Fresh tea leaf quality rapid detection method based on machine vision

Abstract

The invention relates to the technical field of intelligent detection of tea quality, and discloses a method for rapidly detecting the quality of fresh tea leaves based on machine vision. The method comprises the following steps: acquiring a fresh tea leaf image through multispectral imaging and decoupling the fresh tea leaf image in different channels; and simulating a quality decay path by using the characteristic evolution network, generating a prediction vector, and delimiting a suspected deterioration region. A dynamic sampling window is deployed in the region to obtain a local image micro sequence, change rate features are extracted through time domain feature solidification, cross-domain grafting is carried out on the change rate features and static spectral features, and a full-dimension quality description matrix is formed. And determining a quality phase by querying the quality degradation track mapping table, and performing reverse traceability analysis to generate a structured detection report. According to the method, early warning of the dynamic decay trend of the quality is realized, and the accuracy and foresight of quality judgment are improved by fusing static and dynamic characteristics.

Description

Technical Field

[0001] This invention relates to the field of intelligent tea quality detection technology, specifically a rapid tea leaf quality detection method based on machine vision. Background Technology

[0002] Currently, rapid quality assessment of fresh tea leaves primarily relies on machine vision technology. Existing methods generally employ hyperspectral or multispectral imaging to acquire image information of the tea leaf surface. By extracting static visual features such as color, texture, and shape, a correlation model is established between these features and expert sensory scores or certain physicochemical indicators to achieve quality grading or judgment of superiority or inferiority. This type of technology treats the quality of fresh leaves as a static representation, and its analysis depends entirely on the image information captured in a single sampling.

[0003] Existing technical solutions have shortcomings. Static image features can only reflect the apparent state at the moment of sampling, failing to capture the dynamic changes in quality that are taking place, let alone predict subsequent deterioration trends. This results in insufficient sensitivity for detecting fresh leaves in the early stages of quality decay but with inconspicuous apparent changes, easily leading to misjudgments. Furthermore, the features extracted by conventional methods are isolated, lacking a mechanism to deeply correlate and fuse the initial static attributes of the leaf with the subsequent local dynamic change rate. The constructed quality model has a single dimension and limited explanatory power for complex deterioration patterns.

[0004] A technology is needed that can overcome the limitations of static analysis, enabling the simulation and deduction of the natural degradation path of tea quality after acquiring initial images, thereby providing early warnings of potential deterioration. Furthermore, a method is needed to deeply integrate static appearance features with local dynamic temporal change features to form a more comprehensive and sensitive descriptive system of quality status, achieving accurate insight and forward-looking judgment on the dynamic changes in the quality of fresh tea leaves. Summary of the Invention

[0005] The purpose of this invention is to provide a rapid detection method for the quality of fresh tea leaves based on machine vision, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a rapid quality detection method for fresh tea leaves based on machine vision, the method comprising: The multispectral imaging device is activated to acquire an initial image sequence of fresh tea leaves under a predetermined wavelength combination. The initial image sequence was decoupled by channel to extract independent spectral image layers related to the color of tea surface, distribution of water stains on leaf surface and tissue structure. Independent spectral image layers are input into a pre-structured feature evolution network, which generates a quality decay prediction vector by simulating the natural decay path of tea quality. Based on the quality degradation prediction vector, suspected deterioration focus areas are delineated in independent spectral image layers; Dynamic sampling windows were deployed around suspected deterioration focal areas to collect local image micro-sequences that evolved over time; Temporal feature freezing operation is performed on local image micro-sequences to extract a set of freezing features that characterize the rate of quality change; By cross-domain feature grafting of the solidification feature set and the static features of the independent spectral image layer, a full-dimensional quality description matrix is ​​formed. Based on the coordinates of the full-dimensional quality description matrix in the quality space, query the preset quality degradation trajectory mapping table to determine the current quality phase of the fresh tea leaves. Based on the quality phase initiation, the reverse quality tracing analysis includes tracing back along the quality degradation trajectory mapping table to the previous stable quality node, and calculating the feature deviation between the current state and the previous stable quality node. A structured inspection report is generated based on the feature deviation, including quality correction suggestions and the urgency of the handling.

[0007] Preferably, the initial image sequence is decoupled by channel to extract independent spectral image layers related to the tea surface color, leaf surface water stain distribution, and tissue structure, including: From the predetermined wavelength combination of the multispectral imaging device, key wavelengths corresponding to chlorophyll reflectance peaks, cell water absorption valleys, and leaf fiber scattering characteristics were identified. For each key wavelength, the corresponding single-band image is separated from the initial image sequence; Background matrix culling is performed on each single-band image. Background matrix culling refers to removing imaging background noise that is unrelated to the tea subject by using image morphological opening operations. After background matrix removal, the single-band images are classified into color image layer, water stain distribution image layer and tissue structure image layer according to the physical properties they represent, which together constitute an independent spectral image layer.

[0008] Preferably, independent spectral image layers are input into a pre-structured feature evolution network, which generates a quality degradation prediction vector by simulating the natural degradation path of tea quality, including: The feature evolution network contains multiple cascaded simulated decay units, each of which corresponds to a quality state at a time scale. The color image layer, water stain distribution image layer, and tissue structure image layer are fed into the initial simulated decay unit of the feature evolution network as parallel inputs. The feature evolution network iteratively calculates the degree of color decay, water stain diffusion, and loose tissue structure in each simulated decay unit according to the preset time scale order and based on the output of the previous unit and the preset decay rules. After processing by all simulated decay units, a multidimensional vector is output from the end of the network. This multidimensional vector encodes the predicted quality state of fresh tea leaves at multiple future time points, which is the quality decay prediction vector.

[0009] Preferably, based on the quality degradation prediction vector, suspected degradation focus regions are delineated in independent spectral image layers, including: Analyze the quality degradation prediction vector and identify the vector components that predict the degree of quality degradation exceeding the threshold. The vector components are mapped back to their spatial locations in the corresponding color image layer, water stain distribution image layer, or tissue structure image layer. In each independent spectral image layer, a connected region is adaptively expanded, centered on the mapped spatial location and combined with the image gradient field of the independent spectral image layer. By merging the connected regions formed by the expansion in all independent spectral image layers, a comprehensive suspected deterioration focal region is obtained after taking the union.

