Tea form anomaly detection method and system based on visual fusion

Abstract

The invention relates to the technical field of intelligent detection and quality control, in particular to a tea form anomaly detection method and system based on visual fusion. The method comprises the following steps: obtaining a multi-view form response set and carrying out credibility screening; the method comprises the following steps: integrally dividing tea leaves into form units such as trunk trend, edge curl and local collapse and breakage, extracting key points and response features, and establishing cross-view mapping and consistency constraint to construct a structured model; performing time sequence dynamic tracking on each unit, constructing an evolution trajectory through multi-view trajectory alignment and fusion, analyzing an offset trend, and identifying dynamic modes such as stable, slow offset or abnormal rapid offset; constructing a linkage relation network based on the abnormal units to deduce an evolution trend and adaptively generating an intervention strategy; and performing multi-dimensional quantitative scoring and grade judgment on the units and the samples to realize visual output and closed-loop management of abnormities. According to the invention, accurate, dynamic and explainable detection and early warning of the abnormal tea form are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and quality control technology, specifically to a method and system for detecting abnormal tea morphology based on visual fusion. Background Technology

[0002] With the continuous development of computer vision and artificial intelligence technologies, methods such as multi-view 3D reconstruction, dynamic temporal analysis, and structured perception have shown significant potential in fields such as industrial appearance inspection and precision component measurement, providing powerful technical tools for the refined and automated inspection of complex morphological objects. In the quality inspection of agricultural products, especially the morphological analysis of tea, researchers have tried to introduce image processing and machine learning technologies to improve the objectivity and efficiency of inspection.

[0003] Chinese invention patent CN120374870B discloses a vision-based method for detecting defects in Chinese medicinal materials, comprising the following steps: periodically acquiring time-series images of Chinese medicinal material samples and acquiring multi-view images; performing pixel alignment on the time-series images; structurally organizing and associating camera parameters with the multi-view images according to the shooting orientation to generate a registered time-series image sequence and a multi-view image set; by periodically acquiring time-series images of Chinese medicinal material samples and performing pixel alignment, image offsets caused by ambient light fluctuations and equipment jitter are eliminated, ensuring the spatiotemporal consistency of dynamic change calculations; structurally organizing and associating camera parameters with the multi-view images according to the shooting orientation to establish geometric constraint relationships between viewpoints.

[0004] Automatic identification and evaluation of tea morphology is of positive significance for ensuring product quality and improving the level of intelligent production. Current technological development is evolving towards higher precision and stronger adaptability. Against this backdrop, how to effectively integrate multi-view perception, temporal evolution tracking and intelligent early warning mechanisms to build an anomaly detection system that can be real-time, accurate and predictive has become a research direction worthy of in-depth exploration and with good application prospects, taking into account the characteristics of tea morphology, non-rigid posture and subtle changes in microstructure. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a method and system for detecting abnormal tea morphology based on visual fusion.

[0006] The technical solution of this invention: a method for detecting abnormal morphology of tea leaves based on visual fusion, comprising the following specific implementation steps: S1. Multi-view visual acquisition of tea samples, activation of minute structural differences on the surface of tea by introducing multi-directional light perturbation to form visual morphological response, and integration of the acquired multi-view morphological responses into a multi-view morphological response set after credibility screening. S2. Divide the overall morphology of tea leaves into main trend units, edge curling units, and local collapse and breakage units. Extract the key point coordinates and morphological response intensity of each unit from the multi-view morphological response set, establish cross-view mapping relationship and apply consistency constraints to construct a cross-view structured model. S3. Based on a cross-view structured model, initialize the temporal evolution trajectory for each morphological unit. By aligning and dynamically weighting the trajectories from multiple perspectives, a unified fusion trajectory is formed. Analyze the offset trend of the fusion trajectory and identify stable mode, slow offset mode, and abnormally fast offset mode. S4. Based on the morphological units identified as abnormal rapid offset patterns, construct an abnormal linkage network to deduce the abnormal evolution trend, and adaptively generate unit-level intervention strategies in combination with the preset strategy space. S5. Based on the evolution trajectory and anomaly prediction results of each morphological unit, perform multidimensional feature extraction and quantitative scoring to determine the anomaly level at the unit level and the overall tea sample level, and output visualization results and structured reports.

[0007] Preferably, step S1 specifically includes: At least three non-coplanar cameras are deployed to form multiple observation angles. Each camera is calibrated to establish a unified projection reference, and the three-dimensional coordinates of the tea leaves in physical space are projected onto the two-dimensional detection plane of each angle. An adjustable light source group is arranged from the perspective of each camera. By changing the incident direction vector of the light source, images of tea leaves under different lighting directions are acquired, and the morphological response intensity of each lighting direction relative to the reference lighting is calculated. Calculate the morphological response stability index of each pixel under multiple illumination directions, and select reliable morphological response regions based on the preset stability threshold. By combining the credible morphological response regions from each viewpoint with their corresponding two-dimensional projection coordinates, a multi-view morphological response set is generated.

[0008] Preferably, in step S2, extracting the key point coordinates and morphological response intensity of each unit specifically includes: For each morphological unit, extract the two-dimensional coordinates of several key points used to describe its outline and structural backbone; Obtain the morphological response intensity value corresponding to the coordinates of each key point in the multi-view morphological response set; The coordinates of all key points of each morphological unit and their corresponding morphological response intensity values ​​are combined to form the unit feature set of that morphological unit.

[0009] Preferably, in step S2, establishing the cross-perspective mapping relationship specifically involves: For the set of unit features of the same morphological unit under different observation perspectives, a cross-viewpoint mapping function is defined; By minimizing the joint distance function of the spatial position deviation of corresponding key points and the difference in morphological response intensity between different viewpoints, the cross-viewpoint mapping function is solved to achieve accurate cross-viewpoint matching of morphological units.

