Seedling growth state image intelligent monitoring and quality grading method and system

Abstract

The invention discloses a seedling growth state image intelligent monitoring and quality grading method and system, and relates to the field of image monitoring. Synchronously acquiring a multi-view image sequence, an illumination parameter vector and a leaf surface reflection vector; identifying an overlapping region proportion and a contour point displacement vector of a seedling contour between adjacent multi-view image frames under different views, and constructing an overlapping difference atlas; performing cross-view-angle continuous tracking on the target seedling contour in the multi-view-angle image to generate a cross-view-angle structure image set; calculating a boundary stability mean value, a change standard deviation and a contour matching rate of the target seedling structure between adjacent visual angles, and generating a structure stability label; performing compensation processing on the boundary region outside the preset stable interval to generate a candidate structure boundary set of the single seedling; screening each candidate structure boundary to generate a verification structure image set; and extracting a morphological feature data set of the target seedling according to the verification structure image set, and generating a seedling quality grade. And the accuracy and the stability of growth state evaluation and quality grading results are improved.

Description

Technical Field

[0001] This invention relates to the field of image monitoring, specifically to a method and system for intelligent monitoring and quality grading of seedling growth status images. Background Technology

[0002] With the development of image sensor technology, multispectral imaging technology, and data analysis algorithms, image-based non-contact seedling detection methods have gradually replaced traditional manual visual inspection and become the mainstream seedling grading method. By setting up an imaging device above the seedling tray to acquire seedling image information and combining it with characterizing features such as color, morphology, and structure to determine its growth status, this has become a typical intelligent seedling monitoring method.

[0003] In real-world environments where seedling trays are densely packed, existing image-based monitoring methods often fail to accurately distinguish the independent boundaries of each seedling when identifying and separating multiple seedlings in an image. This makes it difficult to clearly extract the structure of a single seedling, which in turn affects the reliability of subsequent growth status assessment and grading. Summary of the Invention

[0004] The purpose of this application is to provide an intelligent monitoring and quality grading method for seedling growth status images. This method effectively reduces the interference of mutual occlusion of multiple seedlings and uneven lighting in high-density seedling tray environments on the stability of single-plant boundary recognition and structure extraction in images, improves the accuracy of single-plant structure separation and morphological feature extraction, and thus enhances the reliability of seedling quality grading results.

[0005] The objective of this application can be achieved through the following technical solution: Firstly, a method for intelligent monitoring and quality grading of seedling growth status images, comprising the following steps:

[0006] Simultaneously acquire multi-view image sequences, light parameter vectors, and leaf reflectance vectors of the target seedling area;

[0007] Based on the multi-view image sequence and the illumination parameter vector, the overlapping area ratio and contour point displacement vector of the seedling outline between adjacent multi-view image frames under different views are identified, and an overlap difference map set is constructed.

[0008] The overlapping difference map is used to continuously track the outline of the target seedling in the multi-view image across different perspectives, and generate a cross-view structure map of the seedling outline.

[0009] Based on the cross-view structure atlas, the boundary stability mean, standard deviation of variation and contour matching rate of the target seedling structure between adjacent views are calculated. Based on the boundary stability mean, standard deviation of variation and contour matching rate within the preset stability interval, a structure stability label is generated.

[0010] By using the cross-view structure diagram and the structure stability label, compensation processing is performed on the boundary region outside the preset stability interval to generate a candidate structure boundary set for a single seedling.

[0011] By comparing the similarity between the reflection direction distribution of the candidate structure boundary and the leaf surface reflection vector, each candidate structure boundary is filtered to generate a set of verification structure images.

[0012] Based on the verification structure image set, extract the morphological feature dataset of the target seedling, and combine it with the structural stability label and the similarity to map it to a preset growth state evaluation interval to generate the seedling quality grade.

[0013] Secondly, the seedling growth status image intelligent monitoring and quality grading system includes the following modules:

[0014] The image monitoring module is used to simultaneously acquire multi-view image sequences, light parameter vectors, and leaf reflectance vectors of the target seedling area;

[0015] The image difference module is used to identify the proportion of overlapping areas and the displacement vector of contour points of seedling outlines under different perspectives between adjacent multi-view image frames based on the multi-view image sequence and the illumination parameter vector, and to construct an overlapping difference map set.

[0016] The image contour module is used to continuously track the contours of target seedlings in multi-view images across different perspectives using the overlapping difference map set, and generate a cross-view structure map set of seedling contours.

[0017] The image contour analysis module is used to calculate the boundary stability mean, variation standard deviation and contour matching rate of the target seedling structure between adjacent views based on the cross-view structure atlas, and generate a structure stability label based on the boundary stability mean, variation standard deviation and contour matching rate within a preset stability interval.

