An image recognition method and system for screening of excellent individual tea plants

Abstract

The application discloses an image recognition method and system for excellent single plant screening of tea trees, and particularly relates to the field of tea tree image recognition, and comprises the following steps: collecting multiple images of a target tea tree under natural wind disturbance conditions, obtaining residual image information formed by branches and leaves under the action of wind disturbance, then performing inter-frame difference processing on the multiple images to form an initial residual image sequence containing residual image tracks; performing layered processing on the initial residual image sequence to form low-frequency residual image information and high-frequency residual image information, and identifying the continuous trend of the branch and leaf residual image tracks in the low-frequency residual image information and the high-frequency residual image information, thereby outputting residual image structure representation. Through dynamic residual image extraction, track layering and coupling analysis on the multiple images of the tea tree under the natural wind disturbance conditions, the time sequence characteristics of the branch and leaf swing are converted into calculable indexes reflecting the toughness and elasticity, so that excellent single plant screening based on biomechanical characteristics is realized.

Description

Technical Field

[0001] This invention relates to the field of tea tree image recognition technology, and more specifically, to an image recognition method and system for screening superior individual tea trees. Background Technology

[0002] In the image recognition process for screening superior individual tea plants, the images collected in the field are generally affected by the natural environment, especially the swaying of branches and leaves under the action of wind. Conventional image processing techniques usually identify the resulting jitter and afterimages as noise and eliminate them through filtering, inter-frame stabilization or motion compensation in order to obtain a clear and static plant morphology.

[0003] However, this "denoising" approach overlooks an important fact: the amplitude and frequency of the swaying of branches and leaves in the wind are not meaningless interference, but directly reflect potential superior traits such as leaf toughness and branch elasticity. In real-world scenarios, these dynamic afterimages carry biomechanical characteristics that are difficult to represent through a single static image. If they are completely removed, the screening process will only focus on comparing surface features such as leaf shape and color, missing the opportunity for a deeper assessment of the core traits of the tea tree.

[0004] Therefore, the key flaw of the current technical approach is that it treats all dynamic afterimages as worthless noise, which makes image recognition unable to effectively capture the structural toughness and functional elasticity of tea trees under natural disturbance. This neglect makes the selection of superior individual plants inherently limited. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an image recognition method and system for screening superior tea plants. By performing dynamic afterimage extraction, trajectory layering, and coupling analysis on multiple frames of tea plant images under natural wind disturbance conditions, the temporal characteristics of branch and leaf swaying are transformed into calculable indicators reflecting toughness and elasticity, thereby achieving the screening of superior tea plants based on biomechanical characteristics, thus solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an image recognition method for screening superior tea plantlets, comprising:

[0007] S1. By acquiring multiple frames of images of the target tea tree under natural wind disturbance conditions, the residual image information formed by the branches and leaves under the action of wind disturbance is obtained. Then, the multiple frames of images are processed to form an initial residual image sequence containing the residual image trajectory.

[0008] S2. Perform layered processing on the initial afterimage sequence to form low-frequency afterimage information and high-frequency afterimage information, and identify the continuous direction of the afterimage trajectory of branches and leaves in the low-frequency afterimage information and the high-frequency afterimage information, thereby outputting the afterimage structure representation;

[0009] S3. By dividing the shadow structure representation into parts, the crown region, lateral branch region and lower leaf region are extracted respectively. Then, the shadow features of each region are compared to form a consistency judgment result of the shadow features inside the plant.

[0010] S4. Perform coupling analysis between the consistency determination result and the branch orientation information obtained based on the static image, and at the same time construct a counterfactual sequence generated by time-series perturbation and perform comparison to solve the environmental interference afterimage and remove it from the afterimage structure representation, thereby forming a tough and elastic evidence chain.

[0011] S5. Perform fusion processing on the toughness and elasticity evidence chain to generate a single-plant afterimage excellence index. When the single-plant afterimage excellence index meets the preset judgment conditions, output the target tea tree as an excellent single plant. If the judgment conditions are not met, return to S3 and re-execute.

[0012] In a preferred embodiment, the process of forming an initial afterimage sequence containing afterimage trajectories includes the following steps:

[0013] S1-1. The imaging device continuously acquires multiple frames of images of the target tea tree at a fixed sampling interval. Each frame of the image is divided into a pixel matrix, and the brightness and color values ​​in the pixel matrix are obtained to construct a basic set of afterimage information including the edges of branches and leaves, leaf vein details and local shadows.

[0014] S1-2. Perform pixel difference operation on adjacent frame images in the afterimage information base set one by one, calculate the brightness difference and color difference of each pixel, and output the pixel difference matrix.

[0015] S1-3. Perform threshold segmentation on the pixel difference matrix to identify the set of pixels that exceed the threshold, form a ghost image candidate region, and record the directionality of pixel changes in the ghost image candidate region;

[0016] S1-4. Perform connectivity analysis on the candidate region of the afterimage, combine adjacent pixels into afterimage segments according to spatial connectivity rules, and then perform trajectory matching on the afterimage segments of adjacent frames based on time order to form a continuous branch and leaf movement path.

[0017] S1-5. Arrange the continuous branch and leaf movement paths sequentially on the time axis and output an initial afterimage sequence containing afterimage trajectories.

[0018] In a preferred embodiment, S2 includes:

[0019] S2-1. The initial afterimage sequence is traversed by a time sliding window of fixed length. A slow-change response map is generated by performing a moving average operation on the brightness and color values ​​of the same pixel in multiple frames within the window. A fast-change energy map is generated by performing absolute value accumulation on the brightness and color differences of the same pixel between adjacent frames.

[0020] S2-2. Perform preset threshold judgment on the slow change response map and the fast change energy map to form a layered result: when the fast change energy is less than the upper limit of the preset threshold and the moving average change within the window is greater than the lower limit of the preset threshold, the corresponding pixel is marked as a low-frequency ghosting pixel, thereby generating low-frequency ghosting information; when the fast change energy is greater than the lower limit of the preset threshold and the count of this state in consecutive frames reaches the lower limit of the preset threshold, the corresponding pixel is marked as a high-frequency ghosting pixel, thereby generating high-frequency ghosting information.

[0021] S2-3. Perform spatial alignment and overlay mapping on the low-frequency and high-frequency afterimage information in the same coordinate system. By performing connectivity filtering on the overlay region and its one-pixel neighborhood, identify the set of ridge pixels with continuous brightness changes. Then, perform thinning and denoising on the ridge pixel set to form a set of candidate trajectory seed points.

[0022] S2-4. Perform continuous direction recognition on the candidate trajectory seed point set: For each candidate trajectory seed point, calculate the gradient direction and direction consistency index within its local window, perform pixel-level growth along the main direction in the current frame, and search for continuation points in the neighborhood of the set radius in the next frame in chronological order. When the direction deviation does not exceed the upper limit of the preset angle, the spatial interval does not exceed the upper limit of the pixel, and the brightness continuity meets the lower limit of the preset limit, the connection is confirmed; otherwise, the current branch is terminated, thus forming a set of branch and leaf afterimage trajectories sorted by time.

[0023] S2-5. Perform structured combination on the set of afterimage trajectories of branches and leaves and output the afterimage structure representation. The afterimage structure representation includes: the time index sequence of each trajectory, the spatial node sequence of each trajectory, the direction sequence of each trajectory, the corresponding low-frequency and high-frequency label distribution, the coverage area mask, and the breakpoint position annotation.

[0024] In a preferred embodiment, in S3, the process of determining the consistency of the afterimage features forming inside the plant includes the following steps:

[0025] S3-1. Perform spatial layering processing on the spatial node sequence represented by the afterimage structure according to the coordinate value range in the vertical direction. Define the upper third of the spatial nodes as the crown region, the middle third of the spatial nodes as the lateral branch region, and the lower third of the spatial nodes as the lower leaf region, thereby completing the division of the crown region, lateral branch region, and lower leaf region.

[0026] S3-2. In each region of the crown region, lateral branch region, and lower leaf region, the afterimage trajectory of branches and leaves is aggregated between frames according to the time index sequence. A weighted average operation is performed on the trajectory nodes of the same frame in the same region to generate a region afterimage feature vector representing the overall swing trend of branches and leaves in that region.

