furthermore, the integrity determination module 14 is used to perform the following method:

[0077] Based on the sample set collected from functional areas, feature association analysis of the sample set is performed according to multi-dimensional features to establish a prediction sub-model for each functional area; the growth and spread relationship of each functional area is analyzed, association weights are configured, and the prediction sub-models of each functional area are connected based on the association weights to construct a feature prediction model; based on the feature fingerprint of the currently identified visible area, the feature prediction model is used to predict the features of the invisible area to obtain the predicted features, and the confidence probability analysis of the predicted features is performed; the confidence probability of the predicted features is corrected according to the structural proportion of the visible area to determine the predicted features and the feature confidence probability.

[0078] Furthermore, the integrity determination module 14 is used to perform the following method:

[0079] For the leaf tip, leaf margin, midrib, and leaf base areas, regional sample sets are collected for each functional region, including sample features and corresponding complete tea leaf structures. The regional sample sets include positive and negative samples. Positive samples are undamaged tea leaves, and negative samples are tea leaves with damage characteristics. For each functional region, multi-dimensional image features such as color, texture, midrib distribution, and edge morphology are extracted from the positive and negative samples. The correspondence between multi-dimensional image features and complete structures, as well as the feature point distribution, feature transition patterns, and abnormal evolution trends within each functional region, are combined to construct a feature evolution path chain and generate a training dataset. Machine learning modeling is performed using the training dataset, and the model parameters are converged and verified to obtain prediction sub-models for each functional region, which are used to predict the complete feature state or abnormal feature manifestations within the functional region point by point.

[0080] Furthermore, the integrity determination module 14 is used to perform the following method:

[0081] Based on the natural structure of fresh tea leaves, functional areas are constructed as nodes in a structural topology graph. The spatial location, growth direction, leaf morphological symmetry, and main vein extension relationship between each functional area are analyzed to establish node connection edges. Each node connection edge is assigned an association weight, which is calculated based on the image feature similarity, spatial distance, and structural connection relationship between regions, and is used to represent the degree of propagation and influence of tea feature information between different functional areas.

[0082] Furthermore, the integrity determination module 14 is used to perform the following method:

[0083] Based on the association weights of adjacent functional areas in the structural topology graph, the confidence level of association prediction is verified. When the similarity of the prediction results of multiple adjacent functional areas reaches a similarity threshold, the confidence probability of the feature is increased; when the difference of the prediction results of adjacent functional areas reaches a difference threshold, the confidence probability of the feature is decreased.

[0084] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0085] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0086] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. An image recognition-based intelligent grading method for tea fresh leaves, characterized in that, The method comprises: The image acquisition device is connected to the area of the tea leaf to be graded to perform forward image acquisition, and a forward image is obtained; According to the structural characteristics of the tea leaf, the forward image is divided into regions to obtain a plurality of functional regions, including a tip region, a margin region, a main vein region, and a base region; Image feature recognition is performed in the functional regions to extract color, texture, shape, and edge region features, and a multi-dimensional feature fingerprint set is constructed; According to the multi-dimensional feature fingerprint set, the integrity of the visible region is judged, the feature confidence probability of the invisible region is predicted, and the missing acquisition angle is determined based on the feature confidence probability to generate angle positioning labeling information; The image acquisition device is controlled based on the angle positioning labeling information to perform compensation acquisition, and a compensation acquisition image is obtained; According to the compensation acquisition image, compensation feature recognition is performed, and the compensation recognition features, the multi-dimensional feature fingerprint set, and the feature confidence probability of the invisible region are fused and analyzed to obtain the grading result of the tea leaf; Wherein, the angle positioning labeling information is generated, including: According to the multi-dimensional feature fingerprint set, the structure proportion of the functional region of the tea leaf is analyzed to obtain the proportion of the visible region; According to the image features of the functional region and the proportion of the visible region, the region integrity is determined to obtain the integrity judgment result; For the tea leaf whose integrity judgment result does not meet the threshold, the feature correlation prediction of the invisible region is performed according to the structure proportion and feature fingerprint of the visible region to obtain the predicted feature and feature confidence probability; For the invisible region whose feature confidence probability does not reach the confidence threshold, the missing acquisition angle is analyzed to obtain the positioning information and the corresponding acquisition angle information, and the angle positioning labeling information is generated; According to the structure proportion and feature fingerprint of the visible region, the feature correlation prediction of the invisible region is performed to obtain the predicted feature and feature confidence probability, including: According to the functional region acquisition region sample set, the region sample set feature correlation analysis is performed according to the multi-dimensional features to establish a prediction sub-model of each functional region; The growth and spread relationship of each functional region is analyzed, the correlation weight is configured, the prediction sub-models of each functional region are connected based on the correlation weight, and a feature prediction model is constructed; According to the feature fingerprint of the visible region, the feature prediction of the invisible region is performed through the feature prediction model to obtain the predicted feature, and the predicted feature confidence probability is analyzed; The predicted feature and feature confidence probability are determined by correcting the predicted feature confidence probability according to the structure proportion of the visible region; Wherein, the growth and spread relationship of each functional region is analyzed, and the correlation weight is configured, including: According to the natural structure of the tea leaf, the functional region is constructed as a node in a structure topology graph, the spatial position, growth direction, leaf shape symmetry, and main vein extension relationship between the functional regions are analyzed, and a node connection edge is established; The correlation weight is allocated to each node connection edge, and the correlation weight is calculated based on the image feature similarity, spatial distance, and structure connection relationship between the regions, and is used to represent the propagation influence degree of the tea leaf feature information between different functional regions.