[0018] The image structure module is used to compensate the boundary region outside the preset stability interval by using the cross-view structure map and the structure stability label to generate a candidate structure boundary set for a single seedling.

[0019] The image verification module is used to filter each candidate structure boundary by comparing the similarity between the reflection direction distribution of the candidate structure boundary and the reflection vector of the leaf surface, and generate a set of verification structure images.

[0020] The quality grading module is used to extract the morphological feature dataset of the target seedling based on the verification structure image set, and combine the structure stability label and the similarity to map it to a preset growth state evaluation interval to generate the seedling quality grade.

[0021] Compared with the prior art, the beneficial effects of this application are:

[0022] 1. In this invention, by constructing a cross-view structural map of seedling contours in a multi-view image sequence, and calculating the boundary stability mean, variation standard deviation and contour matching rate, structural stability labels are generated, which effectively enhances the boundary continuity recognition ability of single seedlings under dense placement and varied posture conditions, and improves the accuracy of contour separation.

[0023] 2. In this invention, by fusing structural stability labels and the similarity of reflection direction distribution of candidate boundaries, the boundary region is jointly screened and verified for reconstruction, generating a set of verification structural images. This achieves accurate extraction of single-plant structures under complex lighting interference, effectively improving the stability and reliability of subsequent growth status assessment and quality grade determination.

[0024] 3. In this invention, by constructing the proportion of contour overlap area and boundary point displacement vector between multi-view image frames, an overlap difference map is formed, which can effectively reflect the trend of seedling boundary change under different viewpoints, provide dynamic basis for contour tracking and error compensation, and enhance the stability of boundary recognition under various postures. Attached Figure Description

[0025] Figure 1 This is a flowchart of the steps of the present invention;

[0026] Figure 2 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0027] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0028] Example 1:

[0029] In traditional seedling image acquisition and growth status grading methods, when faced with high-density seedling tray environments, severe occlusion, shadow blending, and color interference often occur in the images due to the close spatial distribution of multiple seedlings, inconsistent leaf orientation, and complex natural or artificial lighting conditions. This results in blurred boundary information for individual seedlings. Such visual interference significantly reduces the image segmentation algorithm's ability to recognize individual plant structures, further affecting the accuracy of quantitative assessment of seedling growth status and quality grade classification.

[0030] Existing technologies typically rely on single-frame images for edge extraction and feature recognition, lacking modeling of spatial consistency across multi-view images and occlusion compensation mechanisms. They also fail to explicitly model the interference of illumination variation parameters on occluded areas. Especially in greenhouse environments, when oblique light, reflected interference, or localized shadows are present, the contour structure of the same seedling will exhibit inconsistency at different angles. Current image processing methods fail to establish cross-view structural connectivity, leading to structural fusion or recognition breaks among multiple seedlings in the image, severely interfering with morphological feature extraction and growth level assessment.

[0031] For example, during image acquisition of seedling trays at a smart agriculture experimental station, a mobile multi-angle imaging device was used to capture image sequences from the side and above to improve individual plant recognition. However, because some leaves were in shadow or obscured at multiple angles, traditional methods mistakenly identified multiple plants as a single target or a single plant as multiple discontinuous structures, leading to distortions in characteristic parameters such as plant height and leaf width. More seriously, due to the inability to construct a reliable structural compensation model, morphological characteristics exhibited extreme fluctuations under light disturbances, causing abrupt changes in growth status assessment results for the same seedling, undermining the consistency and reliability of the system's quality evaluation.

[0032] Without addressing the aforementioned issues, image-based intelligent monitoring technology for seedling growth will be difficult to apply to high-density seedling environments. Incorrect structural identification results may mislead agricultural intelligent management systems, misclassifying weak plants as healthy or normal plants as unhealthy, leading to inaccurate seedling management decisions and impacting seedling survival rates and subsequent transplanting outcomes. In the long term, the instability of this technology will limit its widespread application in smart agriculture scenarios, forcing producers to still rely on manual visual inspection and experience-based judgment, thus hindering the achievement of large-scale automated and standardized seedling quality management goals.

[0033] like Figure 1 As shown, the intelligent monitoring and quality grading method for seedling growth status images includes the following steps:

[0034] Simultaneously acquire multi-view image sequences, light parameter vectors, and leaf reflectance vectors of the target seedling area;

[0035] Based on the multi-view image sequence and the illumination parameter vector, the overlapping area ratio and contour point displacement vector of the seedling outline between adjacent multi-view image frames under different views are identified, and an overlap difference map set is constructed.