[0010] Preferably, step S3 specifically includes: Each morphological unit is assigned a time series index, and its initial spatial coordinates, initial morphological response intensity values, and cross-view consistency index are recorded to initialize the evolution trajectory of the unit. The trajectories of the same morphological unit under different viewpoints are synchronized in time, and spatial alignment is achieved by affine transformation and rigid rotation and translation. The dynamic weight of each viewpoint in trajectory fusion is calculated based on the cross-view consistency index. The dynamic weight is then used to perform weighted fusion of the aligned viewpoint trajectories to generate a unified fused trajectory. The consistency of the fusion result is then verified. Calculate the spatial offset vector and morphological response intensity change of key points in the fused trajectory at adjacent time steps, and combine the two to calculate the composite offset amplitude of the unit. The composite migration amplitude time series of the cells is smoothed, and its mean migration, maximum migration rate and migration acceleration features are extracted. Based on this, the evolution mode of the cells is identified as stable mode, slow migration mode and abnormally fast migration mode.

[0011] Preferably, step S4 specifically includes: Extract all morphological units identified as abnormal fast offset patterns to form an abnormal unit set; Calculate the spatial distance between any two units in the abnormal unit set and the correlation coefficient of their composite offset amplitude time series. Calculate the abnormal linkage weight between the two based on the spatial distance and the correlation coefficient, and construct an abnormal linkage matrix to form an abnormal association network. For each unit in the abnormal unit set, its abnormal neighborhood set is determined according to the abnormal linkage matrix. A dynamic inference model is dynamically established by combining its own and the current offset of the neighboring units, and the offset of the unit in the future multiple time steps is iteratively predicted. A predefined anomaly response strategy space includes various measures such as adjusting acquisition parameters, implementing local interventions, marking high-risk areas, and dynamic tracking and monitoring. Calculate the matching distance between the features of each anomalous unit and each policy in the policy space, determine the adaptation weight of each policy based on the matching distance, and select the policy with the highest adaptation weight as its unit-level intervention policy for each anomalous unit.

[0012] Preferably, step S5 specifically includes: For each morphological unit, several feature indicators are extracted from its evolution trajectory and prediction results, including current offset, maximum offset rate, cumulative offset, dynamic evolution status label, and abnormal linkage intensity. The feature indicators are weighted and normalized to calculate the unit anomaly score for each morphological unit. Based on the abnormal trend factor of the morphological unit and the linkage weight of the neighboring units, the abnormal unit score is corrected, and the corrected score is mapped to a predefined text level. By combining the corrected anomaly scores of all morphological units and the anomaly linkage weights between units, the overall anomaly score of the tea sample is calculated, and the sample is graded according to the preset grade threshold. Generate a structured dataset containing unit-level and sample-level rating levels, anomaly trend predictions, linkage information, and strategy recommendations, and output it in the form of anomaly heatmaps, anomaly linkage network diagrams, and comprehensive reports.

[0013] Preferably, the anomaly heatmap in the visualization output is used to encode the anomaly level and predicted trend of each morphological unit on the spatial distribution map of tea morphology using different colors; Anomaly linkage network diagrams are used to display the propagation paths and linkage weights between abnormal units in the form of nodes and connections.

[0014] Preferably, the formula for calculating the abnormal linkage weight includes both a spatial distance attenuation term and an offset dynamic correlation coefficient term, and the relative importance of the two terms is adjusted by adjusting the spatial correlation weight.

[0015] The technical solution of this invention: A tea morphology anomaly detection system based on visual fusion, used to execute the aforementioned tea morphology anomaly detection method based on visual fusion, characterized in that it includes: The multi-view visual perception and morphological response acquisition module is used to deploy multi-view cameras and controllable light sources to acquire multi-view images of tea leaves, calculate and filter morphological responses, and generate a multi-view morphological response set. The feature extraction and multidimensional representation module is used to divide the tea morphology into main stem trend, edge curling and local collapse and breakage units, extract the key points and response features of each unit, establish cross-view mapping and consistency constraints, and construct a cross-view structured model. The multi-view trajectory alignment and dynamic fusion module is used to initialize the evolution trajectory for each morphological unit, perform spatiotemporal alignment and weighted fusion of multi-view trajectories, analyze the offset trend, and identify dynamic patterns of stable, slow offset, and abnormally fast offset. The abnormal pattern linkage inference and strategy adaptation module is used to build a linkage relationship network between abnormal units, infer the abnormal evolution trend, and adaptively match and generate unit-level intervention strategies from the preset strategy space. The morphological anomaly quantitative scoring and grading module is used to calculate anomaly scores at the unit and sample levels, determine the grading level, and generate visual charts and structured reports.

[0016] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: This invention designs a visual fusion-based method and system for detecting morphological anomalies in tea leaves. By introducing multi-view acquisition and an active illumination perturbation mechanism, it can fully activate the subtle structural differences on the surface of tea leaves, making potential anomalies such as curling, breakage, and collapse visible at the visual level, significantly improving the sensitivity and comprehensiveness of capturing subtle morphological anomalies. Based on cross-view structured modeling and temporal trajectory fusion methods, it achieves fine depiction and dynamic tracking of local morphological features of tea leaves. It can not only accurately describe the static characteristics of morphological units but also analyze their evolutionary trends over time, thereby supporting early identification and predictive warning of abnormal states and enhancing the forward-looking and intelligent level of the detection system. By constructing an anomaly linkage network and a strategy adaptive matching mechanism, the system can identify the spatial propagation relationship of anomalies and generate targeted intervention suggestions, realizing closed-loop management from detection to warning and strategy response, improving the systematicness and operability of the overall detection process. The quantitative scoring and visualization output functions provided by this invention give the detection results good interpretability and decision support capabilities, providing reliable technical support for the automated and precise detection of tea quality. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for detecting abnormal tea morphology based on visual fusion proposed in this invention. Figure 2 This is a system architecture diagram of a tea morphological anomaly detection system based on visual fusion proposed in this invention. Detailed Implementation