[0010] Preferably, a dynamic sampling window is deployed around the suspected deterioration focus area to collect local image micro-sequences that evolve over time, including: Using the geometric center of the suspected deterioration focal area as the origin, multiple image acquisition points are deployed at a preset distance around it in a radial, equiangular distribution pattern. A miniature imaging unit that can be periodically triggered is fixed at each image acquisition point; All miniature imaging units are controlled to synchronously capture images at uniform time intervals for several cycles, thereby obtaining a series of temporally continuous local images at each acquisition point; A series of local images at the same acquisition point are arranged in chronological order to form a local image micro-sequence of the acquisition point; Local image micro-sequences from all image acquisition points are collected and used as input for subsequent analysis.

[0011] Preferably, a temporal feature freezing operation is performed on the local image micro-sequence to extract a set of frozen features characterizing the rate of quality change, including: For each local image micro-sequence, calculate the pixel value difference matrix between images at adjacent time points; By performing an integral operation on the pixel value difference matrix over the time axis, a cumulative change map representing the cumulative change intensity of the image acquisition points is obtained; Extract pixels whose change intensity exceeds the active threshold from the cumulative change map, and record the coordinates, intensity, and change direction information of the pixels; Perform the above operation on all local image microsequences, summarize the information of all recorded points to form a solidification feature set. The solidification feature set essentially describes the local location and dynamics of significant quality changes around the suspected deterioration focus area within the observation time window.

[0012] Preferably, the solidification feature set is cross-domain feature grafted with the static features of independent spectral image layers to form a full-dimensional quality description matrix, including: From the independent spectral image layer, the static feature values ​​of each pixel in the suspected deterioration focal region in three dimensions of color, water stains and tissue structure are extracted to form a static feature vector field; Each feature point in the solidification feature set is located at its corresponding position in the static feature vector field based on its spatial coordinates. The dynamic change information of the feature points, including the intensity and direction of the change, is used as a new dimension and concatenated with the static feature vector at the corresponding position. After all matching positions are spliced ​​together, a data table is formed in which rows represent spatial locations and columns represent static and dynamic feature dimensions, which is the full-dimensional quality description matrix.

[0013] Preferably, based on the coordinates of the full-dimensional quality description matrix in the quality space, a preset quality degradation trajectory mapping table is queried to determine the current quality phase of the fresh tea leaves, including: The quality space is a multi-dimensional abstract space, where each dimension corresponds to a feature column of the full-dimensional quality description matrix. Aggregate the spatial positions of all feature columns of the full-dimensional quality description matrix and calculate the coordinates of its centroid in the quality space. The quality degradation trajectory mapping table is a pre-built database that records the typical coordinate ranges and their changing trajectories in the quality space for different quality stages from fresh to spoilage. The calculated centroid coordinates are matched with the typical coordinate range in the quality degradation trajectory mapping table to determine the quality stage to which the current fresh tea leaves belong. This quality stage is the quality phase.

[0014] Preferably, the reverse quality tracing analysis is initiated based on the quality phase. This reverse quality tracing analysis includes tracing back along the quality degradation trajectory mapping table to the previous stable quality node and calculating the characteristic deviation between the current state and the previous stable quality node, including: In the quality degradation trajectory mapping table, starting from the coordinate point corresponding to the current quality phase, the tracking proceeds in the opposite direction of the quality degradation trajectory. The node on the trajectory that has the most gradual feature change and is closest to the current coordinate point is defined as the previous stable quality node. Extract the feature vectors corresponding to the current quality phase and the previous stable quality node in the full-dimensional quality description matrix, respectively; Calculate the Euclidean distance between the two feature vectors, and then normalize the Euclidean distance to obtain the feature deviation.

[0015] Preferably, a structured inspection report containing quality correction suggestions and processing urgency is generated based on the feature deviation, including: Establish a lookup table to map the feature deviation to the processing strategy; The calculated feature deviation values ​​are input into a lookup table, which maps to a set of suggested quality correction measures, namely the quality correction recommendations. Meanwhile, based on the magnitude of the feature deviation and the storage time of the fresh tea leaves, a quantitative score of the urgency of the processing is calculated through a linear weighted model. The identification information of fresh tea leaves, the detected quality phases, characteristic deviations, quality correction suggestions, and urgency scores are organized and packaged according to a predetermined format to output a complete structured test report.

[0016] Compared with the prior art, the beneficial effects of the present invention are: A feature evolution network is employed to simulate the natural degradation path of tea quality and generate prediction vectors. This network learns from a large amount of time-series data on normal quality degradation, internally constructing a dynamic model that maps current multispectral features to future state changes. The quality degradation prediction vector output by the network is not a simple classification label, but a latent space encoding indicating the direction and intensity of future quality changes. This allows the system to move beyond simply identifying the current apparent state and to deduce the most likely subsequent quality evolution trend based on inherent patterns. By shifting the detection perspective from "current state judgment" to "trend warning," it can identify potential focal areas where deterioration may occur first, even before obvious anomalies are detected by the naked eye or traditional static features, thus achieving early perception of quality risks.