2. The image recognition based intelligent grading method of tea fresh leaves as claimed in claim 1 wherein, After the compensation feature recognition based on the compensation acquisition image, The compensation identification feature is spliced with a multi-dimensional feature fingerprint set to reset the feature fingerprint set; The reset feature fingerprint set is used for feature integrity and main identification feature evaluation to establish a fingerprint evaluation list; According to the fingerprint evaluation list, combined with the forward image acquisition angle and the compensation acquisition angle, the acquisition angle analysis is performed to determine whether there is a feature missing area; If there is a residual feature missing area, the acquisition position and angle analysis are performed to control the image acquisition device to perform re-compensation acquisition.

3. The image recognition based intelligent grading method of tea fresh leaves as claimed in claim 2 wherein, If there is a residual feature missing area, the acquisition position and angle analysis are performed to control the image acquisition device to perform re-compensation acquisition, including: Obtain the angle adjustment range of the image acquisition device; With the angle adjustment range as a constraint, the residual feature missing area is searched for angle compensation, and when the compensation relationship is satisfied, the acquisition position and angle generation control instruction are determined to control the image acquisition device to perform re-compensation acquisition.

4. The image recognition based intelligent grading method of tea fresh leaves as claimed in claim 3 wherein, The residual feature missing area is searched for angle compensation, and the following steps are further included: When the compensation relationship is not satisfied, the current feature evaluation confidence probability is analyzed according to the reset feature fingerprint set to obtain a hierarchical confidence; When the hierarchical confidence reaches a preset threshold, the current feature fingerprint is determined; When the hierarchical confidence does not reach the preset threshold, the tea fresh leaf distribution area is located, and a sorting isolation instruction is sent to control the sorting device to sort and isolate according to the located area.

5. The image recognition based intelligent grading method of tea fresh leaves as claimed in claim 1 wherein, According to the function area acquisition area sample set, the area sample set feature correlation analysis is performed according to the multi-dimensional feature to establish each function area prediction sub-model, including: For the tip area, the edge area, the main vein area and the leaf base area, the area sample set of each function area is collected, including sample features and corresponding complete structure tea fresh leaves, the area sample set includes positive samples and negative samples, the positive samples are undamaged tea fresh leaves, and the negative samples are tea fresh leaves with damage features; For each function area, color features, texture features, main vein distribution features and edge morphology features multi-dimensional image features are extracted from the positive and negative samples, and combined with the corresponding relationship between the multi-dimensional image features and the complete structure, as well as the feature point distribution, feature transition mode and abnormal evolution trend inside each function area, a feature evolution path chain is constructed to generate a training data set; Machine learning modeling is performed using the training data set, and model parameter convergence verification is performed to obtain the each function area prediction sub-model for point-by-point deduction of the complete feature state or abnormal feature performance inside the function area.

6. The image recognition based intelligent grading method of tea fresh leaves as claimed in claim 1 wherein, Determining the predicted feature and the feature confidence probability further includes: Based on the correlation weight of the adjacent function areas in the structure topology graph, the correlation prediction confidence is verified, when the prediction results of multiple adjacent function areas reach a similarity threshold, the feature confidence probability is improved; When the prediction results of adjacent function areas reach a difference threshold, the feature confidence probability is reduced.

7. An image recognition based intelligent grading system for tea fresh leaves characterized by, The system is used to implement the image recognition based tea fresh leaf intelligent grading method of any one of claims 1-6, and the system comprises: An image acquisition module: connect the image acquisition device to the region of the tea leaf to be graded to acquire a forward image; A region division module: according to the structural characteristics of the tea leaf, the forward image is divided into a plurality of functional regions, including a tip region, a margin region, a main vein region and a base region; A feature extraction module: image feature recognition is performed in the functional region, and color, texture, shape and edge region features are extracted to construct a multi-dimensional feature fingerprint set; An integrity judgment module: according to the multi-dimensional feature fingerprint set, the integrity of the visible region is judged, the feature confidence probability of the invisible region is predicted, and the missing collection angle is based on the feature confidence probability to generate angle positioning label information; A compensation acquisition module: based on the angle positioning label information, the image acquisition device is controlled to perform compensation acquisition to obtain a compensation acquisition image; A grading evaluation module: according to the compensation acquisition image, the compensation feature recognition is performed, the compensation recognition feature, the multi-dimensional feature fingerprint set and the feature confidence probability of the invisible region are fused and analyzed to obtain the grading level result of the tea leaf.

Images