[0027] S3-3. Standardize the mean brightness gradient, direction sequence similarity, and low-frequency to high-frequency label ratio contained in the afterimage feature vectors of each region to form a feature index set for inter-region comparison.

[0028] In a preferred embodiment, in S3, the process of determining the consistency of the afterimage features forming inside the plant further includes the following steps:

[0029] S3-4. Using the feature index set of the crown region as a reference, perform Euclidean distance calculation and direction deviation calculation on the feature index sets of the lateral branch region and the lower leaf region respectively, and generate the difference matrix between each region.

[0030] S3-5. Perform normalization operation on the difference matrix and calculate its average deviation. When the average deviation is lower than the preset threshold, output the consistency judgment result of the afterimage feature inside the plant as consistent; otherwise, output the inconsistent state.

[0031] In a preferred embodiment, the process of forming a chain of evidence of resilience and elasticity includes the following steps:

[0032] S4-1. By extracting the green channel image from the RGB color space of the static image and performing contrast stretching and noise suppression, an enhanced image for structure extraction is obtained; edge detection, connected component extraction and skeleton thinning are performed on the enhanced image to construct a candidate set of branch centerlines composed of connected curves.

[0033] S4-2. Filter the candidate set of branch centerlines according to the lower limit of length and the upper limit of curvature change rate, and retain curves whose length is not lower than the lower limit of length and whose curvature change rate is not higher than the upper limit of curvature change rate; calculate the tangential direction point by point along the spatial node sequence of each retained curve to form branch orientation information composed of curve identifier, spatial node sequence and corresponding tangential direction.

[0034] S4-3. Perform spatial registration processing on the afterimage structure representation corresponding to the branch orientation information and the consistency judgment result in the same image coordinate system, so that the spatial node sequence of the afterimage trajectory and the spatial node sequence of the branch orientation information are established in a one-to-one correspondence, forming a spatial coupling reference.

[0035] S4-4. On the spatial coupling reference, perform projection matching node by node along the time index sequence for each afterimage trajectory: with the current afterimage trajectory node as input, solve the spatial interval from the node to the center line of the corresponding branch and the directional deviation between the node's direction and the corresponding tangential direction; at the trajectory level, count the proportion of nodes whose directional deviation is not higher than the upper limit of the directional deviation and whose spatial interval is not higher than the upper limit of the spatial interval, and generate a coupling coefficient map.

[0036] In a preferred embodiment, the process of forming a chain of evidence for resilience and elasticity further includes the following steps:

[0037] S4-5. While keeping the spatial node sequence of the afterimage trajectory unchanged, the original time index sequence is shuffled to obtain the counterfactual sequence of time order perturbation. The projection matching and statistical process of S4-4 is reused to generate the comparison coupling coefficient map. The coupling coefficient map is compared with the comparison coupling coefficient map position by position to identify the afterimage trajectory segments that still maintain the target coupling coefficient after the time order perturbation, and they are marked as environmental interference afterimages.

[0038] S4-6. Remove the environmental interference afterimage from the afterimage structure representation, re-statistically evaluate the three indicators of directional stability, deformation amplitude, and time duration on the afterimage trajectory after removal, generate a toughness and elasticity evidence chain consisting of trajectory identifier, directional stability, deformation amplitude, time duration, and corresponding spatial overlay mask, and output it.

[0039] In a preferred embodiment, the process of generating and determining the excellence index of a single plant's afterimage includes the following steps:

[0040] S5-1. The three indicators of directional stability, deformation amplitude and time duration in the toughness and elasticity evidence chain are fused and calculated according to the trajectory identifier: the weighted value of deformation amplitude is subtracted from the weighted value of directional stability and the weighted value of time duration to generate an intermediate fusion value, and the intermediate fusion value of all trajectories is normalized to form a single-plant afterimage excellence index.

[0041] S5-2. Compare the single-plant afterimage quality index with the preset judgment threshold. When the single-plant afterimage quality index is not lower than the preset judgment threshold, output the judgment result that the corresponding tea tree is a good single plant. When the single-plant afterimage quality index is lower than the judgment threshold, return to step S3 and execute again.

[0042] An image recognition system for screening superior tea plants includes a residual image extraction module, a trajectory layering module, a feature comparison module, an interference removal module, and a superior plant determination module.

[0043] The afterimage extraction module acquires multiple frames of images of the target tea tree under natural wind disturbance conditions, obtains afterimage information formed by branches and leaves under wind disturbance, and then performs inter-frame difference processing on the multiple frames to form an initial afterimage sequence containing afterimage trajectories.

[0044] The trajectory layering module is used to perform layering processing on the initial afterimage sequence to form low-frequency afterimage information and high-frequency afterimage information, and to identify the continuous direction of the branch and leaf afterimage trajectory in the low-frequency afterimage information and the high-frequency afterimage information, thereby outputting the afterimage structure representation;

[0045] The feature comparison module divides the afterimage structure representation into parts, extracting the crown region, lateral branch region, and lower leaf region respectively, and then compares the afterimage features of each region to form a consistency judgment result of the afterimage features within the plant.

[0046] The interference removal module is used to perform coupled analysis on the consistency judgment result and the branch orientation information solved based on the static image. At the same time, it constructs a counterfactual sequence generated by time-series perturbation and performs comparison to solve the environmental interference afterimage and remove it from the afterimage structure representation, thereby forming a tough and elastic evidence chain.

[0047] The excellent judgment module is used to perform fusion processing on the toughness and elasticity evidence chain to generate a single-plant afterimage excellent index. When the single-plant afterimage excellent index meets the preset judgment conditions, the target tea tree single plant is output as an excellent single plant. When the judgment conditions are not met, the feature comparison module is returned to re-execute.

[0048] The technical effects and advantages of this invention are as follows:

[0049] This invention acquires multiple frames of tea tree images under natural wind disturbance conditions, transforming the afterimages of branches and leaves, which are considered noise in traditional identification, into the object of analysis. This achieves a fundamental shift from "removing interference" to "utilizing interference," enabling the system to extract the toughness and elasticity information of tea tree branches and leaves based on the dynamic features of the afterimages. This solves the technical problem that existing technologies cannot capture the core traits of superior individual plants.

[0050] This invention restores the swaying trajectory of branches and leaves in a time series into a computable set of dynamic paths through inter-frame difference calculation and structuring of afterimage trajectory. It also separates the overall swaying and local jitter by using low-frequency and high-frequency afterimage layering, enabling the system to identify the response characteristics of branches and leaves at different structural levels and improving the accuracy and stability of dynamic feature extraction.

[0051] This invention establishes a difference matrix between the crown, lateral branches, and lower leaf regions by comparing and calculating the spatial node layering and regional features of the afterimage structure. This enables a quantitative determination of the consistency of afterimage features within the plant, thus providing a quantifiable analytical basis for the balance and overall coordination of the tea tree's growth structure.

[0052] This invention achieves adaptive removal of environmental interference afterimages by spatially coupling analysis of static branch orientation information and dynamic afterimage structure, combined with counterfactual sequences generated by time perturbation. This enables the system to distinguish between real structural responses and external noise responses in complex natural environments, thereby improving the robustness of dynamic recognition.

[0053] This invention constructs a chain of evidence for toughness and elasticity by recalculating and fusing three-dimensional indicators of directional stability, deformation amplitude and time persistence, and generates a single-plant residual image quality index in a weighted normalization manner, realizing the transformation from dynamic response data to quantitative judgment of quality, and making the selection of superior single plants scientific and automated. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method steps of the present invention.

[0055] Figure 2 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0056] 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.

[0057] Refer to the instruction manual appendix Figure 1-2 An embodiment of the present invention provides an image recognition method for screening superior tea plantlets, comprising:

[0058] S1. By acquiring multiple frames of images of the target tea tree under natural wind disturbance conditions, the residual image information formed by the branches and leaves under the action of wind disturbance is obtained. Then, the multiple frames of images are processed to form an initial residual image sequence containing the residual image trajectory.

[0059] S2. Perform layered processing on the initial afterimage sequence to form low-frequency afterimage information and high-frequency afterimage information, and identify the continuous direction of the afterimage trajectory of branches and leaves in the low-frequency afterimage information and the high-frequency afterimage information, thereby outputting the afterimage structure representation;

[0060] S3. By dividing the shadow structure representation into parts, the crown region, lateral branch region and lower leaf region are extracted respectively. Then, the shadow features of each region are compared to form a consistency judgment result of the shadow features inside the plant.