[0017] After delineating the suspected area, a dynamic sampling window is deployed to collect local image micro-sequences, and a temporal feature freezing operation is performed on them to extract a set of frozen features representing the rate of quality change. These frozen features are then cross-domain grafted with the initial static spectral features. The temporal feature freezing operation essentially compresses and enhances the features of the local micro-sequences in the time dimension, extracting dynamic indicators such as color change gradients and texture variation rates. Cross-domain feature grafting organically integrates the dynamic rate features representing the "speed of change" with the static original features representing "how it originally was" at the feature level. This constructs a full-dimensional quality description matrix that simultaneously contains original state information and dynamic evolution information. This enriches the dimensions and connotation of the quality description, ensuring that the final quality judgment is based not only on "what" the object is, but also on "how it is changing." This improves the detection accuracy and reliability of complex deterioration processes, especially those with slight initial appearances but rapid internal changes, making the system's quality assessment more in-depth and predictive. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the rapid quality detection method for fresh tea leaves based on machine vision as described in this invention. Figure 2 A flowchart of the channel decoupling steps; Figure 3 A flowchart for generating quality degradation prediction vectors for a feature evolution network; Figure 4 A multi-dimensional statistical comparison chart of coagulation feature sets in the quality detection of fresh tea leaves; Figure 5 A graph showing the relationship between the urgency score and storage time in the quality testing of fresh tea leaves. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1This invention provides a rapid quality detection method for fresh tea leaves based on machine vision. The method includes: activating a multispectral imaging device to acquire an initial image sequence of fresh tea leaves under a predetermined wavelength combination; subsequently, performing a channel-by-channel decoupling operation on the initial image sequence to extract independent spectral image layers related to the surface color, leaf surface water stain distribution, and tissue structure of the tea leaves. The obtained independent spectral image layers are input into a pre-structured feature evolution network, which generates a quality decay prediction vector by simulating the natural decay path of tea quality. Based on this quality decay prediction vector, suspected deterioration focus regions are delineated in the independent spectral image layers. Dynamic sampling windows are deployed around the delineated suspected deterioration focus regions to collect local image micro-sequences that evolve over time. Temporal feature solidification is performed on the collected local image micro-sequences to extract a solidified feature set characterizing the rate of quality change. Then, the solidified feature set is cross-domain feature grafted with the static features of the independent spectral image layers to form a full-dimensional quality description matrix. Based on the coordinates of the full-dimensional quality description matrix in the quality space, a pre-defined quality degradation trajectory mapping table is queried to determine the current quality phase of the fresh tea leaves. Based on this determined quality phase, a corresponding reverse quality tracing analysis is initiated. This analysis includes tracing back along the quality degradation trajectory mapping table to the previous stable quality node and calculating the feature deviation between the current state and the previous stable quality node. Finally, a structured inspection report containing quality correction suggestions and the urgency of the action is generated based on the calculated feature deviation.

[0021] Example 1: See Figure 2 From a predetermined wavelength combination of a multispectral imaging device, key wavelengths corresponding to chlorophyll reflectance peaks, cell water absorption valleys, and leaf fiber scattering characteristics are identified. For each identified key wavelength, a corresponding single-band image is separated from the initial image sequence. Background matrix removal is performed on each single-band image; background matrix removal refers to removing imaging background noise unrelated to the tea leaf subject using image morphological opening operations. The single-band images after background matrix removal are categorized into color image layer, water stain distribution image layer, and tissue structure image layer according to the physical properties they represent. These image layers together constitute independent spectral image layers. In specific implementation, the channel decoupling step in the machine vision-based rapid detection method for fresh tea leaf quality is implemented, involving the accurate separation of image information corresponding to key physical properties from the raw data acquired by the multispectral imaging device. The multispectral imaging device is configured to capture surface images of fresh tea leaves in multiple discrete narrow bands. These bands together constitute the initial image sequence, and each image frame of the initial image sequence corresponds to a center wavelength and bandwidth.

[0022] In some embodiments, identifying key wavelengths from a predetermined wavelength combination of the multispectral imaging device is the initial operation for channel decoupling. The predetermined wavelength combination is pre-designed and includes a series of bands covering the visible and near-infrared regions. Key wavelength identification is based on the inherent spectral characteristics of fresh tea leaf tissue. Specifically, the key wavelength corresponding to the chlorophyll reflectance peak is located around 680 nm, the key wavelength corresponding to the cell water absorption valley is located around 970 nm, and the key wavelength corresponding to the leaf fiber scattering characteristics is located around 1300 nm. Using spectral analysis software or a pre-defined wavelength index table, the image data indexes corresponding to these three characteristic wavelengths are accurately located and extracted from the band list output by the multispectral imaging device. For each identified key wavelength, the system separates the corresponding single-band image from the stored initial image sequence. For example, when the key wavelengths are identified as 680 nm, 970 nm, and 1300 nm, the system will extract single-band images with a center wavelength of 680 nm, 970 nm, and 1300 nm from the initial image sequence, respectively. Each single-band image is a two-dimensional grayscale matrix, and each pixel value in the matrix reflects the intensity of reflection or absorption of the corresponding wavelength on the surface of fresh tea leaves.

[0023] For each separated single-band image, the system performs a background matrix culling operation. This operation utilizes the opening operation from mathematical morphology to process the single-band image. The opening operation effectively removes imaging background noise unrelated to the tea leaf subject, such as reflections from the tray or ambient stray light. The opening operation is defined as a process of erosion followed by dilation, and its mathematical expression is:

[0024] in: Represents the input single-band image pixel matrix. Represents a predefined structuring element. The shape and size are set according to the typical width of the tea leaf edge, symbol Represents morphological erosion operation, symbol This represents the morphological dilation operation, while This refers to a clean single-band image after background matrix removal.

[0025] In practice, classifying the single-band images after background matrix removal into independent image layers is the final step in constructing independent spectral image layers. The system automatically classifies images based on the physical properties they represent. Single-band images with a center wavelength of 680 nm primarily reflect chlorophyll content and distribution and are classified into the color image layer; single-band images with a center wavelength of 970 nm are sensitive to moisture and are classified into the water stain distribution image layer; single-band images with a center wavelength of 1300 nm can penetrate the surface to reflect internal structures and are classified into the tissue structure image layer. In some embodiments, the color image layer, water stain distribution image layer, and tissue structure image layer are stored in the system as three independent, spatially aligned two-dimensional data matrices. These three data matrices together constitute the independent spectral image layers required for subsequent analysis.

[0026] It can be understood that each pixel in the independent spectral image layer is associated with a set of three-dimensional physical attribute values, which are derived from the grayscale values ​​of the color image layer, water stain distribution image layer, and tissue structure image layer at the same location. This data organization provides a structured input for subsequent feature extraction and fusion. Through the above-mentioned channel decoupling process, the original multispectral image data is transformed into independent spectral image layer data that directly corresponds to the key quality factors of fresh tea leaves and has removed background interference.