[0018] Example 1, as Figure 1 As shown, the present invention proposes a method for detecting anomalies in tea morphology based on visual fusion, and its specific implementation steps are as follows: S1. By actively introducing multi-view and multi-directional lighting perturbations during the acquisition phase, the inherent morphological differences of tea leaves are fully activated at the visual level. Under a unified geometric benchmark, the morphological responses from different viewpoints are structured and organized, ensuring that the acquired image data is not only abundant but also possesses a foundation for fusion in terms of structural consistency and physical orientation. This provides stable, reliable, and discriminative original morphological input for subsequent morphological alignment, evolutionary analysis, and anomaly detection. The specific implementation process is as follows: S11. By setting up multiple non-coplanar observation angles within the detection area and establishing a unified projection benchmark, the morphology of tea leaves under different angles becomes comparable, avoiding direct interference from angle differences in subsequent analysis. Specifically: On the tea conveyor line, drying platform or testing station, the testing area is spatially planned. According to the natural spreading shape and conveying direction of the tea, at least three non-coplanar cameras are set up to form a frontal view, an oblique side view and a close-range macro view respectively. Each camera must be precisely calibrated for tilt during installation. and azimuth This ensures that tea images acquired from different perspectives can be mapped to the same reference coordinate system; Based on this, a morphological observation reference plane is defined for each camera, and the three-dimensional coordinates O(x,y,z) of the tea leaves in physical space are projected onto the two-dimensional detection plane: ; ; in, This represents a comprehensive description of the i-th visual perception perspective; This represents the pitch angle (elevation / depression) of the i-th viewpoint relative to the tea-shaped plane; This represents the azimuth angle of the i-th viewpoint around the vertical axis; This represents the set of three-dimensional coordinate points of tea leaves in physical space. This represents the projection operator, which characterizes the projection of three-dimensional physical points. Projected onto a two-dimensional plane from the i-th viewpoint; This represents the coordinates of the tea leaves projected onto the two-dimensional detection plane from the i-th viewpoint; N represents the number of multi-view cameras. S12. By introducing multi-directional incident light from various observation angles, and utilizing the sensitivity of tea surface structure to changes in light, the visual response differences of morphological features such as curling, breakage, and collapse are actively amplified, making potential anomalies explicit. Specifically: An adjustable light source group is arranged from the perspective of each camera, and the direction vector of the light source is denoted as... By changing the incident angle to irradiate the surface of the tea leaves locally, the outline, wrinkles, broken edges, or local collapse of the tea leaves will produce obvious differences in reflection. The morphological response of tea leaves under different light directions is defined as follows: ; in, This represents the illumination vector at the j-th illumination direction from the i-th viewpoint; The azimuth angle (direction of incidence of light source) is the angle in the horizontal direction. This indicates the angle (elevation / depression) of the incident light source in the vertical direction; This represents the brightness value of the acquired image from the i-th viewpoint and the j-th illumination direction; This represents the brightness value collected under reference lighting conditions; This represents the morphological response intensity under the i-th viewpoint and j-th illumination condition; S13. By evaluating the stability of morphological response under multiple illumination conditions, unreliable responses caused by occasional reflections or noise are filtered out, and only morphological information dominated by the true structure is retained to ensure the physical reliability of the acquisition results. Specifically: Since illumination disturbances can introduce reflections, shadows, or noise, the response data undergoes reliability screening; a response stability index for each pixel under multiple illumination conditions is defined: ; Set a stability threshold based on stability indicators. Filter out the reliable morphological response regions: ; Where M represents the number of lighting directions used in each viewpoint; This represents the average response value under all lighting conditions at the i-th viewpoint; This indicates the stability index of the response; This represents the credible morphological response region from the i-th perspective; S14. Unify and integrate reliable morphological responses from different perspectives into a multi-view morphological response set, ensuring consistency in spatial structure and response intensity of the output results. This provides direct and standardized input for subsequent cross-view fusion and morphological anomaly analysis. Specifically: The filtered reliable response regions from each perspective With projected coordinates Combined, they form a unified set of multi-perspective morphological responses: ; Accordingly: This set unifies the geometric information and activation response information from each viewpoint, providing a complete data foundation for subsequent cross-view morphological unit extraction and structural alignment; in, This represents a set of multi-view morphological responses, including the projected coordinates and corresponding reliable response intensities for all viewpoints. This indicates that a union operation is performed on all viewpoints i, integrating the morphological responses from different viewpoints into the same set. In step S11, N cameras have been defined to ensure that information from multiple angles is included in the analysis.