[0061] S4. Perform coupling analysis between the consistency determination result and the branch orientation information obtained based on the static image, and at the same time construct a counterfactual sequence generated by time-series perturbation and perform comparison to solve the environmental interference afterimage and remove it from the afterimage structure representation, thereby forming a tough and elastic evidence chain.

[0062] S5. Perform fusion processing on the toughness and elasticity evidence chain to generate a single-plant afterimage excellence index. When the single-plant afterimage excellence index meets the preset judgment conditions, output the target tea tree as an excellent single plant. If the judgment conditions are not met, return to S3 and re-execute.

[0063] The process of forming an initial afterimage sequence containing afterimage trajectories includes the following steps:

[0064] S1-1. Multiple frames of images of the target tea tree are continuously acquired by an imaging device at a fixed sampling interval. Each frame is divided into a pixel matrix, and the brightness and color values ​​in the pixel matrix are obtained to construct a basic set of afterimage information including leaf edges, leaf vein details, and local shadows. The imaging device includes an optical lens for acquiring multiple frames of images of the target tea tree, an image sensor, and a signal acquisition and storage component. Dividing each frame into a pixel matrix means decomposing the image into a two-dimensional array of individual pixels in the order of rows and columns. Each pixel corresponds to a brightness and color value at a specific location. The pixel matrix is ​​essentially a mathematical representation of the image in digital form, directly generated after digitization of the raw data acquired by the image sensor.

[0065] In addition, the brightness values ​​in the pixel matrix construct the edge information of branches and leaves through abrupt changes in the grayscale gradient, the color values ​​depict the details of leaf veins through changes in color channel differences, and the decrease in brightness values ​​in local areas represents shadows. Thus, the combination of the three forms the basic set of afterimage information.

[0066] S1-2. Perform pixel difference operation on adjacent frame images in the afterimage information base set one by one, calculate the brightness difference and color difference of each pixel, and output the pixel difference matrix.

[0067] S1-3. Perform threshold segmentation on the pixel difference matrix to identify the set of pixels exceeding the threshold, forming a candidate region for afterimages. Record the directionality of pixel changes in the candidate region for subsequent trajectory tracking. Threshold segmentation refers to an image processing method that separates pixels in the pixel matrix that are higher or lower than a preset value, thereby distinguishing the foreground region from the background region. The directionality of pixel changes refers to the spatial trend of pixel intensity or color difference in adjacent frames, such as continuous increase or decrease along the horizontal, vertical, or diagonal direction, used to characterize the trajectory of the afterimage of branches and leaves in motion.

[0068] S1-4. Perform connectivity analysis on the candidate region of the afterimage, combine adjacent pixels into afterimage segments according to spatial connectivity rules, and then perform trajectory matching on the afterimage segments of adjacent frames based on time order to form a continuous branch and leaf movement path.

[0069] It should be noted that connectivity analysis detects the relationship between adjacent pixels in an image to determine which pixels belong to the same continuous region; spatial connectivity rules are the methods for determining pixel adjacency, including but not limited to four-connectivity (adjacent vertically and horizontally) and eight-connectivity (including diagonal adjacency).

[0070] When combining images into a fragmented image according to spatial connectivity rules, the connectivity relationship of pixels in the candidate region is checked point by point, and the set of pixels that satisfy the connectivity rules is merged into a whole region, which is the fragmented image.

[0071] In addition, trajectory matching refers to associating the image fragments of different frames with each other based on the similarity of their position, shape and direction in adjacent frames. When the image fragments of adjacent frames are successfully matched, these fragments are strung together in time order to obtain the continuous motion path of the image fragments of the branches and leaves swaying in the wind in the image sequence.

[0072] S1-5. Arrange the continuous branch and leaf movement paths sequentially on the time axis, output an initial afterimage sequence containing afterimage trajectories, and use the initial afterimage sequence as input for subsequent afterimage layering processing.

[0073] S2 includes:

[0074] S2-1. The initial afterimage sequence is traversed through a fixed-length time sliding window. A slow-change response map is generated by performing a moving average operation on the brightness and color values ​​of the same pixel in multiple frames within the window. A fast-change energy map is generated by accumulating the absolute values ​​of the brightness and color differences of the same pixel between adjacent frames. It should be noted that the moving average operation refers to taking the average value of the brightness or color value of the same pixel position for a fixed number of frames in a continuous image to smooth random fluctuations and extract the trend of slow changes over time. The absolute value accumulation refers to calculating the absolute value of the difference in brightness or color of the same pixel in adjacent frames and accumulating it frame by frame in the time dimension to highlight pixel areas with rapid changes or strong jitter.

[0075] In addition, the fast change energy map is essentially a two-dimensional mapping that reflects the intensity of brightness or color changes of pixels in a time series, and is used to characterize areas of rapid jitter or instantaneous motion; the slow change response map is essentially a two-dimensional mapping that describes the smooth change trend of pixels in the time dimension, and is used to characterize areas of slow displacement or overall oscillation.

[0076] S2-2. Perform preset threshold judgment on the slow change response map and the fast change energy map to form a layered result: When the fast change energy is less than the upper limit of the preset threshold and the moving average change within the window is greater than the lower limit of the preset threshold, the corresponding pixel is marked as a low-frequency ghosting pixel, thereby generating low-frequency ghosting information; when the fast change energy is greater than the lower limit of the preset threshold and the count of this state in consecutive frames reaches the lower limit of the preset threshold, the corresponding pixel is marked as a high-frequency ghosting pixel, thereby generating high-frequency ghosting information; where fast change energy is a metric in the fast change energy map, representing the cumulative intensity of the absolute value of the difference in brightness or color value of the same pixel in consecutive frames in the time dimension, used to highlight high-energy areas of local shaking or instantaneous movement of branches and leaves in the fast change energy map;

[0077] Moving average change refers to the difference between the moving average values ​​of the same pixel over a continuous time window, and is used to measure the magnitude of a pixel's brightness or color change over time.

[0078] Additionally, in S2-2, this state refers to the state in which the pixel's rapid change energy is greater than the preset threshold lower limit in consecutive frames. That is, the pixel is continuously at a high energy change level in the time series, which is used to determine that it belongs to the high-frequency ghosting region.

[0079] Counting consecutive frames refers to the total number of frames in an image sequence in which a certain pixel appears consecutively and meets preset conditions (including: the fast change energy exceeds a preset threshold).

[0080] S2-3. Perform spatial alignment and overlay mapping on low-frequency and high-frequency afterimage information in the same coordinate system. By performing connectivity filtering on the overlay region and its one-pixel neighborhood, identify the set of ridge pixels with continuous brightness changes, and perform thinning and denoising on the ridge pixel set to form a set of candidate trajectory seed points. Spatial alignment and overlay mapping refers to performing position correction on low-frequency and high-frequency afterimage information in the same spatial coordinate system, so that the corresponding pixels correspond one-to-one, and then overlaying the pixel values ​​or feature values ​​of the two to express the overall sway and local jitter features on the same image plane at the same time.

[0081] It should be noted that S2-3 means that by performing connectivity filtering on the superimposed region and its one-pixel neighborhood, in the image after the low-frequency and high-frequency afterimage information are superimposed, the brightness continuity of the one-pixel neighborhood around each pixel is detected with each pixel as the center, and pixels with the same brightness change direction and intensity difference below the preset threshold are classified into the same connected region.

[0082] When identifying a set of ridge pixels with continuous brightness changes, the brightness gradient direction field is calculated in each connected region, and pixels with stable gradient direction and local maximum gradient magnitude are extracted as ridge pixels. In addition, when performing thinning and denoising on the ridge pixel set, the process includes iteratively removing branch pixels with non-endpoint neighborhood degree greater than two and removing short chain segments with length less than the minimum threshold, thereby preserving the central path and forming a set of candidate trajectory seed points.

[0083] In this scheme, the ridge pixel set is essentially a long, continuous chain of pixels composed of pixels with stable brightness gradient direction and local maximum gradient magnitude in the image, used to represent the central path of brightness change in the afterimage of branches and leaves; the candidate trajectory seed point set is essentially a set of structural core pixels retained from the ridge pixel set, representing the starting and continuing positions of the afterimage of branches and leaves that may form a continuous motion trajectory in the time series.