[0036] The overlapping difference map is used to continuously track the outline of the target seedling in the multi-view image across different perspectives, and generate a cross-view structure map of the seedling outline.

[0037] Based on the cross-view structure atlas, the boundary stability mean, standard deviation of variation and contour matching rate of the target seedling structure between adjacent views are calculated. Based on the boundary stability mean, standard deviation of variation and contour matching rate within the preset stability interval, a structure stability label is generated.

[0038] By using the cross-view structure diagram and the structure stability label, compensation processing is performed on the boundary region outside the preset stability interval to generate a candidate structure boundary set for a single seedling.

[0039] By comparing the similarity between the reflection direction distribution of the candidate structure boundary and the leaf surface reflection vector, each candidate structure boundary is filtered to generate a set of verification structure images.

[0040] Based on the verification structure image set, extract the morphological feature dataset of the target seedling, and combine it with the structural stability label and the similarity to map it to a preset growth state evaluation interval to generate the seedling quality grade.

[0041] This invention utilizes a structural stability recognition mechanism that integrates multi-view images and light reflection information to identify and assess the structure and growth quality of individual seedlings in high-density seedling trays. First, multi-view image sequences, light parameter vectors, and leaf reflectance vectors of the target seedling area are acquired simultaneously, providing a basis for light consistency calibration for subsequent identification. Based on this, the light parameter vectors are combined to identify the proportion of overlapping areas and the displacement vectors of contour points in adjacent multi-view image frames under different viewpoints, constructing an overlap difference map reflecting changes in structural stability.

[0042] Furthermore, the overlapping difference atlas is used to continuously track the outline of the target seedling across multiple viewpoints in the multi-view images, generating a cross-viewpoint structural atlas with spatiotemporal consistency. Based on this structural atlas, the boundary stability mean, standard deviation of variation, and outline matching rate between adjacent viewpoints are extracted, and these are compared with a preset stability interval to form a structural stability label reflecting the clarity and consistency of the structure.

[0043] Subsequently, based on the structural labels and structural atlas, structural compensation is performed on boundary regions with insufficient stability to generate a candidate structural boundary set with complete structural contour representation. To enhance the accuracy of structural recognition, the similarity between the candidate structural boundaries and the actual leaf surface reflection vectors in terms of directional reflection distribution is further compared, and regions that meet the spectral reflectance consistency condition are selected, ultimately generating a verification structural image set.

[0044] Based on this, the seedling morphological feature dataset is extracted from the verification structure image, and combined with the structural stability level and reflection similarity, it is mapped to a preset growth state evaluation interval, thereby obtaining the growth quality level result of the target seedling.

[0045] Through the above-mentioned multi-view fusion and stability calibration mechanism, the present invention can improve the accuracy of structural identification and quality assessment of individual seedlings in high-density seedling trays. Especially under complex environments such as uneven lighting, overlapping leaves and posture interference, it can still achieve stable and reliable structural contour extraction and grade determination.

[0046] The core innovation of this application lies in constructing a method for identifying the structural stability and grading the quality of seedling images by integrating multi-source information. The overall technical solution revolves around the synchronous acquisition of multi-view image sequences, illumination parameter vectors, and leaf reflection vectors. First, by constructing an overlapping difference map set between multi-view image frames, the cross-viewpoint displacement relationship of the seedling contour is extracted, and an illumination perturbation correction mechanism is introduced to effectively reduce contour errors caused by different shooting angles and illumination conditions. Subsequently, a structural map set is generated through continuous cross-view tracking, and boundary stability indices (including boundary stability mean, standard deviation of variation, and contour matching rate) are further calculated and mapped to a preset stability interval to establish a quantifiable structural stability label. Based on this, by combining the structural stability label with the similarity of reflection direction distribution, candidate boundary structures with highly consistent physical properties are selected to form a verification structural image set, ensuring the accuracy and representativeness of subsequent morphological feature extraction. Finally, the extracted morphological feature data, stability level, and reflection similarity are mapped together to the growth state evaluation interval to complete the quality grade determination of the target seedling. A fusion recognition process based on multi-view contour stability, spectral reflectance features and structural compensation correction was realized in complex seedling raising scenarios, which significantly improved the stability, accuracy and physical consistency of single seedling grading assessment.

[0047] This application further proposes that the specific steps for simultaneously acquiring multi-view image sequences, light parameter vectors, and leaf reflectance vectors of the target seedling area include:

[0048] The purpose of acquiring image data of the target seedling tray from multiple shooting angles and the shooting time is to collect image data covering the target seedling tray from different angles to ensure sufficient geometric depth information and structural details of the occluded areas. Industrial cameras mounted on tracks or robotic arms can be used for automatic multi-view shooting, with preset viewing angle intervals (e.g., 10°~20°). The acquired images are typically in RGB format (resolution recommended to be no less than 1920×1080) to meet the accuracy requirements of subsequent structural boundary analysis.