[0019] S2. By dividing the multi-view morphological response set obtained in step S1 into functional units and establishing key point mapping and consistency constraints across perspectives, a structured cross-view model is constructed to achieve fine analysis of local morphology of tea leaves and sensitive capture of potential anomalies, providing a reliable foundation for subsequent morphological evolution and anomaly judgment. The specific implementation process is as follows: S21. The overall morphology of the tea leaves is divided into main stem trend units, edge curling units, and local collapse / fragmentation units. A set of unit features is formed through key point extraction and response value encoding, enabling independent description of local structures and providing analytical granularity for cross-perspective correspondence. Specifically: For the multi-view morphological response set F output in step S1, the tea leaves as a whole are divided into several logical units to facilitate detailed analysis of each local structure; each morphological unit includes: Main trend unit: used to describe the overall extension direction and curvature of the tea leaves; Edge curling unit: used to depict leaf curling, folding, or tip twisting; Localized collapse / fracture unit: Used to capture areas of slight fracture or localized collapse; For each unit, extract several key points. To describe the unit profile and structural backbone, and to add the response values ​​collected in step S1. Encode: ; Where u and v represent the morphological unit number and the key point number, respectively; The coordinates of the v-th key point within unit u are represented by K; K represents the number of key points in each morphological unit. Represents key points from the i-th perspective. The response value (morphological response intensity); This represents the complete feature set of the u-th morphological unit from the i-th viewpoint, including keypoint coordinates and response values; S22. Establish the correspondence between key points of morphological units from multiple perspectives. By jointly minimizing the mapping function of spatial position deviation and response difference, achieve accurate matching of units across perspectives, ensuring that the same unit maintains structural and response consistency under different perspectives. Specifically: The same morphological unit may exhibit scale differences, rotational offsets, or projection distortions under different viewpoints. To establish a reliable correspondence, a cross-viewpoint mapping function is defined. : ; Keypoint matching is achieved by minimizing the joint distance function of spatial location deviation and response difference: ; in, This represents the mapping function between the i-th and j-th viewpoints of the u-th unit; Weighting coefficients representing the spatial location error and response difference; S23. Perform cross-view consistency quantization on the mapped morphological units, calculate the deviation between the unit spatial coordinates and response characteristics to form a consistency index, which is used to evaluate unit stability and mark potential abnormal regions, thereby improving the sensitivity and reliability of anomaly detection. After completing the unit mapping, the coordinates and responses of each unit under multiple perspectives are quantized for consistency, forming a stability constraint index: ; in, The weighting coefficient representing the difference in response in the consistency constraint; The cross-view consistency index represents the morphological unit u, and the spatial and response stability of the unit under multiple views; S24. Integrate the feature sets, cross-perspective mappings, and consistency indicators of all units to construct a complete cross-perspective structured model. This provides a directly usable data foundation for subsequent morphological evolution trajectory analysis and offset trend judgment, enabling structured and interpretable anomaly analysis support, namely: ; Where S represents the cross-perspective structured model set, which includes the features, mapping relationships and consistency indicators of all units; U represents the total number of morphological units into which the tea plant is divided, that is, how many analysis objects the tea plant is divided into, and each object has its own set of key points, cross-perspective mapping and consistency indicators.