[0084] S2-4. Perform continuous direction recognition on the candidate trajectory seed point set: For each candidate trajectory seed point, calculate the gradient direction and direction consistency index within its local window, perform pixel-level growth along the main direction in the current frame, and search for continuation points in the neighborhood of the set radius in the next frame in chronological order. When the direction deviation does not exceed the upper limit of the preset angle, the spatial interval does not exceed the upper limit of the pixel, and the brightness continuity meets the lower limit of the preset limit, the connection is confirmed; otherwise, the current branch is terminated, thus forming a set of branch and leaf afterimage trajectories sorted by time.

[0085] The gradient direction and direction consistency index within a local window refers to calculating the gradient direction of each pixel's brightness change in its neighborhood, centered on the candidate trajectory seed point, and then calculating the angle difference between the gradient direction of each pixel in the neighborhood and the main direction, based on the main direction. When the average value of these angle differences is lower than the set upper limit of the angle, the direction consistency index takes a high value, indicating that the pixel brightness change direction in this area is concentrated and consistent.

[0086] In S2-4, the main direction refers to the average vector direction of all pixel gradient directions in the local window of the candidate trajectory seed point, which is used to represent the main extension direction of brightness change or branch movement.

[0087] Pixel-level growth refers to starting from the current pixel and extending along the main direction pixel by pixel to search for adjacent pixels. When the brightness continuity and directional deviation of adjacent pixels meet the set conditions, the pixel is incorporated into the trajectory path to achieve point-by-point expansion of the trajectory.

[0088] Searching for a continuation point within a set radius neighborhood in the next frame in chronological order means that in the next frame of the time series, with the current trajectory's end pixel as the center, searching within a predetermined radius neighborhood for the pixel whose brightness distribution and orientation features best match the current point, and using it as the continuation position of the trajectory in time.

[0089] In addition, spatial interval refers to the Euclidean distance between the current trajectory end pixel and the candidate continuation pixel on the image plane, which is used to limit the spatial continuity of trajectory connection; directional deviation refers to the angle between the main direction of the current trajectory and the brightness gradient direction of the candidate continuation pixel, which is used to measure the consistency of the trajectory extension direction; brightness continuity refers to the fact that the amplitude of brightness value change between adjacent pixels remains within a preset threshold range, reflecting the smoothness of brightness change in local areas of the image.

[0090] S2-5. Perform structured combination on the set of branch and leaf afterimage trajectories and output the afterimage structure representation. The afterimage structure representation includes: the time index sequence of each trajectory, the spatial node sequence of each trajectory, the direction sequence of each trajectory, the corresponding low-frequency and high-frequency label distribution, the coverage area mask, and the breakpoint position label. The afterimage structure representation is used as the input for the subsequent part segmentation step.

[0091] In S2-5, it should be noted that the scattered trajectory information of the branch and leaf afterimage trajectory set is transformed into a computable data structure by structuring the data. Specifically, this includes: First, renumbering each branch and leaf afterimage trajectory according to the order of its appearance in the frames to form a time index sequence representing time changes; then extracting the spatial coordinates of each pixel in the trajectory and arranging them sequentially to form a spatial node sequence; next, calculating the brightness gradient direction between each node and continuously recording the direction values ​​of adjacent nodes to form a direction sequence; simultaneously, attaching corresponding low-frequency or high-frequency labels to each node of the trajectory based on whether the frame source of the trajectory is a low-frequency or high-frequency afterimage; then generating a coverage area mask based on the boundary range of all nodes in the trajectory to represent the coverage area of ​​the branch and leaf movement in the image; finally, detecting nodes in the trajectory whose brightness or direction changes exceed a threshold and marking these positions as breakpoints. After the above steps are integrated, the output afterimage structure representation includes the time index sequence, spatial node sequence, direction sequence, label distribution, coverage mask, and breakpoint annotations, which are used as input for subsequent part segmentation.

[0092] In S3, the process for determining the consistency of the residual features within the plant includes the following steps:

[0093] S3-1. Perform spatial layering processing on the spatial node sequence represented by the afterimage structure according to the vertical coordinate value range. Define the upper third of the spatial nodes as the crown region, the middle third of the spatial nodes as the lateral branch region, and the lower third of the spatial nodes as the lower leaf region, thereby completing the division of the crown region, lateral branch region, and lower leaf region. The spatial node sequence refers to the spatial node sequence of each trajectory in S2-5 above. Its meaning is: in a branch and leaf afterimage trajectory, according to the movement order of the trajectory in the image plane, the set of all pixel coordinate points or feature nodes are recorded in sequence, which is used to describe the continuous path of the branch and leaf afterimage in spatial position.

[0094] In S3-1, it should be further explained that in the actual application of this scheme, by extracting the vertical coordinate values ​​of the spatial nodes in the image coordinate system, calculating the maximum and minimum values ​​of all spatial nodes, and then dividing them into three segments at equal intervals according to the numerical range: the upper segment corresponds to the crown area, the middle segment corresponds to the lateral branch area, and the lower segment corresponds to the lower leaf area, thereby realizing the layered division of the crown, lateral branches, and lower leaves.

[0095] S3-2. Within each region of the crown region, lateral branch region, and lower leaf region, the afterimage trajectories of branches and leaves are aggregated between frames based on the time index sequence. A weighted average operation is performed on the trajectory nodes of the same frame within the same region to generate a region afterimage feature vector representing the overall swaying trend of branches and leaves in that region. The time index sequence refers to the ordered index set formed by numbering each spatial node in each afterimage trajectory of branches and leaves according to the time order of its appearance frames, which is used to record the temporal position distribution of the trajectory in the image sequence.

[0096] Inter-frame aggregation refers to summarizing multiple afterimage trajectory nodes belonging to the same region within the same time frame according to the time index, and performing weighted averaging or statistical synthesis on their brightness values, directional features, etc., to obtain the overall afterimage feature expression of the region within the time period.

[0097] In addition, the weighted average operation refers to taking multiple trajectory nodes in the same frame within the same region, assigning different weights to each node based on the importance of its features (such as the magnitude of brightness change or directional stability), and then calculating the weighted average value to obtain a vector that comprehensively reflects the motion characteristics of all nodes in the region. This operation is used to balance local differences so that the output regional afterimage feature vector accurately expresses the overall swaying trend of the branches and leaves in the region.

[0098] S3-3. Standardize the mean brightness gradient, direction sequence similarity, and low-frequency to high-frequency label ratio contained in the feature vector of each region's afterimage to form a feature index set for inter-region comparison; where the mean brightness gradient refers to the average value of the brightness gradient amplitude of all trajectory nodes in the same region, which is used to reflect the overall brightness change intensity of the afterimage of branches and leaves in that region.

[0099] Directional sequence similarity refers to the degree of similarity between the directional sequences of trajectory nodes in different regions. It is used to measure the consistency of the movement direction of branches and leaves in each region by calculating the cosine similarity or correlation coefficient of the directional angle sequence.

[0100] The ratio of low-frequency to high-frequency tags refers to the ratio of the number of low-frequency afterimage nodes to the number of high-frequency afterimage nodes in the same area, which is used to reflect the relative proportion of overall swaying and local shaking characteristics in the movement of branches and leaves in that area.

[0101] In S3, the process of determining the consistency of the afterimage features formed inside the plant also includes the following steps:

[0102] S3-4. Using the feature index set of the crown region as a reference, perform Euclidean distance calculation and direction deviation calculation on the feature index sets of the lateral branch region and the lower leaf region respectively, and generate a difference matrix between each region. The Euclidean distance calculation is to calculate the square difference of the numerical elements of the feature index set of each region in the same dimension and take the square root to quantify the numerical difference of the feature intensity between different regions. The direction deviation calculation is to compare the angle difference of the average direction angle between regions to measure the consistency of the direction of branch and leaf movement. After obtaining the feature intensity difference and direction difference respectively, arrange the Euclidean distance value and direction deviation value of each region in a matrix form to generate the difference matrix between each region. This matrix contains the numerical difference and direction difference between the crown, lateral branch and lower leaf regions, thereby describing the consistent distribution of the afterimage features inside the plant.