[0049] Image data from the same shooting angle are arranged sequentially according to the shooting time to obtain image data groups. These image data groups are then arranged sequentially according to the shooting angle from smallest to largest to form a multi-view image sequence.

[0050] Acquire ambient lighting parameter data for each shooting angle, including the incident angle of the light source, azimuth angle, illuminance intensity, and time stamp for each shooting angle. These parameters can be recorded in real time by synchronously installed illuminance meters or environmental sensors. The ambient lighting parameter data is used to construct a lighting parameter vector. ,in, The incident angle of the light source, It is the azimuth angle. Illuminance intensity, This refers to the shooting time.

[0051] Acquire multi-channel reflectance spectral data to construct leaf reflectance vector sets. ,in, This represents the reflectance value in the nth spectral channel; it is typically acquired using a multi-channel hyperspectral camera or multi-channel sensor, including channels for blue, green, red, and near-infrared light. The reflectance spectra of each channel reflect the leaf surface structure and health status of the seedlings and can be used to verify the physical consistency of the identification results.

[0052] For example, when photographing a 4×6 seedling tray, eight different viewing angles were selected, and a high-resolution image was captured from each angle. Simultaneously, the illuminance at each viewing angle was recorded as 870 lx, the incident angle as 45°, the azimuth as 60°, and the time as 16:23:07. Reflectance was collected using a multi-channel sensor in the blue, green, red, and near-infrared bands, forming a set of 4D leaf surface reflectance vectors, for example, R=[0.23,0.45,0.52,0.76]. These data will collectively constitute the basic input for structural stability identification and quality level assessment.

[0053] This application's solution simultaneously collects image data of the target seedling tray from multiple shooting angles, ambient light parameter data corresponding to each angle, and multi-channel reflectance spectral data. This constructs a multi-view image sequence, light parameter vectors, and leaf reflectance vector sets, forming a unified data foundation for structure recognition and verification. The multi-view image sequence enables complete and continuous modeling of the seedling structure in space, improving the coherence and completeness of contour extraction. The incident direction, illuminance value, and shooting timestamp included in the light parameter vector can be used to model the light perturbation effect in the image, providing quantitative support for subsequent contour point displacement correction, occlusion removal, and identification of areas with abnormal brightness. The leaf reflectance vector set constructed from the multi-channel reflectance spectral data can be used to verify the physical authenticity of candidate boundaries, effectively eliminating misjudgments of non-vegetation areas. Simultaneously, the introduction of shooting timestamps ensures the synchronous correspondence of image data, light data, and spectral data in the time dimension, enhancing stability, accuracy, and adaptability in complex environments.

[0054] This application further proposes that the specific steps for identifying the proportion of overlapping areas and the displacement vectors of contour points of seedling outlines under different viewpoints between adjacent multi-view image frames include:

[0055] Time to obtain the i-th frame image Synchronous shooting time and viewing angle in the illumination parameter vector Build an index table ,in, Indicates the shooting perspective index. This represents the normalized illumination parameter vector of the i-th frame image.

[0056] Preset time tolerance threshold (For example, (Set within the range of 0.1s to 1.0s), when the following conditions are met:

[0057] , and At that time, among them, To preset the visual tolerance threshold (e.g., (Set within the range of 5° to 15°) For illumination parameter thresholds (e.g., (Set within the range of 0.05 to 0.15) to filter out image frames taken under different lighting conditions but at similar times, thereby reducing structural mismatches caused by lighting changes.

[0058] Frame i and frame j are considered as adjacent viewpoint matching frame pairs, generating an image frame index set. .

[0059] The original image data is loaded into each set of matched image frames, and a lightweight segmentation model based on a multi-scale attention mechanism (such as MSA-LiteSeg) is used to segment the images in each frame. Perform semantic segmentation and output a binary mask. , where x represents the horizontal pixel coordinate in the image; y represents the vertical pixel coordinate in the image; and a pixel with a value of 1 indicates that it belongs to the seedling boundary area.

[0060] Next, in the binary mask Contour extraction algorithms (such as Canny edge detection and boundary tracking) are used to extract contour point sets. ,in, This represents the set of contour points extracted from the i-th frame of the image; This represents the horizontal pixel coordinate of the k-th contour point; This represents the vertical pixel coordinate of the k-th contour point; and records the pixel gradient direction of each point. and grayscale gradient magnitude .