[0027] Example 2: See Figure 3 The feature evolution network comprises multiple cascaded simulated decay units, each corresponding to a quality state at a given time scale. Color image layer, water stain distribution image layer, and tissue structure image layer are fed into the initial simulated decay unit of the feature evolution network as parallel inputs. Following a preset time scale order, within each simulated decay unit, the feature evolution network iteratively calculates the degree of color decay, water stain diffusion, and tissue structure loosening based on the output of the previous unit and preset decay rules. After processing by all simulated decay units, a multi-dimensional vector is output from the network's end. This multi-dimensional vector encodes the predicted quality state of fresh tea leaves at multiple future time points, i.e., the quality decay prediction vector. The quality decay prediction vector is analyzed to identify vector components whose predicted quality state deterioration exceeds a threshold. These vector components are mapped back to their spatial locations in the corresponding color image layer, water stain distribution image layer, or tissue structure image layer. In each independent spectral image layer, a connected region is adaptively expanded around the mapped spatial location, combined with the image gradient field of that independent spectral image layer. By merging the connected regions formed by the expansion in all independent spectral image layers, a comprehensive suspected deterioration focal region is obtained after taking the union.

[0028] In practical implementation, the steps of generating quality decay prediction vectors and delineating suspected deterioration focus areas in the rapid quality detection method for fresh tea leaves based on machine vision involve a computational model simulating a biochemical process and spatial analysis of the prediction results. A pre-structured feature evolution network is a core computational module, designed based on the general patterns of color, moisture content, and cell structure changes in fresh tea leaves over time after harvesting. In some embodiments, the feature evolution network includes multiple cascaded simulated decay units, the number of which strictly corresponds to the number of preset time scales. For example, if it is necessary to predict the quality status at three future time points, the feature evolution network includes three cascaded simulated decay units, with the output of each simulated decay unit representing the predicted state at a specific future moment. The system feeds the color image layer, water stain distribution image layer, and tissue structure image layer from independent spectral image layers as three sets of parallel input data into the initial simulated decay unit of the feature evolution network. The data from each image layer is input in the form of a two-dimensional matrix.

[0029] The feature evolution network operates according to a preset time scale from the current time to future times, performing iterative calculations within each simulated decay unit. The calculations are based on the output of the previous simulated decay unit and decay rules embedded in the network parameters. These decay rules are defined as weight matrices and nonlinear transformation functions. For data in the color image layer, the decay rule simulates the reflectance change caused by chlorophyll degradation; for data in the water stain distribution image layer, the decay rule simulates the water evaporation and diffusion process; for data in the tissue structure image layer, the decay rule simulates the changes in scattering properties caused by cell wall relaxation. In specific implementations, the calculation within the t-th simulated decay unit can be represented as a feature transformation, which takes the form of a linear weighted combination and activation.

[0030] in: The feature map representing the output of the (t-1)th simulated decay unit in the initial unit. This refers to the tensor obtained by stitching together the independent spectral image layers from the input. It is a learnable weight matrix whose parameters encode the quality decay rule from time t-1 to time t. It is the corresponding bias vector. This represents a nonlinear activation function used to introduce nonlinearity into the model, while This refers to the predicted quality feature map output by the t-th simulated decay unit, corresponding to the t-th time scale in the future.

[0031] After processing by all cascaded simulated decay units in the feature evolution network, a multidimensional vector is output from the end of the network. This multidimensional vector is then processed by global pooling from the feature map output by the last simulated decay unit. The multidimensional vector extracted is the quality degradation prediction vector, which compactly encodes the predicted quality state of fresh tea leaves at multiple consecutive time points in the future. In specific implementation, delineating the suspected deterioration focus area based on the quality degradation prediction vector is a spatial localization process. The system analyzes each component of the quality degradation prediction vector and identifies the vector components whose predicted quality state deterioration exceeds a preset threshold. Each vector component is associated with a quality attribute and a prediction time point; for example, a component may correspond to "the looseness of the tissue structure at the second prediction time point." The identified out-of-range vector components are back-mapped back to their original spatial locations in the corresponding color image layer, water stain distribution image layer, or tissue structure image layer through the spatial mapping relationship recorded within the feature evolution network. The mapping relationship is determined by the receptive field of the convolutional kernel in the network.

[0032] In each mapped independent spectral image layer, the system adaptively expands to form a connected region, using the mapped spatial coordinates as the center point and incorporating the image gradient field information of that independent spectral image layer. The expansion logic involves growing the region along the direction of weaker gradients, as weak gradients represent uniform properties, and suspected degradation may occur within this uniform region. The region growth stops when an edge with a gradient value greater than a preset threshold is encountered. This step can be understood as being executed in parallel within each independent spectral image layer, potentially yielding one region in the color image layer, another in the water stain distribution image layer, and a third in the tissue structure image layer. Optionally, the system ultimately merges the connected regions formed by the above method in all independent spectral image layers, taking the union of all regions to obtain a comprehensive spatial range covering all suspected degradation properties. This spatial range is marked as the suspected degradation focal region, which is a binary mask image used to guide subsequent dynamic sampling.

[0033] Example 3: Using the geometric center of the suspected deterioration focal area as the origin, multiple image acquisition points are deployed at predetermined distances around the perimeter in a radial, equiangular distribution pattern. A periodically triggerable miniature imaging unit is fixed at each image acquisition point. All miniature imaging units are controlled to synchronously capture images at uniform time intervals for several cycles, obtaining a series of temporally continuous local images at each acquisition point. The series of local images at the same acquisition point are arranged chronologically to form a local image micro-sequence for that acquisition point. The local image micro-sequences of all image acquisition points are collected as input for subsequent analysis. For each local image micro-sequence, the pixel value difference matrix between images at adjacent time points is calculated. The pixel value difference matrix is ​​integrated over time to obtain a cumulative change map characterizing the cumulative change intensity of that image acquisition point. Pixels with change intensity exceeding an active threshold are extracted from the cumulative change map, and their coordinates, intensity, and direction of change are recorded. Perform the above operation on all local image microsequences, summarize the information of all recorded points to form a solidification feature set. The solidification feature set essentially describes the local location and dynamics of significant quality changes around the suspected deterioration focus area within the observation time window.