[0103] In this scheme, the difference matrix includes a two-dimensional numerical table with regions as rows and regions as columns. Each matrix element represents the Euclidean distance and directional deviation between two regions (such as crown and lateral branch, crown and lower leaf, lateral branch and lower leaf). The difference matrix is ​​essentially a symmetric numerical matrix used to describe the difference relationship of afterimage features between regions.

[0104] S3-5. Perform normalization on the difference matrix and calculate its average deviation. When the average deviation is lower than a preset threshold, output the consistency judgment result of the afterimage features inside the plant as consistent; otherwise, output an inconsistent state. The normalization operation refers to linearly scaling each value in the difference matrix according to its maximum and minimum values ​​to unify its range to between 0 and 1 for comparability analysis. The average deviation is calculated by taking the average of the absolute differences between all elements and the mean of the matrix in the normalized matrix, which is used to measure the overall dispersion of the afterimage feature differences in each region.

[0105] The process of forming a chain of evidence demonstrating resilience and flexibility includes the following steps:

[0106] S4-1. An enhanced image for structure extraction is obtained by extracting the green channel image from the RGB color space of the static image and performing contrast stretching and noise suppression. Edge detection, connected component extraction, and skeleton thinning are performed on the enhanced image to construct a candidate set of branch centerlines composed of slender connected curves. The green channel image is extracted because, under natural visible light conditions, the reflection intensity of tea tree branches and leaves is highest in the green band, and the green channel can most clearly present the brightness difference between branches and leaves, thus facilitating the detection of structural edges. In addition, contrast stretching redistributes the image grayscale values ​​to a wider dynamic range through linear mapping to enhance structural details, while noise suppression removes isolated noise points by applying medium-mean-mean-filter or Gaussian filtering, making the continuity of branch edges more stable in subsequent processing.

[0107] Furthermore, in S4-1, edge detection is performed by calculating the gradient change of pixel brightness in the image in spatial location, identifying pixels whose gradient magnitude exceeds a preset threshold, and marking these pixels as boundary pixels, thereby extracting the contour line between the branches and the background.

[0108] Connected component extraction is the process of identifying a set of interconnected boundary pixels as an independent region in a binarized edge image based on the spatial adjacency relationship of pixels (such as 4-connected or 8-connected), with each region representing a continuous branch structure.

[0109] Skeleton thinning is an iterative process that gradually removes edge pixels and retains central connected pixels within a connected region. By continuously peeling away the outer layers while keeping the topology unchanged, each branch region eventually converges to a centerline with a width of one pixel, which is used to express the spatial movement of the branch.

[0110] S4-2. Filter the candidate set of branch centerlines according to the lower limit of length and the upper limit of curvature change rate, and retain curves whose length is not lower than the lower limit of length and whose curvature change rate is not higher than the upper limit of curvature change rate; calculate the tangential direction point by point along the spatial node sequence of each retained curve to form branch orientation information composed of curve identifier, spatial node sequence and corresponding tangential direction.

[0111] In S4-2, it should be noted that: geometric screening and direction calculation are performed on the candidate set of branch centerlines to generate branch orientation information that can be used for spatial coupling analysis; specifically, each curve in the candidate set of branch centerlines is regarded as an ordered set of points formed by connecting several spatial nodes in coordinate order. First, the total length of the curve is calculated and compared with the set lower limit of length. When the total length is not lower than the lower limit of length, it is determined that the length condition is met; then, the curvature value is calculated on the curve in units of three consecutive spatial nodes, and the rate of curvature change is statistically analyzed within the entire curve range. When the rate of curvature change is not higher than the upper limit of curvature change, it is determined to be a branch curve with a smooth shape and is retained.

[0112] In addition, for each selected curve, the tangential direction of the line connecting adjacent nodes is calculated point by point along its spatial node sequence. The obtained direction angles are recorded in the order of the nodes to form a data set consisting of curve identifier, spatial node sequence and corresponding tangential direction. This data set is the branch orientation information, which is used for spatial coupling matching with the afterimage trajectory in subsequent steps.

[0113] S4-3. Perform spatial registration processing on the afterimage structure representation corresponding to the branch orientation information and the consistency judgment result in the same image coordinate system, so that the spatial node sequence of the afterimage trajectory and the spatial node sequence of the branch orientation information are established in a one-to-one correspondence, forming a spatial coupling reference; wherein the spatial registration processing refers to aligning the branch orientation information and the afterimage structure representation in the same image coordinate system, and by calculating the translation, rotation and scale difference of the two sets of spatial node sequences in the position coordinates, applying geometric transformation (such as affine or perspective transformation) to map the two into a unified reference frame;

[0114] Establishing a one-to-one correspondence means that in the same coordinate system after registration, taking each spatial node of the afterimage trajectory as a reference, searching for the node with the closest distance and the lowest direction difference in the branch orientation information and pairing them up, thereby achieving a unique correspondence between nodes.

[0115] The final spatial coupling reference includes the coordinates of the centerline of the registered branches, the coordinates of the corresponding afterimage trajectory nodes, and their direction matching relationships, which are used for subsequent coupling coefficient calculation and environmental interference determination.

[0116] S4-4. On the spatial coupling reference, perform projection matching node by node along the time index sequence for each afterimage trajectory: taking the current afterimage trajectory node as input, solve the spatial interval from the node to the corresponding branch centerline and the directional deviation between the node's direction and the corresponding tangential direction; statistically analyze the proportion of nodes whose directional deviation is not higher than the upper limit of directional deviation and whose spatial interval is not higher than the upper limit of spatial interval at the trajectory level, and generate a coupling coefficient map, where the coupling coefficient corresponding to each spatial node position is the value of the proportion of nodes that meet the above upper limit conditions; in addition, the coupling coefficient map is a two-dimensional numerical distribution map constructed with spatial coordinates as the horizontal and vertical axes, and its content comes from the statistical results obtained after performing point-by-point matching between all afterimage trajectory nodes and branch centerline nodes on the spatial coupling reference;

[0117] At each spatial node location, the ratio of the number of nodes with directional deviations no higher than the upper limit of directional deviations and spatial intervals no higher than the upper limit of spatial intervals to the total number of nodes at that location is calculated, and this ratio is used as the coupling coefficient of that node; then, the coupling coefficients of all nodes are mapped onto the image plane according to their spatial coordinate positions, thereby forming a coupling coefficient map.

[0118] The coupling coefficient represents the degree of consistency between the local afterimage trajectory and the branch direction, while the coupling coefficient diagram is a visual representation of the overall spatial distribution of these coupling coefficients, used to reflect the global characteristics of the matching between the movement of branches and leaves and the branch structure.

[0119] The process of forming a chain of evidence demonstrating resilience and flexibility also includes the following steps:

[0120] S4-5. While keeping the spatial node sequence of the afterimage trajectory unchanged, the original time index sequence is shuffled to obtain the counterfactual sequence of time order perturbation. The projection matching and statistical process of S4-4 is reused to generate a comparison coupling coefficient map. The coupling coefficient map is compared with the comparison coupling coefficient map position by position to identify the afterimage trajectory segments that still maintain a high coupling coefficient of the target after the time order perturbation, and these segments are marked as environmental interference afterimages.

[0121] In S4-5, it should be noted that the identification and removal of environmental interference afterimages are achieved by constructing a counterfactual sequence of temporal perturbations and comparing it position-by-position with the original coupling coefficient map. Specifically, while keeping the spatial node sequence of the afterimage trajectory unchanged, the original time index sequence is randomly reordered so that the spatial positions of the trajectory nodes remain consistent but their temporal order is disrupted, thereby generating a counterfactual sequence of temporal perturbations. This sequence is used to simulate the hypothetical situation where there is no real branch and leaf movement law. Subsequently, using the counterfactual sequence as input, the projection matching and statistical process defined in S4-4 is reused, that is, each trajectory node is recalculated and... The spatial interval and directional deviation between the nodes of the corresponding branch centerline are calculated, and the proportion of nodes that satisfy the condition that the directional deviation is not higher than the upper limit of the directional deviation and the spatial interval is not higher than the upper limit of the spatial interval is statistically analyzed at the trajectory level to generate the corresponding reference coupling coefficient map. Then, the original coupling coefficient map and the reference coupling coefficient map are compared node by node under the same spatial coordinates. When a trajectory segment still maintains a coupling coefficient higher than the preset high coupling threshold after counterfactual perturbation, it is determined that the high coupling of the segment does not originate from the actual time series of branch and leaf movement, but is caused by stable external disturbances (such as fixed wind speed direction or background shaking). Therefore, the trajectory segment is marked as an environmental disturbance afterimage.