[0020] S3. Based on the cross-perspective structured model output from step S2, time-series dynamic tracking of each unit of tea is performed to construct morphological evolution trajectories and quantify offset trends. This provides a reliable basis for early detection and grade assessment of minor anomalies. The overall process includes trajectory initialization, multi-perspective fusion, offset trend analysis, and dynamic pattern recognition, ensuring continuous, precise, and practical analysis. The specific implementation process is as follows: S31. Assign a time series index to each unit and initialize the evolution trajectory, record the spatial coordinates of key points, response values, and cross-perspective consistency indicators to form a multi-dimensional trajectory vector, providing basic data and initial state for subsequent dynamic analysis, ensuring that the trajectory highly corresponds to the actual tea structure, specifically: Assign a unique time series index t0 to each unit u in the cross-perspective structured model S, and initialize the evolutionary trajectory. This forms a time series framework; Record the initial spatial coordinates of key points in the unit Response value and cross-perspective consistency indicators ; To establish a multi-dimensional trajectory vector for subsequent dynamic tracking: ; Where t0 and t represent time series indices, t0 represents the initial time point, and t represents any subsequent time point; This represents the two-dimensional spatial coordinates of the v-th key point within cell u from the i-th viewpoint. ; This represents the morphological response value of the key point at time t0 and viewpoint i; This represents the cross-view consistency index of a unit, which quantifies the consistency of key point coordinates and responses of the same unit under different viewpoints; The vector representing the evolution trajectory of unit u at time t0; S32. The trajectories of each unit under different perspectives are time-series aligned and dynamically weighted and fused. Weights are adjusted using consistency indices and perspective reliability. Simultaneously, deviations are corrected through local rigidity or affine transformation to form a stable, unified trajectory, providing a precise basis for the analysis of minute changes. Specifically: The trajectory of each unit u under each viewpoint i Preprocessing is performed to obtain the aligned view trajectory. ; The preprocessing process includes, but is not limited to: Key point coordinates Normalization is performed to eliminate scale bias caused by differences in camera position and lens focal length; the response values ​​are then normalized. Illumination normalization and noise filtering are performed to reduce the impact of external environmental changes; the trajectory is synchronized in time to ensure that key points at each viewpoint correspond at the same time step t; initial alignment is achieved using affine transformation or rigid rotation and translation. ; in, The affine transformation matrix representing viewpoint i is used to correct for rotation, scaling, and shearing distortions. This represents the translation vector of viewpoint i, used to correct camera offset. This indicates the coordinates of the key points after initial alignment; After preprocessing (such as affine transformation or rigid rotation and translation alignment) of each view trajectory, the aligned view trajectory is obtained. Calculate dynamic weights To assess its contribution to the integration: ; The aligned trajectories are then fused according to dynamic weights to generate a unified multi-view trajectory: ; ; Forming fusion trajectory vectors : ; Perform consistency checks on the fused trajectories and calculate the average deviation: ; like If the threshold is exceeded, it is marked as a possible abnormal viewpoint or noise interference, and the weights need to be readjusted or local corrections need to be made. in, This represents the dynamic weight of viewpoint i at time t, measuring its contribution to the fusion. This represents the adjustment coefficient, which controls the sensitivity of the consistency index to the weights. The consistency index of unit u at viewpoint i; The coordinates of the key points after fusion are obtained through dynamic weighted multi-view combination; This represents the key point response value after fusion, reflecting the comprehensive multi-view morphological characteristics; This represents the fused unit trajectory vector, which includes fused spatial coordinates and response information; This represents the fusion trajectory verification index, which measures the average deviation between key points from each viewpoint and the fusion trajectory. S33. Calculate the spatial offset and response changes of key points in the calculation unit to form a composite offset index. Analyze the offset trend through time-series moving average or weighted smoothing to quantify the dynamic change rate and direction of the unit, providing quantifiable evidence for early anomaly identification. Specifically: For each keypoint v=1,…,K of cell u, calculate the spatial offset vector between adjacent time steps: ; Key point response values Calculate time series changes: ; By combining the spatial offset of key points with the response offset, the overall offset magnitude of the cell is calculated: ; Perform trend analysis on the cell offset time series to identify the rate of change and anomalous abrupt changes: ; ; ; in, This represents the spatial offset vector of keypoint v between time t and t-1; This indicates the change in the keypoint response value over adjacent time steps; The weighting coefficients representing spatial offset and response offset are used to unify the dimensions and control the impact of response offset on the composite index. The value represents the composite offset magnitude of cell u at time t; W represents the sliding window length, used to calculate the offset moving average. This represents the moving average offset value, reflecting the local average change trend of the element at time t; This represents the cumulative value of the cell offset; This represents the offset value after smoothing the weighted exponent; Represents the exponential smoothing coefficient. ; S34. Based on the offset time series, dynamic evolution pattern recognition is performed, dividing the units into "stable", "slowly offset" or "abnormally fast offset", outputting the evolution trajectory model and state labels, and visualizing them to provide structured data and intuitive reference for subsequent anomaly judgment and level assessment. Specifically: Offset time series for each cell u Extracting dynamic features: Mean offset: ; Maximum offset rate: ; Offset acceleration (abrupt acceleration can identify rapid anomalies): ; Based on the extracted feature vector The dynamic pattern classification procedure is executed using the following strategies: ; For each unit u, output a structured evolution trajectory model: ; Overall system output set: ; An interface for generating anomaly alerts based on trajectory status: triggers warning signals for abnormally fast-moving units; provides trend predictions for slowly-moving units; and can adaptively generate reports that record the location, offset, and evolution trend of abnormal units. in, The time average value of the unit offset represents the average magnitude of the overall morphological evolution; T represents the total length of the time series (total number of sampling frames minus 1); t0 represents the starting time point or the starting sampling frame of the time series; t T This indicates the end time point of the time series or the last sampled frame, which is the last time point of the offset data sequence; This represents the maximum rate of unit displacement, i.e., the fastest morphological change during evolution; It represents the acceleration of the unit offset (the change in the rate of offset change), capturing sudden change trends; The normal threshold representing the mean of cell offset is obtained statistically from training or experimental samples; The normal threshold representing the maximum rate of cell offset is determined by training or experimental statistics; This represents the normal threshold for unit offset acceleration, controlling the sensitivity to sudden changes. The classification results of the evolution state of unit u are represented, including "stable unit", "slowly shifting unit" and "abnormally fast shifting unit"; Represents a structured evolutionary trajectory; This represents the set of structured evolution trajectories of U units in the entire tea sample.