[0034] In practical implementation, deploying dynamic sampling windows to acquire local image micro-sequences involves spatiotemporal high-resolution monitoring of the area surrounding suspected deterioration focal points. The system uses the geometric center coordinates of the suspected deterioration focal point as the deployment origin. Multiple image acquisition points are evenly distributed radially at equal angles within a preset radius. For example, eight image acquisition points can be set, with an azimuth interval of 45 degrees between adjacent points. The spatial position of each image acquisition point is defined using polar coordinates. At each deployed image acquisition point, a periodically triggerable micro-imaging unit is fixedly installed. This micro-imaging unit has a fixed focal length and field of view, covering a small local area centered on the image acquisition point. The system sends synchronous trigger commands to all micro-imaging units via a central controller, controlling them to synchronously capture images at uniformly set time intervals. Image capture continues for several preset cycles, thereby obtaining a series of continuously arranged local images in the time dimension at each image acquisition point. A series of local images acquired at the same image acquisition point are arranged strictly according to their capture timestamp order to form a local image micro-sequence for that image acquisition point. This local image micro-sequence is a three-dimensional data block, with two dimensions representing the spatial size of the image and the third dimension representing time. The local image micro-sequences from all image acquisition points are then collected and used as input data for subsequent temporal feature freezing operations.

[0035] The temporal feature freezing operation on local image micro-sequences aims to extract dynamic features characterizing the rate of quality change. The system performs calculations for each input local image micro-sequence. For a local image micro-sequence containing T time-point images obtained from an image acquisition point, the pixel value difference matrix between adjacent time-point images is calculated. This difference matrix reflects the spectral or brightness change at each spatial location within the time interval. In some embodiments, the core of the temporal feature freezing operation is calculating the cumulative change intensity at each spatial location over the entire observation time window. This calculation is achieved by integrating the pixel value difference matrix along the time axis. The mathematical expression of the integration operation is as follows:

[0036] in: Represents the spatial coordinates of a pixel within a local image micro-sequence. This represents a time point index, with values ​​ranging from 1 to T-1. Represents a point in time Time is located at coordinates Pixel intensity value at that location This indicates the absolute value operation. It is the calculated coordinates The cumulative change intensity value at a given location is calculated, and this calculation is performed on all pixels to obtain a cumulative change map characterizing the cumulative change intensity of that image acquisition point throughout the entire observation period.

[0037] From the calculated cumulative change map, the system extracts pixels whose cumulative change intensity exceeds a preset active threshold, which is a threshold value set based on historical data or statistical characteristics. For each extracted pixel, the system records its spatial coordinates. Its corresponding cumulative change intensity value The system also calculates and records the main direction of change of the pixel over time, which can be determined by analyzing the sign trend of the difference matrix between adjacent frames. Optionally, the system repeats the above calculation, extraction, and recording operations on the local image micro-sequences corresponding to all image acquisition points, summarizing all feature point information extracted from one image acquisition point with feature point information extracted from all other image acquisition points to form a complete solidification feature set. It can be understood that the solidification feature set is essentially a structured list or database, where each record describes a specific spatial location within the observation time window, around a suspected deterioration focus area, where a significant quality change occurs, and the dynamics of the intensity and direction of the change observed at that location.

[0038] In practice, the processing of local image micro-sequences is continuous and automated. The system immediately initiates a temporal feature solidification operation after acquiring a sufficient number of image cycles, and the generation of the solidified feature set does not depend on operator intervention. In some embodiments, the sampling time interval of the micro-imaging units can be adjusted according to the tea variety or environmental conditions; shorter intervals are used to capture rapid changes, while longer intervals are suitable for monitoring slow decay. It can be understood that through this deployment and solidification method, the system can capture microscopic, time-evolving change signals from the seemingly static surface of fresh tea leaves. These signals are key to characterizing the rate of quality change. Optionally, the acquisition of local image micro-sequences can be performed using imaging equipment with different but compatible spectral response ranges than that used for acquiring independent spectral image layers, ensuring that dynamic and static features are physically comparable.

[0039] See Figure 4 This is a multi-dimensional statistical comparison chart of the coagulation feature set in the quality testing of fresh tea leaves. The core analysis focuses on the correlation between the number of feature points and the average intensity of change at different sampling points. Sampling points 4 and 5 show a significantly higher number of feature points than other sampling points (over 400), indicating more dense microscopic changes in the areas corresponding to these two sampling points. The number of feature points at sampling point 8 drops sharply, possibly corresponding to an area with a lower degree of deterioration. A continuous upward trend is observed, especially with accelerated growth at sampling points 6-8, indicating that the average degree of deterioration gradually increases from sampling points 1 to 8. The combination of "local peaks + overall increase" in the number of feature points and the average intensity of change reflects the synergistic characteristics of "range expansion" and "deepening of degree" in the deteriorated areas; some areas initially show dense microscopic changes, while the degree of change in subsequent areas continues to increase. This type of chart is used in the feature analysis stage of fresh tea leaf quality testing. It can quickly locate the sampling points with the most significant deterioration (such as sampling points 5 and 8), providing crucial microscopic change evidence for subsequent quality phase determination and traceability analysis, and assisting in determining the core area and development trend of deterioration in fresh tea leaves.

[0040] Example 4: From the independent spectral image layer, extract the static feature values ​​of each pixel in the suspected deterioration focus area in three dimensions: color, water stains, and tissue structure, to construct a static feature vector field. Position each feature point in the solidification feature set to its corresponding position in the static feature vector field based on its spatial coordinates. Incorporate the dynamic change information of the feature points, including the intensity and direction of change, as a new dimension and concatenate it with the corresponding static feature vector. After concatenating all matching positions, a data table is formed where rows represent spatial positions and columns represent static and dynamic feature dimensions; this is the full-dimensional quality description matrix. The quality space is a multi-dimensional abstract space, with each dimension corresponding to a feature column of the full-dimensional quality description matrix. Aggregate the spatial positions of all feature columns of the full-dimensional quality description matrix and calculate the coordinates of their centroids in the quality space. The quality degradation trajectory mapping table is a pre-established database that records the typical coordinate ranges and change trajectories of different quality stages from fresh to spoilage in the quality space. The calculated centroid coordinates are matched with the typical coordinate range in the quality degradation trajectory mapping table to determine the quality stage to which the current fresh tea leaves belong. This quality stage is the quality phase.