[0122] S4-6. Remove the environmental interference afterimage from the afterimage structure representation, re-statistically evaluate the three indicators of directional stability, deformation amplitude, and time duration on the afterimage trajectory after removal, generate a toughness and elasticity evidence chain consisting of trajectory identifier, directional stability, deformation amplitude, time duration, and corresponding spatial overlay mask, and output it.

[0123] It should be noted that directional stability is based on the brightness gradient directional angle of each node in the afterimage trajectory. It is calculated by averaging the difference in directional angle of the same node in adjacent frames, taking the absolute value, and then defining stability in an inverse proportional form, that is: directional stability = 1 ÷ (average value of the absolute value of the difference in directional angle between adjacent frames). The smaller the average difference, the more consistent the swaying direction of the branches and leaves, and the higher the directional stability value.

[0124] Deformation amplitude is based on the spatial position change of the trajectory between adjacent frames. First, the Euclidean distance between the coordinates of the same node in adjacent frames is calculated, and then the average value of the distance sequence of the entire trajectory is calculated. It is defined as: Deformation amplitude = average value of Euclidean distance between all nodes in adjacent frames. The smaller the value of deformation amplitude, the more slight the change in the spatial morphology of the branches and leaves, and the more stable the structure.

[0125] Temporal persistence is based on the frame coverage of the trajectory in the time index sequence. The number of consecutive frames in which the trajectory appears is counted and divided by the total number of frames. It is defined as: Temporal persistence = Number of consecutive frames of trajectory ÷ Total number of frames. This ratio is used to reflect the continuous and stable existence time of the trajectory in the complete image sequence. The higher the ratio, the more coherent the branch and leaf movement.

[0126] Additionally, the trajectory identifier is a number assigned to each trajectory by the system when generating the afterimage trajectory. It is used to distinguish the movement paths of different branches and leaves in subsequent calculations and comparisons, and to establish a correspondence with its related feature indicators (such as directional stability, deformation amplitude, and temporal persistence). The spatial coverage mask is a binary layer generated based on the coordinate range of all spatial nodes of the trajectory in the image plane. In this layer, the pixels in the trajectory coverage area are assigned a value of 1, and other areas are assigned a value of 0. It is used to represent the coverage range and distribution pattern of the afterimage of branches and leaves in space.

[0127] In S4-6, it is important to explain as a whole that, in practical applications, this involves performing removal and recalculation operations on the afterimage structure representation to construct a resilience and elasticity evidence chain that can quantitatively reflect the dynamic characteristics of branches and leaves. Specifically, firstly, using the trajectory identifiers of environmental interference afterimages marked in the previous step as indices, all spatial node sequences and temporal index sequences of the corresponding trajectories are precisely deleted from the afterimage structure representation to ensure that subsequent statistics are based only on valid trajectories formed by real branch and leaf movements. Subsequently, for each retained afterimage trajectory, feature statistics are re-executed according to the order of adjacent frames in its temporal index sequence: using the absolute difference of the brightness gradient direction angle of the same node in adjacent frames as input, the average of the direction angle differences of all nodes is calculated, and then its reciprocal is taken to obtain the direction stability; using the absolute difference of the brightness gradient direction angle of the same node in adjacent frames as input, the direction stability is obtained by calculating the average of the direction angle differences of all nodes. The deformation amplitude is obtained by averaging the Euclidean distance of the spatial coordinates of the nodes. The temporal persistence is obtained by dividing the number of consecutive frames of the trajectory by the total number of frames of the trajectory in the entire sequence. After calculating the three indicators, a unique trajectory identifier is retained for each trajectory, and a rectangular boundary region is generated based on the minimum and maximum coordinate values ​​of all spatial nodes of the trajectory. The pixel values ​​corresponding to the positions of all nodes in the region are set to 1, and the rest are set to 0, thus forming a spatial overlay mask. Finally, the trajectory identifier, directional stability, deformation amplitude, temporal persistence and spatial overlay mask of each trajectory are combined into a complete data unit. All data units are summarized in the order of trajectory identifier to construct and output the toughness and elasticity evidence chain for subsequent calculation of excellent single plants.

[0128] The process of generating and determining the excellence index of a single plant's afterimage includes the following steps:

[0129] S5-1. The three indicators of directional stability, deformation amplitude and time duration in the toughness and elasticity evidence chain are fused and calculated according to the trajectory identifier: the weighted value of deformation amplitude is subtracted from the weighted value of directional stability and the weighted value of time duration to generate an intermediate fusion value for measuring the stability and morphological recovery ability of branches and leaves, and the intermediate fusion value of all trajectories is normalized to form a single plant afterimage excellence index.

[0130] S5-2. Compare the single-plant afterimage quality index with the preset judgment threshold. When the single-plant afterimage quality index is not lower than the preset judgment threshold, output the judgment result that the corresponding tea tree single plant is a good single plant. When the single-plant afterimage quality index is lower than the judgment threshold, return to step S3 to re-execute the part division and consistency judgment process, and re-analyze and correct the afterimage trajectory to ensure the stability and reliability of the final good judgment.

[0131] It should be noted in S5-1 that by performing per-trajectory fusion calculations on the numerical structures of the toughness and elasticity evidence chains, a quantifiable goodness evaluation index is established. Specifically, the system first uses the trajectory identifier as an index to sequentially extract three indicators for each trajectory from the evidence chain: the direction stability, deformation amplitude, and time persistence. The direction stability is used to measure the stability of the movement direction of the branches and leaves, and the higher the value, the more constant the swinging direction. The time persistence reflects the proportion of the trajectory continuously existing in the time series, and the higher the value, the more persistent and stable the movement of the branches and leaves. The deformation amplitude represents the average change amount of the trajectory in the spatial form, and the smaller the value, the more complete the shape of the branches is maintained.

[0132] In the calculation process, to ensure that the relative contributions of each indicator conform to physiological characteristics, the system performs weighted fusion on the three indicators according to preset weights: multiply the direction stability by the weight W1, the time persistence by the weight W2, and the deformation amplitude by the weight W3, and calculate the intermediate fusion value F = (W1 × direction stability + W2 × time persistence - W3 × deformation amplitude). Subsequently, a normalization operation is performed on the F values of all trajectories. Taking the maximum and minimum values of all intermediate fusion values as the interval endpoints, it is adjusted to the standardized range of 0 to 1 using a linear mapping method. The result after normalization is used as the excellent index of the single-plant afterimage, which is used to reflect the dynamic balance ability and morphological recovery ability of the branches and leaves of the entire tea plant.

[0133] It should be noted in S5-2 that by performing quantitative determination and dynamic return processing on the excellent index of the single-plant afterimage, the adaptive recognition of excellent single plants is achieved. Specifically, the system numerically compares the calculated excellent index of the single-plant afterimage with the preset determination threshold T. When the excellent index ≥ T, it means that the branches and leaves of this single plant have high direction stability, low deformation amplitude, and strong time persistence, and the system directly outputs the determination result of "the target tea tree single plant is an excellent single plant". When the excellent index < T, the system triggers the return logic and automatically returns to step S3 to perform the part division and consistency determination process again, and re-corrects the trajectory stratification and regional characteristics. This return mechanism ensures that in the case of determination deviations caused by changes in light, wind speed, or image noise, the system can maintain the stability and reliability of the excellent determination result through re-analysis and correction processes.

[0134] An image recognition system for screening excellent single tea plants includes an afterimage extraction module, a trajectory stratification module, a feature comparison module, an interference elimination module, and an excellent determination module.

[0135] The afterimage extraction module acquires the afterimage information formed by the branches and leaves under the action of wind disturbance by collecting multiple frames of images of the target tea tree under natural wind disturbance conditions, and then performs inter-frame difference processing on the multiple frames of images to form an initial afterimage sequence containing afterimage trajectories.