[0021] S4. Based on the structured trajectory model and dynamic evolution state information output in step S3, by constructing abnormal linkage relationships, inferring abnormal development trends, generating adaptive strategies, and visualizing the output, the linkage prediction, risk warning, and dynamic intervention of tea morphological abnormalities are realized. The specific implementation process is as follows: S41. Extract the anomalous fast-migrating units from step S3, construct an anomalous linkage matrix by combining spatial proximity and dynamic correlation of migration, quantify the possibility of anomalous propagation between units, form an anomalous association network, provide a quantitative basis for subsequent evolution trend inference, and realize the identification of potential linkages between microstructural anomalies, specifically: From the set of structured evolutionary trajectories Extracting abnormal units : ; Construct an anomaly linkage matrix L to quantify the anomaly impact relationships between units: ; in, This represents the set of abnormal units, which is the set of all morphological units that were determined to be "abnormal fast offset units" in step S3, and is used to limit the scope of the abnormal linkage analysis. The abnormal linkage weight coefficient represents the linkage strength or influence between unit u and unit v in the abnormal morphological evolution. The larger the value, the more likely the abnormal development of the two is to influence each other. This represents the spatial distance between unit u and unit v, used to characterize whether anomalies have spatial proximity; This represents the spatial scale control parameter, used to adjust the decay rate of spatial distance in abnormal linkage calculations. The smaller the value, the more sensitive it is to very close units only. The time series correlation coefficient between unit u and unit v represents the degree of synchronization between the two in terms of morphological change trends. and These represent the composite offset time series of elements u and v, respectively, combining spatial offset and response offset to reflect the intensity of morphological changes. This represents the spatial correlation weight adjustment coefficient, which is used to balance the relative importance of spatial proximity and dynamic offset correlation in abnormal linkage, and its value ranges from 0 to 1. S42. Based on the abnormal linkage matrix and unit offset dynamics, a unit-level dynamic deduction model is established to iteratively predict future offset sequences, identify high-risk or potentially risky units, and achieve forward-looking early warning of the evolution trend of minute morphological anomalies, providing a scientific basis for strategy generation. Specifically: For each abnormal unit Establish a dynamic simulation model: ; The offset is calculated iteratively for the next H time steps to form the predicted time series: ; Threshold determination and risk labeling: like These are marked as high-risk abnormal units; Units that evolve slowly but show a continuously rising trend are marked as potential risk units, prompting early intervention; in, This represents the offset predicted by cell u in the next time step, used to determine whether the anomaly continues or worsens. The abnormal neighborhood set of unit u is determined by the abnormal linkage matrix L and includes other abnormal units that have a strong linkage relationship with u. This represents the anomaly propagation function, used to describe the influence of the neighboring cell offset on the evolution of the current cell. It can be a linear or nonlinear mapping. This represents a random disturbance term, used to characterize the unavoidable small uncertainties in real-world environments; This represents the critical anomaly threshold. When the predicted offset exceeds this threshold, the cell is considered to have entered a high-risk anomaly state. S43. Define the policy space, calculate policy matching weights by combining anomaly prediction sequences and unit features, adaptively select the optimal policy to generate unit-level intervention schemes, achieve precise, hierarchical, and dynamic anomaly control and intervention, and improve the system's intelligence and feasibility. Specifically: Define the strategy space S, including but not limited to: adjusting the visual acquisition angle or lighting parameters to improve the ability to capture micro-anomalies; local support or intervention operations to prevent the spread of damage; prioritizing the marking of high-risk areas for manual review; and dynamic tracking strategies for anomaly units for cumulative trend monitoring. Calculate the strategy matching weight: ; Choose the strategy with the highest weight. As unit-level strategies, a strategy matrix is ​​formed: ; Where S represents the set of strategy spaces, which contains a variety of predefined exception handling strategies of the system; This represents the adaptation weight of unit u to strategy s. The larger the value, the more suitable the strategy is for the abnormal state of the current unit. It represents the policy matching distance, characterizing the degree of mismatch between policy s and the abnormal features of unit u (offset, rate, acceleration, linkage strength, etc.), and is a metric used to quantify the degree of mismatch between policy s and the abnormal features of unit u. This represents the weight sensitivity coefficient, which is used to adjust the strength of the influence of the policy matching distance on the weight distribution. The larger the value, the more sensitive it is to policy differences. This represents the set of strategy adaptation results, which is the final strategy combination selected for each unit; S44. Visualize the results of anomaly linkage, evolution trend, and strategy adaptation, including anomaly spatial distribution, linkage network, and strategy indications. Simultaneously, generate a structured dataset to support real-time monitoring, trend analysis, and decision execution. Specifically, visualize the spatial distribution of anomaly units, using color coding to indicate anomaly levels and predicted trends; visualize the anomaly linkage network, displaying the propagation paths and weight strengths between units; and visualize the strategy adaptation results, using arrows or labels to annotate unit strategies and priorities. Output a structured data set: ; in, This represents the final system output data set, which includes the evolution trajectory of each unit, anomaly prediction results, and corresponding strategies, for visualization or system calls.

[0022] S5. The tea unit anomaly information, linkage relationships, and strategy adaptation results output in step S4 are subjected to multi-dimensional quantitative processing to achieve unit-level scoring, grade determination, and overall sample anomaly assessment. A closed-loop management system is formed through visualization and report output, providing an operable, traceable, and implementable evaluation system for tea morphological anomaly detection. The specific implementation process is as follows: S51. For each tea leaf unit, extract feature indicators such as offset, maximum offset rate, cumulative offset, dynamic evolution state, and abnormal linkage intensity. Calculate the unit anomaly score using a weighted normalization function to reflect the degree of microstructural anomaly and support trend and local perturbation smoothing. Specifically: For structured trajectory models and predicted offset sequence Extracted feature metric: Current offset Maximum offset rate Cumulative offset Dynamic evolution state Linkage strength ; Define the unit anomaly scoring function: ; in, The unit anomaly score quantifies the degree of morphological abnormality of each tea unit, with a value range of [0,1]. The larger the value, the more obvious the abnormality. , , and The unit score weights represent the contributions of offset, maximum rate, cumulative offset, and linkage strength to the anomaly score, respectively. The weights can be determined through experimental data or historical samples, and the sum of the weights is 1. S52. Convert unit anomaly scores into text-based ratings to achieve understandable and actionable anomaly classification. Simultaneously, dynamically adjust the scores by incorporating trend evolution and neighborhood interaction to make rating determination more forward-looking and sensitive. Specifically: Unit anomaly scoring Corrections are made to consider abnormal trends within consecutive time periods of a given cell and the impact of abnormal linkages between neighboring cells: ; ; According to the revised rating Map units to text levels: ; Among them, T u This represents the trend factor, i.e., the frequency of anomalies occurring in consecutive time steps or the cumulative offset. In this embodiment, it is represented by a normalized representation using the number of historical offsets exceeding the threshold; L u The neighbor linkage weight represents the potential influence of surrounding abnormal units on the unit; N(v) represents the neighborhood of unit v. and This represents an adjustable coefficient that controls the weighting effect of trends and correlations on the score; This indicates the corrected unit anomaly score; S53. All unit scores are weighted and summed using anomaly linkage weights to generate an overall sample score. Sample levels are then categorized based on set thresholds. Simultaneously, intervention suggestions are generated through strategy adaptation, achieving a closed-loop evaluation from microstructure to macroscopic sample, quantifying the impact of micro-anomalies on the overall sample. Specifically: The overall sample score is generated by combining the anomaly score of the integrated unit with the linkage weights. ; Set level thresholds based on overall scores. , and Classify samples into grades: ; in, The overall anomaly score of the sample is represented by |U|, which combines the scores of all units and their interrelationships to reflect the overall degree of anomaly in the tea sample; U represents the set of all morphological units in the tea sample; |U| represents the total number of units. This represents the linkage adjustment coefficient, used to determine the contribution of abnormal linkage of the control unit to the overall score; Table of sample grades; S54. The unit-level scores and grades, as well as the overall sample scores and grades, are output in a structured manner and presented intuitively through anomaly heatmaps, linkage network diagrams, and reports. Simultaneously, strategy suggestions and traceable records are generated, supporting automated system decision-making and manually interpretable analysis, forming a complete closed-loop management system. This includes outputting a unit-level score and grade matrix. Output sample overall score and grade ; Generate a complete report that records unit anomaly scores, levels, trend predictions, linkage information, and strategy recommendations.