[0041] In practical implementation, the steps of cross-domain feature grafting and quality phase determination in the rapid quality detection method for fresh tea leaves based on machine vision involve aligning and fusing dynamic and static features spatially, and mapping the fused features to an abstract metric space. The system first extracts static feature values ​​from independent spectral image layers, which include color image layers, water stain distribution image layers, and tissue structure image layers. The data from these image layers are spatially aligned with the binary mask image of the suspected deterioration focus area. In practice, the system iterates through every pixel marked as "suspected deterioration" within the suspected deterioration focal area. For each such pixel, it reads its grayscale value in the color image layer, the grayscale value in the water stain distribution image layer, and the grayscale value in the tissue structure image layer. These three grayscale values ​​together constitute a three-dimensional static feature vector, representing the static attributes of the spatial location in the three dimensions of color, water stain, and tissue structure. After performing this operation on all pixels located within the suspected deterioration focal area, a static feature vector field covering the entire area is formed. The static feature vector field is a three-dimensional array, where two dimensions are spatial coordinates and the third dimension is the feature dimension.

[0042] The system then spatially matches each feature point in the solidification feature set with the static feature vector field. Each feature point in the solidification feature set contains its precise spatial coordinate information. Based on the spatial coordinates of the feature point, the system locates its corresponding spatial position in the static feature vector field. If the coordinates match perfectly, the system locates it directly; otherwise, it uses nearest neighbor interpolation to determine the corresponding position in the static feature vector field. After successful localization, the system uses the dynamic change information carried by the feature point as a new feature dimension. The dynamic change information includes at least the intensity and direction of change. This dimension is then concatenated with the three-dimensional static feature vector of the corresponding position extracted from the static feature vector field. The concatenation operation generates an extended feature vector, which simultaneously contains the static attributes and the observed dynamic change attributes of that position. For example, the extended feature vector of a feature point might be in the form of [color value, water stain value, tissue structure value, change intensity, change direction]. In practice, the system performs the aforementioned localization and splicing operations on feature points in all solidification feature sets, organizes all generated extended feature vectors according to their corresponding spatial location indices, and finally forms a structured data table. This data table is defined as a full-dimensional quality description matrix, where rows represent different spatial locations (i.e., feature points) and columns represent different feature dimensions. See Table 1 for a possible example of a full-dimensional quality description matrix.

[0043] Table 1: Structure of the Full-Dimensional Quality Description Matrix

[0044] In some embodiments, determining the quality phase based on the full-dimensional quality description matrix involves a predefined abstract space and mapping query process. The quality space is a multi-dimensional abstract mathematical space, with the number of dimensions matching the number of feature columns in the full-dimensional quality description matrix. Each coordinate axis in the quality space corresponds to a feature column in the full-dimensional quality description matrix; for example, the first coordinate axis corresponds to "color feature value," the second to "water stain feature value," and so on. The system aggregates the spatial locations of all rows (i.e., feature vectors at all spatial locations) of the full-dimensional quality description matrix and represents the overall quality state of the current tea leaf sample by calculating the centroid coordinates of these feature vectors in the multi-dimensional quality space. The formula for calculating the centroid coordinates is:

[0045] in: This represents the total number of rows in the full-dimensional quality description matrix, i.e., the total number of feature points. This represents the extended feature vector corresponding to the i-th row in the full-dimensional quality description matrix. This represents the calculated centroid vector, i.e., the coordinates of the current sample in the quality space.

[0046] The system then queries a pre-defined quality degradation trajectory mapping table. This table is a knowledge base stored in a database. It records, in the form of a data table, a series of continuous or discrete quality stages of fresh tea leaves from their fresh state to their spoiled state. Each quality stage has a typical coordinate range in the quality space, and the evolutionary relationship between each stage constitutes the quality degradation trajectory. The system then calculates the centroid coordinates of the current sample. The calculation involves matching the coordinates of each quality stage with the typical coordinate ranges recorded in the quality degradation trajectory mapping table. This matching calculation typically includes... The Mahalanobis or Euclidean distance to the center point of the typical coordinate range of each stage is calculated, and then the quality stage with the smallest distance is selected as the matching result. This matched quality stage is defined as the current quality phase of the fresh tea leaves. The quality phase is a discrete or continuous label used to summarize the quality grade or decay state of the fresh tea leaves at the current moment. Optionally, the quality degradation trajectory mapping table can be built based on a large amount of historical multispectral image data and its simultaneously measured physicochemical quality indicators, learned through methods such as clustering, regression, or time series modeling. In some embodiments, the coordinates of the quality space can be dimensionality-reduced through principal component analysis, allowing the construction and querying of the quality degradation trajectory mapping table to be performed in a lower-dimensional space, thereby improving computational efficiency.

[0047] Example 5: In the quality degradation trajectory mapping table, starting from the coordinate point corresponding to the current quality phase, the trajectory is traced in the opposite direction. The tracing continues until the node with the most gradual feature change closest to the current coordinate point is reached; this node is defined as the previous stable quality node. Feature vectors corresponding to the current quality phase and the previous stable quality node are extracted from the full-dimensional quality description matrix. The Euclidean distance between these two feature vectors is calculated and normalized to obtain the feature deviation. A lookup table is established to correlate feature deviation with processing strategies. The calculated feature deviation values ​​are input into the lookup table, mapping to a set of suggested quality correction measures, which are the quality correction recommendations. Simultaneously, based on the magnitude of the feature deviation and the storage time of the fresh tea leaves, a quantitative score for processing urgency is calculated using a linear weighted model. The identification information of the fresh tea leaves, the detected quality phase, the feature deviation, the quality correction recommendations, and the processing urgency score are organized and packaged according to a predetermined format, outputting a complete structured inspection report.

[0048] In practical implementation, the reverse quality tracing analysis and generation of structured inspection reports in the machine vision-based rapid quality detection method for fresh tea leaves begin with the backtracking of identified quality phases. The system locates the coordinate point or coordinate range corresponding to the current quality phase of the fresh tea leaves in the quality degradation trajectory mapping table. The quality degradation trajectory mapping table not only defines the spatial range of each quality phase but also encodes the quality degradation trajectory evolving from the fresh phase to the spoilage phase in the form of a directed graph or ordered list. Starting from the coordinate point corresponding to the current quality phase, the system gradually tracks in the reverse direction defined by the quality degradation trajectory. The tracking algorithm performs reverse retrieval based on the topological relationships between adjacent phases recorded in the quality degradation trajectory mapping table. The tracking process continues until a node with a gradual feature change is located on the quality degradation trajectory. Nodes with a gradual feature change are defined in the quality degradation trajectory mapping table by a threshold difference in feature vectors between adjacent phases. When the Euclidean distance between the feature vectors of two adjacent nodes is less than the preset gradual change threshold during the reverse tracking process, the system determines that the node is the previous stable quality node. The previous stable quality node represents a stage in the quality degradation process where the quality of fresh tea leaves remained relatively stable. In some embodiments, the system extracts the aggregated feature vector corresponding to the current quality phase in the full-dimensional quality description matrix and the typical feature vector of the previous stable quality node stored in the quality degradation trajectory mapping table. The extraction operation involves reading pre-stored vector data from the database or calculating new feature vectors from a subset of the full-dimensional quality description matrix according to the phase definition.