[0136] The trajectory layering module is used to perform layering processing on the initial afterimage sequence to form low-frequency afterimage information and high-frequency afterimage information, and to identify the continuous direction of the branch and leaf afterimage trajectory in the low-frequency afterimage information and the high-frequency afterimage information, thereby outputting the afterimage structure representation;

[0137] The feature comparison module divides the afterimage structure representation into parts, extracting the crown region, lateral branch region, and lower leaf region respectively, and then compares the afterimage features of each region to form a consistency judgment result of the afterimage features within the plant.

[0138] The interference removal module is used to perform coupled analysis on the consistency judgment result and the branch orientation information solved based on the static image. At the same time, it constructs a counterfactual sequence generated by time-series perturbation and performs comparison to solve the environmental interference afterimage and remove it from the afterimage structure representation, thereby forming a tough and elastic evidence chain.

[0139] The excellent judgment module is used to perform fusion processing on the toughness and elasticity evidence chain to generate a single-plant afterimage excellent index. When the single-plant afterimage excellent index meets the preset judgment conditions, the target tea tree single plant is output as an excellent single plant. When the judgment conditions are not met, the feature comparison module is returned to re-execute.

[0140] The working principle of this scheme is as follows: First, multiple frames of images of the target tea tree are continuously acquired under natural wind disturbance. The brightness and color changes of the branches and leaves caused by wind are extracted as afterimage information through the inter-frame differences. After connectivity and time matching, an initial afterimage sequence containing continuous motion paths is obtained.

[0141] Subsequently, the sequence was decomposed into two types of information according to the time window: slow change (overall sway) and fast change (local jitter). These were then aligned and superimposed in the same coordinate system. The afterimage trajectory was identified by connectivity filtering and ridge line refinement. The trajectory was then structured into an afterimage structure representation (including time index sequence, spatial node sequence, direction sequence, low / high frequency label distribution, coverage area mask, and breakpoint annotation). Based on this, the spatial nodes were divided into three regions according to the vertical coordinate: crown, lateral branches, and lower leaves. The trajectory nodes of the same frame in each region were weighted and aggregated to form a regional afterimage feature vector. The feature index set was constructed using the mean of brightness gradient, direction sequence similarity, and low / high frequency label ratio. A difference matrix was generated by comparing each region pairwise and the average deviation was calculated by normalization. This resulted in the consistency judgment result of the afterimage features within the plant.

[0142] Next, the branch centerline is extracted from the static image and the branch direction information is generated. Spatial registration is completed in the same coordinate system as the afterimage structure. A one-to-one correspondence between trajectory nodes and branch centerlines is established as a spatial coupling reference. The spatial interval and directional deviation to the centerline are calculated for each trajectory node along the time index. The proportion that meets the upper limit condition is statistically analyzed and mapped to the coupling coefficient map.

[0143] To verify the true temporal correlation, the time index is shuffled to construct a counterfactual sequence, and the same projection matching process is reused to generate a comparison coupling coefficient map. The two maps are compared position by position, and segments that still maintain high coupling after perturbation are marked as environmental interference afterimages and removed from the afterimage structure representation. The directional stability, deformation amplitude, and temporal persistence of the removed trajectories are recalculated, and combined with trajectory markers and spatial overlay masks to form a toughness and elasticity evidence chain. Finally, the trajectory indicators in the evidence chain are fused and normalized according to preset weights to form a single-plant afterimage excellence index. This index is compared with a preset judgment threshold. If the index is not lower than the threshold, the tea tree single plant is output as an excellent single plant. Otherwise, it returns to the part division and consistency judgment stage for re-analysis and correction, thus forming the operation process of "afterimage extraction - hierarchical identification - regional self-comparison - structural coupling and counterfactual noise removal - evidence chain fusion judgment".

[0144] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image recognition method for screening of elite individual tea plants, characterized in that, The method comprises the following steps: S1, by collecting multiple frames of images of target tea trees under natural wind disturbance conditions, obtaining residual shadow information formed by branches and leaves under the action of wind disturbance, and then performing inter-frame difference processing on the multiple frames of images to form an initial residual shadow sequence containing residual shadow tracks; S2, performing hierarchical processing on the initial residual shadow sequence to form low-frequency residual shadow information and high-frequency residual shadow information, and identifying the continuous direction of the branch and leaf residual shadow track in the low-frequency residual shadow information and the high-frequency residual shadow information, thereby outputting a residual shadow structure representation; S3, by dividing the residual shadow structure representation into parts, the crown area, the lateral branch area and the lower leaf area are respectively solved, and then the residual shadow characteristics of each area are compared to form a consistency determination result of the residual shadow characteristics inside the plant; S4, coupling analysis is performed on the consistency determination result and the branch direction information solved based on the static image, and a counterfactual sequence generated by time sequence disturbance is constructed and compared to solve the environmental disturbance residual shadow and remove it from the residual shadow structure representation, thereby forming a resilience and elasticity evidence chain; S41, by extracting a green channel image in the RGB color space of the static image and performing contrast stretching and noise suppression, an enhanced image for structure extraction is obtained; Edge detection, connected domain extraction and skeleton thinning are performed on the enhanced image to construct a branch center line candidate set composed of connected curves; S42, the branch center line candidate set is screened according to the length lower limit and the curvature change rate upper limit, and curves with a length not lower than the length lower limit and a curvature change rate not higher than the curvature change rate upper limit are retained; the tangential direction is calculated point by point along the spatial node sequence of each retained curve to form branch direction information; S43, the branch direction information and the residual shadow structure representation corresponding to the consistency determination result are subjected to spatial registration processing in the same image coordinate system, so that the spatial node sequence of the residual shadow track and the spatial node sequence of the branch direction information establish a one-to-one correspondence relationship, forming a spatial coupling reference; S44, on the spatial coupling reference, projection matching is performed on each residual shadow track node by node along its time index sequence to generate a coupling coefficient map; S45, under the condition of keeping the spatial node sequence of the residual shadow track unchanged, the original time index sequence is disturbed to obtain a counterfactual sequence of time sequence disturbance, and the projection matching and statistical process of S44 are reused to generate a control coupling coefficient map; The coupling coefficient map and the control coupling coefficient map are compared position by position to identify the residual shadow track segment that still maintains the target coupling coefficient after the time sequence disturbance, and mark it as environmental disturbance residual shadow; S46, the environmental disturbance residual shadow is removed from the residual shadow structure representation, and the three indexes of direction stability, deformation amplitude and time persistence are recalculated for the residual shadow track after removal, to generate a resilience and elasticity evidence chain and output; S5, the resilience and elasticity evidence chain is subjected to fusion processing to generate a single-plant residual shadow excellent index, and when the single-plant residual shadow excellent index meets the preset determination condition, the target tea tree single plant is output as an excellent single plant, and when the determination condition is not met, it returns to S3 for re-execution.

2. The image recognition method for screening excellent tea plants according to claim 1, characterized in that: The process of forming an initial residual image sequence containing residual image tracks comprises the following steps: S11. Continuously collecting multiple frames of images of the target tea plant at a fixed sampling interval by an imaging device, dividing each frame of image into a pixel matrix, and obtaining the brightness value and color value in the pixel matrix for constructing a residual image information base set containing branch and leaf edges, leaf vein details, and local shadows; S12. Performing pixel difference operation on adjacent frames of images in the residual image information base set one by one, calculating the brightness difference and color difference of each pixel point, and outputting a pixel difference matrix; S13. Performing threshold segmentation on the pixel difference matrix, identifying the pixel point set that exceeds the threshold, forming a residual image candidate region, and recording the directionality of pixel changes in the residual image candidate region; S14. Performing connectivity analysis on the residual image candidate region, combining adjacent pixel points into residual image segments according to spatial connectivity rules, and then performing track matching on residual image segments of adjacent frames based on time sequence to form continuous branch and leaf movement paths; S15. Arranging the continuous branch and leaf movement paths in time sequence to output an initial residual image sequence containing residual image tracks.