[0023] Example 2, as Figure 2 As shown, the present invention proposes a tea morphology anomaly detection system based on visual fusion, which is used to execute a tea morphology anomaly detection method based on visual fusion proposed in Embodiment 1. The system includes: a multi-view visual perception and morphology response acquisition module, a feature extraction and multi-dimensional representation module, a multi-view trajectory alignment and dynamic fusion module, an anomaly pattern linkage deduction and strategy adaptation module, and a morphology anomaly quantitative scoring and level determination module.

[0024] The multi-view visual perception and morphological response acquisition module is used to perform multi-view, high-resolution visual acquisition of tea samples to ensure complete capture of morphological features such as tea leaves, buds, and curling. Through a multi-camera array or a rotatable shooting device, it can acquire images of tea from different perspectives. At the same time, it combines light source adjustment and depth sensing sensors to capture the three-dimensional morphology and minute structural changes of tea leaves. The feature extraction and multidimensional representation module is used to perform unified feature extraction and multidimensional representation of multi-view images and depth information to generate structured data of tea morphology; it integrates geometric morphology features, texture features and dynamic response features, and uses convolutional neural networks and adaptive coding methods to construct feature vectors for each tea unit; at the same time, it extracts the instantaneous offset, cumulative change, local curling and edge morphology information of the unit and generates a unified representation. The multi-view trajectory alignment and dynamic fusion module spatially aligns and fuses time series data from different viewpoints to generate a unified dynamic morphological trajectory model. It uses spatial coordinate calibration and time synchronization mechanisms to match the offset, rate and evolution trend captured by each viewpoint, while considering the linkage relationship of local anomalies and possible occlusion effects to achieve a globally consistent dynamic morphological description. The abnormal pattern linkage inference and strategy adaptation module is used to analyze and infer abnormal patterns based on dynamic trajectory models, identify the evolution trend and potential spread of abnormal tea morphology; based on the linkage strength and evolution rate between units, it infers the overall morphological changes that local abnormalities may cause, and provides response suggestions for abnormal situations in combination with preset strategies or adaptive strategies. The morphological anomaly quantification scoring and grading module integrates unit-level scores, linkage information, and predicted trends to generate anomaly scores and grading for both unit and overall samples, providing visual output and report generation. It performs multi-dimensional scoring on each tea unit, combining trend correction and linkage weighting to quantify anomalies into grades, and then comprehensively generates the overall sample score and grade. The final results are presented intuitively through heatmaps, linkage network diagrams, and overall grade indicators, and can generate automated reports to support operation and decision-making.

[0025] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A tea leaf shape abnormality detection method based on visual fusion, characterized in that, The embodiment comprises the following specific steps: S1, multi-view visual acquisition is performed on the tea leaf sample, the microstructure difference of the tea leaf surface is activated by introducing multi-directional light disturbance to form a visual morphological response, and the obtained multi-view morphological response is integrated into a multi-view morphological response set after credibility screening; S2, the overall morphology of the tea leaf is divided into a trunk trend unit, an edge curling unit and a local collapse and crushing unit, the key point coordinates and morphological response intensities of each unit are extracted from the multi-view morphological response set, a mapping relationship across views is established and consistency constraints are applied, and a cross-view structured model is constructed; S3, based on the cross-view structured model, a time sequence evolution track is initialized for each morphological unit, the tracks under multiple views are aligned and dynamically weighted and fused to form a unified fusion track, the offset trend of the fusion track is analyzed, and stable mode, slow offset mode and abnormal rapid offset mode are identified; S4, based on the morphological unit identified as the abnormal rapid offset mode, an abnormal linkage relationship network is constructed to deduce the abnormal evolution trend, and a unit-level intervention strategy is adaptively generated in combination with a preset strategy space; S5, based on the evolution track and abnormal prediction result of each morphological unit, multi-dimensional feature extraction and quantitative scoring are performed to realize abnormal level judgment of the unit level and the overall level of the tea leaf sample, and visual results and structured reports are output.

2. The tea shape abnormality detection method based on visual fusion according to claim 1, characterized in that, Step S1 specifically comprises: At least three cameras that do not share the same plane are arranged to form multiple observation angles, each camera is calibrated to establish a unified projection reference, and the three-dimensional coordinates of the tea leaf in the physical space are projected onto the two-dimensional detection plane of each view; Adjustable direction light source groups are arranged under each camera view, the incident direction vector of the light source is changed, the images of the tea leaf under different light directions are collected, and the morphological response intensity of each light direction relative to the reference light is calculated; The morphological response stability index of each pixel point under multiple light directions is calculated, and the reliable morphological response area is selected according to the preset stability threshold; The reliable morphological response area under each view is combined with its corresponding two-dimensional projection coordinates to generate a multi-view morphological response set.

3. The tea leaf shape abnormality detection method based on visual fusion according to claim 2, characterized in that, In step S2, the extraction of the key point coordinates and the morphological response intensity of each unit specifically comprises: For each morphological unit, the two-dimensional coordinates of several key points used to describe the outline and structural trunk of the unit are extracted; The morphological response intensity value corresponding to each key point coordinate in the multi-view morphological response set is obtained; All key point coordinates and their corresponding morphological response intensity values of each morphological unit are combined to form a unit feature set of the morphological unit.