[0049] The system calculates the Euclidean distance between the eigenvector of the current quality phase and the eigenvector of the previous stable quality node. The Euclidean distance measures the straight-line distance between two vectors in the multidimensional quality space. To obtain a standardized metric, the system normalizes the calculated raw Euclidean distance. The normalization process uses the maximum distance between the eigenvectors of the fresh and decaying phases recorded in the quality degradation trajectory mapping table as a benchmark. The calculated normalized value is defined as the eigenvalue deviation, a scalar value between 0 and 1. The formula for calculating the eigenvalue deviation is as follows:

[0050] in: The feature vector representing the current quality phase, The feature vector representing the previous stable quality node, with the symbol... This indicates that the L2 norm of a vector, i.e., the Euclidean distance, is being calculated. The baseline Euclidean distance between the eigenvectors of the fresh phase and the decay phase, as defined in the quality degradation trajectory mapping table, represents the distance between them. This is the calculated normalized feature deviation.

[0051] It is understandable that the system generates the final output based on the calculated feature deviation, and the generation process relies on a pre-established lookup table mapping feature deviation to processing strategies. The system will then quantify the feature deviation... Using the input key, a lookup table mapping feature deviation to processing strategies is retrieved, resulting in one or more sets of suggested quality correction measures. These measures constitute quality correction recommendations. The content of these recommendations may include specific operational guidelines such as "immediately spread out to dry," "reduce ambient humidity," or "recommend prioritizing processing."

[0052] The system initiates a quantitative score calculation for processing urgency. The urgency score is determined by both the magnitude of the feature deviation and the length of time the fresh tea leaves have been stored. The length of time the fresh tea leaves have been stored is calculated using the system clock and the timestamp of the fresh leaf entry record. The linear weighted model can be expressed as:

[0053] in: It is the feature deviation. This is a normalized value representing the length of time fresh tea leaves have been stored. and These are pre-defined weighting coefficients used to balance the contribution of feature deviation and storage time to processing urgency.

[0054] In some embodiments, the system finally organizes and encapsulates the identification information of the fresh tea leaves, the detected quality phase, the calculated feature deviation, the quality correction suggestions obtained from the query, and the calculated processing urgency score according to a predetermined structured format. The predetermined structured format can be XML, JSON, or a specific text template. The encapsulated data package is then output as a complete structured inspection report. The structured inspection report can be displayed through a human-machine interface or transmitted to a warehouse management system or production scheduling system to guide subsequent decisions regarding the processing of fresh tea leaves.

[0055] See Figure 5This is a graph showing the relationship between the urgency score and storage time in the quality inspection of fresh tea leaves. The core concept is the trend of urgency over storage time and the current status. The urgency score increases linearly with storage time, indicating that the longer the fresh tea leaves are stored, the higher the risk of quality deterioration and the greater the urgency of handling. The current storage time is 48 hours, corresponding to an urgency score in the medium range (approximately 0.3). Combined with the scoring rules, this reflects the overall urgency of the current degree of quality deterioration and the storage duration. This type of chart is used in the reporting stage of fresh tea leaf quality inspection, visually demonstrating the impact of storage time on handling urgency, helping operators quickly determine the current processing priority of fresh leaves, and providing a time-based decision-making basis for warehouse management.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rapid detection of tea leaf quality based on machine vision, characterized in that, The following stages are executed in sequence: The multispectral imaging device is activated to acquire an initial image sequence of fresh tea leaves under a predetermined wavelength combination. The initial image sequence was decoupled by channel to extract independent spectral image layers related to the color of tea surface, distribution of water stains on leaf surface and tissue structure. Independent spectral image layers are input into a pre-structured feature evolution network, which generates a quality decay prediction vector by simulating the natural decay path of tea quality. Based on the quality degradation prediction vector, suspected deterioration focus areas are delineated in independent spectral image layers; Dynamic sampling windows were deployed around suspected deterioration focal areas to collect local image micro-sequences that evolved over time; Temporal feature freezing operation is performed on local image micro-sequences to extract a set of freezing features that characterize the rate of quality change; By cross-domain feature grafting of the solidification feature set and the static features of the independent spectral image layer, a full-dimensional quality description matrix is ​​formed. Based on the coordinates of the full-dimensional quality description matrix in the quality space, query the preset quality degradation trajectory mapping table to determine the current quality phase of the fresh tea leaves. Based on the quality phase initiation, the reverse quality tracing analysis includes tracing back along the quality degradation trajectory mapping table to the previous stable quality node, and calculating the feature deviation between the current state and the previous stable quality node. A structured inspection report is generated based on the feature deviation, including quality correction suggestions and the urgency of the handling.

2. The method for rapid detection of tea leaf quality based on machine vision according to claim 1, characterized in that, The initial image sequence was decoupled by channel to extract independent spectral image layers related to the tea leaf surface color, water stain distribution, and tissue structure, including: From the predetermined wavelength combination of the multispectral imaging device, key wavelengths corresponding to chlorophyll reflectance peaks, cell water absorption valleys, and leaf fiber scattering characteristics were identified. For each key wavelength, the corresponding single-band image is separated from the initial image sequence; Background matrix culling is performed on each single-band image. Background matrix culling refers to removing imaging background noise that is unrelated to the tea subject by using image morphological opening operations. After background matrix removal, the single-band images are classified into color image layer, water stain distribution image layer and tissue structure image layer according to the physical properties they represent, which together constitute an independent spectral image layer.