3. The image recognition method for screening excellent tea plants according to claim 2, characterized in that: In S2, it comprises: S21. Iterating the initial residual image sequence by a fixed length time sliding window, generating a slow variable response map by performing moving average operation on the multiple frame brightness values and color values of the same pixel within the window, and generating a fast variable energy map by performing absolute value accumulation on the brightness difference and color difference between the same pixel in adjacent frames; S22. Performing preset threshold judgment on the slow variable response map and the fast variable energy map to form a hierarchical result: when the fast variable energy is less than the upper limit of the preset threshold and the moving average change within the window is greater than the lower limit of the preset threshold, the corresponding pixel is labeled as a low-frequency residual image pixel, thereby generating low-frequency residual image information; when the state that the fast variable energy is greater than the lower limit of the preset threshold reaches the preset lower limit in the count of consecutive frames, the corresponding pixel is labeled as a high-frequency residual image pixel, thereby generating high-frequency residual image information; S23. Performing spatial alignment and superposition mapping of the low-frequency residual image information and the high-frequency residual image information in the same coordinate system, identifying the ridge line pixel set with continuous brightness changes by performing connectivity screening on the superposition region and its one-pixel neighborhood, and performing thinning and denoising on the ridge line pixel set to form a candidate track seed point set; S24. Performing continuous direction recognition on the candidate track seed point set: for each candidate track seed point, counting the gradient direction and direction consistency index within its local window, performing pixel-level growth in the current frame along the main direction and searching for a continuation point within the set radius neighborhood in the next frame in time sequence, confirming the connection when the direction deviation does not exceed the upper limit of the preset angle, the spatial interval does not exceed the upper limit of the pixel, and the brightness continuity meets the preset lower limit, otherwise terminating the current branch, thereby forming a time-ordered branch and leaf residual image track set. S25, performing structured combination on the branch and leaf residual trajectory set and outputting a residual structure representation, the residual structure representation comprising: a time index sequence of each trajectory, a spatial node sequence of each trajectory, a direction sequence of each trajectory, corresponding low-frequency and high-frequency label distributions, an overlay area mask, and a breakpoint position annotation.

4. The image recognition method for screening excellent tea plant individual according to claim 3, characterized in that: In S3, the determination process of the consistency determination result of the residual feature in the plant interior comprises the following steps: S31, performing spatial layering processing on the spatial node sequence of the residual structure representation according to the coordinate value range in the vertical direction, defining the upper third of the spatial nodes as the crown region, the middle third of the spatial nodes as the lateral branch region, and the lower third of the spatial nodes as the lower leaf region, thereby completing the division of the crown region, the lateral branch region, and the lower leaf region; S32, in each of the crown region, the lateral branch region, and the lower leaf region, performing inter-frame aggregation on the branch and leaf residual trajectory according to the time index sequence, and performing weighted average operation on the trajectory nodes of the same frame in the same region to generate a regional residual feature vector representing the overall swing trend of the branches and leaves in the region; S33, performing standardization processing on the brightness gradient mean value, the direction sequence similarity, and the low-frequency and high-frequency label proportion contained in each regional residual feature vector to form a feature index set for inter-regional comparison.

5. The image recognition method for screening excellent tea plant individual according to claim 4, characterized in that: In S3, the determination process of the consistency determination result of the residual feature in the plant interior further comprises the following steps: S34, taking the feature index set of the crown region as a reference, performing Euclidean distance calculation and direction deviation calculation on the feature index sets of the lateral branch region and the lower leaf region respectively, and generating a difference matrix between the regions; S35, performing normalization operation on the difference matrix to calculate the average deviation amount, and outputting the consistency determination result of the residual feature in the plant interior as consistent state when the average deviation amount is lower than the preset threshold, otherwise outputting as inconsistent state.

6. The image recognition method for screening excellent tea plant individual according to claim 5, characterized in that: In the process of generating the individual residual excellent index and making the determination, the following steps are included: S51, fusing and calculating the direction stability, the deformation amplitude, and the time persistence in the toughness and elasticity evidence chain according to the trajectory identification, taking the sum of the weighted value of the direction stability and the weighted value of the time persistence minus the weighted value of the deformation amplitude to generate an intermediate fusion value, and performing normalization processing on the intermediate fusion values of all trajectories to form the individual residual excellent index; S52, comparing the individual residual excellent index with the preset determination threshold, outputting the determination result that the corresponding tea plant individual is an excellent individual when the individual residual excellent index is not lower than the preset determination threshold, and returning to step S3 for re-execution when the individual residual excellent index is lower than the determination threshold.

7. The image recognition method for tea plant excellent individual screening according to claim 1, characterized in that: In S42, the branch direction information is composed of curve identification, spatial node sequence and corresponding tangent direction.

8. The image recognition method for tea plant excellent individual screening according to claim 1, characterized in that: In S44, the process of generating coupling coefficient map by performing projection matching includes: taking the current residual shadow track node as input, solving the spatial interval of the node to the corresponding branch center line and the directional deviation of the node direction and the corresponding tangent direction; at the track level, the proportion of nodes with directional deviation not higher than the upper limit of directional deviation and spatial interval not higher than the upper limit of spatial interval is counted to generate the coupling coefficient map.

9. The image recognition method for tea plant excellent individual screening according to claim 1, characterized in that: In S46, the flexibility and elasticity evidence chain is composed of track identification, direction stability, deformation amplitude, time duration and corresponding spatial coverage mask.

10. An image recognition system for tea plant excellent individual screening, comprising a residual shadow extraction module, a track layering module, a feature comparison module, an interference elimination module and an excellent judgment module, characterized in that: The residual shadow extraction module acquires residual shadow information formed by branches and leaves under the action of wind disturbance by collecting multiple images of the target tea plant under natural wind disturbance, and then performs inter-frame difference processing on the multiple images to form an initial residual shadow sequence containing residual shadow tracks; The track layering module is used to perform layering processing on the initial residual shadow sequence to form low-frequency residual shadow information and high-frequency residual shadow information, and identify the continuous direction of branch and leaf residual shadow tracks in the low-frequency residual shadow information and the high-frequency residual shadow information, thereby outputting a residual shadow structure representation; The feature comparison module divides the residual shadow structure representation into parts, and then solves the crown area, lateral branch area and lower leaf area respectively, and then performs comparison on the residual shadow features of each area to form a consistency judgment result of the residual shadow features inside the plant; The interference elimination module is used to perform coupling analysis on the consistency judgment result and the branch direction information solved based on the static image, construct an anti-fact sequence generated by time sequence disturbance and perform comparison to solve environmental interference residual shadow and eliminate it from the residual shadow structure representation, thereby forming a flexibility and elasticity evidence chain; S41, an enhanced image for structure extraction is obtained by extracting a green channel image in the RGB color space of a static image and performing contrast stretching and noise suppression; Edge detection, connected domain extraction and skeleton thinning are performed on the enhanced image to construct a branch center line candidate set composed of connected curves; S42, the branch center line candidate set is screened according to the length lower limit and the curvature change rate upper limit, and curves with length not lower than the length lower limit and curvature change rate not higher than the curvature change rate upper limit are retained; the tangent direction is calculated along the spatial node sequence of each retained curve point by point to form branch direction information. S43, the branch direction information and the corresponding residual shadow structure of the consistency determination result is expressed in the same image coordinate system to perform spatial registration processing, so that the spatial node sequence of the residual shadow track and the spatial node sequence of the branch direction information establish one-to-one correspondence, and form a spatial coupling reference; S44, on the spatial coupling reference, the projection matching is performed on each residual shadow track along the time index sequence of the residual shadow track, and a coupling coefficient map is generated; S45, under the condition of keeping the spatial node sequence of the residual shadow track unchanged, the original time index sequence is disturbed to obtain an anti-fact sequence of time sequence disturbance, and the projection matching and the statistical process of S44 are reused to generate a contrast coupling coefficient map; The coupling coefficient map and the contrast coupling coefficient map are compared position by position, the residual shadow track segment that still maintains the target coupling coefficient after the time sequence disturbance is identified, and is marked as an environmental interference residual shadow; S46, the environmental interference residual shadow is removed from the residual shadow structure representation, the direction stability, the deformation amplitude and the time persistence of the residual shadow track after the removal are re-counted, a toughness and elasticity evidence chain is generated, and output is generated; The excellent determination module is used for performing fusion processing on the toughness and elasticity evidence chain, generating a single-plant residual shadow excellent index, and outputting a target tea tree single plant as an excellent single plant when the single-plant residual shadow excellent index meets a preset determination condition, and returning to the feature comparison module for re-execution when the determination condition is not met.

Images