4. The tea shape abnormality detection method based on visual fusion according to claim 3, characterized in that, In step S2, the mapping relationship across views is specifically established as follows: For the unit feature sets of the same morphological unit under different observation angles, a cross-view mapping function is defined; By minimizing the joint distance function of the spatial position deviation and the morphological response intensity difference of the corresponding key points between different views, the cross-view mapping function is solved to realize the accurate matching of the morphological unit across views.

5. The tea leaf shape abnormality detection method based on visual fusion according to claim 4, characterized in that, Step S3 specifically comprises: Each morphological unit is given a time sequence index, the initial spatial coordinates of the key points, the initial morphological response intensity value and the cross-view consistency index are recorded, and the evolution track of the unit is initialized; The trajectories of the same morphological units under each view angle are time-synchronized, and affine transformation, rigid rotation and translation are used for spatial alignment; The dynamic weight of each view angle in trajectory fusion is calculated according to the cross-view consistency index, the aligned view angle trajectories are weighted and fused by using the dynamic weight, a unified fused trajectory is generated, and the consistency of the fusion result is verified; The spatial offset vector and the morphological response intensity change of the key points in the adjacent time steps in the fused trajectory are calculated, and the compound offset amplitude of the unit is calculated by combining the two; The time sequence of the compound offset amplitude of the unit is smoothed, the mean offset, the maximum offset rate and the offset acceleration characteristics are extracted, and the evolution mode of the unit is identified as a stable mode, a slow offset mode or an abnormal rapid offset mode.

6. The tea shape abnormality detection method based on visual fusion according to claim 5, characterized in that, Step S4 specifically includes: Extract all morphological units identified as abnormal rapid offset mode to form an abnormal unit set; The spatial distance between any two units in the abnormal unit set and the correlation coefficient of their compound offset amplitude time sequence are calculated, the abnormal linkage weight between the two is calculated based on the spatial distance and the correlation coefficient, an abnormal linkage matrix is constructed to form an abnormal association network; For each unit in the abnormal unit set, determine its abnormal neighborhood set according to the abnormal linkage matrix, establish a dynamics deduction model combining its own and the current offset dynamics of the neighborhood units, and iteratively predict the offset of the unit at multiple future time steps; An abnormal response strategy space including adjusting the acquisition parameters, implementing local intervention, marking high-risk areas and dynamic tracking and monitoring is predefined; The matching distance between the characteristics of each abnormal unit and each strategy in the strategy space is calculated, the adaptation weight of each strategy is determined according to the matching distance, and the strategy with the highest adaptation weight is selected as the unit-level intervention strategy for each abnormal unit.

7. The tea leaf shape abnormality detection method based on visual fusion according to claim 6, characterized in that, Step S5 specifically includes: For each morphological unit, extract several feature indicators such as current offset, maximum offset rate, cumulative offset, dynamic evolution state label and abnormal linkage strength from its evolution trajectory and prediction result; The feature indicators are normalized and weighted, and the unit abnormal score of each morphological unit is calculated; According to the abnormal trend factor of the morphological unit and the linkage weight of the neighborhood units, the unit abnormal score is corrected, and the corrected score is mapped to a predefined word level; The overall abnormal score of the tea sample is calculated by combining the corrected abnormal scores of all morphological units and the abnormal linkage weights between the units, and the sample level is divided according to the preset level threshold; A structured data set containing unit-level and sample-level score levels, abnormal trend prediction, linkage relationship information and strategy suggestions is generated, and is visualized as an abnormal heat map, an abnormal linkage network diagram and a comprehensive report. 8.The tea shape abnormality detection method based on visual fusion according to claim 7, characterized in that, The abnormal heat map in the visual output is used to encode the abnormal levels and prediction trends of each morphological unit using different colors on the tea morphology spatial distribution map; The abnormal linkage network diagram is used to display the propagation path and linkage weight strength between abnormal units in the form of nodes and lines. 9.The tea shape abnormality detection method based on visual fusion according to claim 6, characterized in that, The calculation formula of the abnormal linkage weight contains a spatial distance attenuation term and a bias movement correlation coefficient term, and balances the relative importance of the two terms through a spatial correlation weight adjustment coefficient.

10. A visual fusion based tea leaf shape abnormality detection system for performing the visual fusion based tea leaf shape abnormality detection method according to any one of claims 1-9. The method comprises the following steps: A multi-view visual perception and morphological response acquisition module is used to lay out multi-view cameras and controllable light sources, acquire multi-view images of tea leaves, calculate and filter morphological responses, and generate a multi-view morphological response set; A feature extraction and multi-dimensional representation module is used to divide the tea leaf morphology into a main trend, edge curling, and local collapse and fragmentation units, extract key points and response features of each unit, establish cross-view mapping and consistency constraints, and construct a cross-view structured model; A multi-view trajectory alignment and dynamic fusion module is used to initialize evolution trajectories for each morphological unit, perform spatio-temporal alignment and weighted fusion of multi-view trajectories, analyze deviation trends, and identify stable, slow deviation, and abnormal rapid deviation dynamic patterns; An abnormal pattern linkage deduction and strategy adaptation module is used to construct a linkage relationship network between abnormal units, deduce abnormal evolution trends, and adaptively match unit-level intervention strategies from a pre-set strategy space; A morphological abnormality quantitative scoring and grade determination module is used to calculate unit-level and sample-level abnormal scores, determine grades, and generate visual charts and structured reports.

Images