3. The method for rapid detection of tea leaf quality based on machine vision according to claim 2, characterized in that, Independent spectral image layers are input into a pre-structured feature evolution network, which generates a quality degradation prediction vector by simulating the natural degradation path of tea quality, including: The feature evolution network contains multiple cascaded simulated decay units, each of which corresponds to a quality state at a time scale. The color image layer, water stain distribution image layer, and tissue structure image layer are fed into the initial simulated decay unit of the feature evolution network as parallel inputs. The feature evolution network iteratively calculates the degree of color decay, water stain diffusion, and loose tissue structure in each simulated decay unit according to the preset time scale order and based on the output of the previous unit and the preset decay rules. After processing by all simulated decay units, a multidimensional vector is output from the end of the network. This multidimensional vector encodes the predicted quality state of fresh tea leaves at multiple future time points, which is the quality decay prediction vector.

4. The method for rapid detection of tea leaf quality based on machine vision according to claim 1, characterized in that, Based on the quality degradation prediction vector, suspected degradation focus regions are delineated in independent spectral image layers, including: Analyze the quality degradation prediction vector and identify the vector components that predict the degree of quality degradation exceeding the threshold. The vector components are mapped back to their spatial locations in the corresponding color image layer, water stain distribution image layer, or tissue structure image layer. In each independent spectral image layer, a connected region is adaptively expanded, centered on the mapped spatial location and combined with the image gradient field of the independent spectral image layer. By merging the connected regions formed by the expansion in all independent spectral image layers, a comprehensive suspected deterioration focal region is obtained after taking the union.

5. The machine vision based method for rapid detection of tea leaf quality as claimed in claim 4, wherein, Dynamic sampling windows were deployed around the suspected deterioration focal area to collect local image micro-sequences that evolved over time, including: Using the geometric center of the suspected deterioration focal area as the origin, multiple image acquisition points are deployed at a preset distance around it in a radial, equiangular distribution pattern. A miniature imaging unit that can be periodically triggered is fixed at each image acquisition point; All miniature imaging units are controlled to synchronously capture images at uniform time intervals for several cycles, thereby obtaining a series of temporally continuous local images at each acquisition point; A series of local images at the same acquisition point are arranged in chronological order to form a local image micro-sequence of the acquisition point; Local image micro-sequences from all image acquisition points are collected and used as input for subsequent analysis.

6. The method for rapid detection of tea leaf quality based on machine vision according to claim 1, characterized in that, Temporal feature freezing is performed on local image micro-sequences to extract a set of frozen features characterizing the rate of quality change, including: For each local image micro-sequence, calculate the pixel value difference matrix between images at adjacent time points; By performing an integral operation on the pixel value difference matrix over the time axis, a cumulative change map representing the cumulative change intensity of the image acquisition points is obtained; Extract pixels whose change intensity exceeds the active threshold from the cumulative change map, and record the coordinates, intensity, and change direction information of the pixels; Perform the above operation on all local image microsequences, summarize the information of all recorded points to form a solidification feature set. The solidification feature set essentially describes the local location and dynamics of significant quality changes around the suspected deterioration focus area within the observation time window.

7. The machine vision based rapid detection method of tea leaf quality as claimed in claim 1, wherein, By cross-domain feature grafting between the solidification feature set and the static features of independent spectral image layers, a full-dimensional quality description matrix is ​​formed, including: From the independent spectral image layer, the static feature values ​​of each pixel in the suspected deterioration focal region in three dimensions of color, water stains and tissue structure are extracted to form a static feature vector field; Each feature point in the solidification feature set is located at its corresponding position in the static feature vector field based on its spatial coordinates. The dynamic change information of the feature points, including the intensity and direction of the change, is used as a new dimension and concatenated with the static feature vector at the corresponding position. After all matching positions are spliced ​​together, a data table is formed in which rows represent spatial locations and columns represent static and dynamic feature dimensions, which is the full-dimensional quality description matrix.

8. The machine vision based rapid detection method of tea leaf quality as claimed in claim 1, wherein, Based on the coordinates of the full-dimensional quality description matrix in the quality space, a pre-defined quality degradation trajectory mapping table is queried to determine the current quality phase of the fresh tea leaves, including: The quality space is a multi-dimensional abstract space, where each dimension corresponds to a feature column of the full-dimensional quality description matrix. Aggregate the spatial positions of all feature columns of the full-dimensional quality description matrix and calculate the coordinates of its centroid in the quality space. The quality degradation trajectory mapping table is a pre-built database that records the typical coordinate ranges and their changing trajectories in the quality space for different quality stages from fresh to spoilage. The calculated centroid coordinates are matched with the typical coordinate range in the quality degradation trajectory mapping table to determine the quality stage to which the current fresh tea leaves belong. This quality stage is the quality phase. 9.The machine vision based method for rapid detection of tea leaf quality according to claim 1, wherein, Based on the quality phase initiation, a reverse quality tracing analysis is performed. This analysis includes tracing back along the quality degradation trajectory mapping table to the previous stable quality node and calculating the feature deviation between the current state and the previous stable quality node, including: In the quality degradation trajectory mapping table, starting from the coordinate point corresponding to the current quality phase, the tracking proceeds in the opposite direction of the quality degradation trajectory. The node on the trajectory that has the most gradual feature change and is closest to the current coordinate point is defined as the previous stable quality node. Extract the feature vectors corresponding to the current quality phase and the previous stable quality node in the full-dimensional quality description matrix, respectively; Calculate the Euclidean distance between the two feature vectors, and then normalize the Euclidean distance to obtain the feature deviation.

10. The machine vision based rapid detection method of tea leaf quality as claimed in claim 1 wherein, Generate a structured inspection report based on the feature deviation, including quality correction suggestions and the urgency of handling, including: Establish a lookup table to map the feature deviation to the processing strategy; The calculated feature deviation values ​​are input into a lookup table, which maps to a set of suggested quality correction measures, namely the quality correction recommendations. Meanwhile, based on the magnitude of the feature deviation and the storage time of the fresh tea leaves, a quantitative score of the urgency of the processing is calculated through a linear weighted model. The identification information of fresh tea leaves, the detected quality phases, characteristic deviations, quality correction suggestions, and urgency scores are organized and packaged according to a predetermined format to output a complete structured test report.

Images