[0061] Preset spatial neighborhood radius (For example, 3 to 10 pixels), iterate through each point in the contour point set. Construct a local neighborhood centered on it:

[0062] ;in, It refers to Other contour points in the surrounding area;

[0063] For the set of all points that satisfy the adjacency condition The data is divided into a set of connected contour segments using density clustering algorithms (such as DBSCAN) or region growing methods, and the following three local feature parameters are calculated for the points in each group:

[0064] Local principal direction vector : Calculate the first principal axis direction of the current contour point set point based on principal component analysis (PCA);

[0065] Average gradient magnitude : Represents the points in a set of points The mean;

[0066] Boundary point density ,in, Represents a set of points; The preset spatial neighborhood radius; This represents the number of points within a unit neighborhood.

[0067] Each contour point group obtained through spatial clustering Store the corresponding frame's view profile subset: , where m is the total number of contour subsets in the i-th frame image; Represents a group of outline dots All contour pixels, along with their local geometric feature parameter sets. .

[0068] Contour subset for frame i In the (i+1)th frame, find the subset whose spatial location is closest and whose angle between their main direction vectors is less than a preset maximum deviation threshold angle. Where l represents the number of the l-th contour subset in the (i+1)-th frame, and potential matching pairs are constructed. Subsequently, the intersection-to-union ratio (IoU) under the standard projection plane is calculated as the proportion of the overlapping region. :

[0069] ;

[0070] in, This represents the overlap ratio of the contour regions between subset k in frame i and subset l in frame i+1. The "∩" and "∪" operations are implemented by superimposing the contour boundaries after image frame normalization. If the overlap threshold is greater than or equal to the preset threshold, then the pair of contour regions have a stable structural correspondence in both viewpoints.

[0071] Between successfully matched contour pairs, corresponding contour point pairs are established based on the perspective geometric mapping relationship between the two frames. ,in, This represents the coordinates of the j-th contour point in the i-th frame. Let represent the coordinates of the (j+1)th contour point in the i-th frame, and calculate its two-dimensional displacement vector:

[0072] ;

[0073] in, This represents the position coordinates of the j-th contour point in the i-th frame of the image; This represents the position coordinates of the corresponding matching point in the (i+1)th frame of the image; This represents the displacement of the j-th contour point between the i-th frame and the (i+1)-th frame;

[0074] Simultaneously record the magnitude of each set of displacement vectors. Direction angle Generate a set of contour point displacement vectors: ,in, This represents the set of displacement moduli of all matching points in the k-th contour subset (the l-th contour structure in frame i); This indicates the number of successfully paired boundary points in the current contour subset pair;

[0075] All matching contour subsets that meet the conditions and their corresponding... and The results are stored in the cross-view matching map.

[0076] The proposed solution improves the registration accuracy between multi-angle images by aligning timestamps and matching viewpoint indices between adjacent viewpoint image frames based on multi-view image sequences and illumination parameter vectors. This ensures temporal consistency and viewpoint comparability of observation data for the same seedling from different angles, reducing image mismatch issues caused by illumination changes or shooting delays. Furthermore, by extracting contour point sets from corresponding viewpoint image frames and performing clustering based on spatial neighborhood radii, the spatial aggregation and structural continuity of contour points under complex lighting conditions are significantly improved, avoiding boundary recognition failures due to background noise or contour fragmentation. Calculating the overlap ratio of contour regions and the displacement vector of contour points between adjacent viewpoints based on viewpoint contour subsets not only enables fine-grained tracking of the three-dimensional morphological changes of the target seedling but also quantifies structural consistency and boundary offset trends under different viewpoints, enhancing the perception and robustness of the seedling's multi-view structural behavior.

[0077] This application further proposes that the specific steps for generating the overlap difference atlas include:

[0078] Since changes in illumination may cause shifts in the grayscale gradient of target edges and abnormal contrast in an image, leading to mismatch of contour points, this embodiment addresses this issue.

[0079] Contour point set matching pairs extracted from adjacent viewpoint images The initial displacement vector is calculated as follows:

[0080] ;

[0081] However, due to variations in the incident light direction and illuminance level at different shooting times, distortion of the contour boundaries in the image plane may occur. Therefore, a perturbation weight correction function is introduced, and the illumination perturbation weight factor is calculated first:

[0082] ;

[0083] in, This is the weighting coefficient for the direction of illumination, and ; This is the weighting coefficient for light intensity, and ; Let be the normalized illumination direction vector of the i-th frame image; The normalized illumination direction vector of the (i+1)th frame image; The angle between the incident light direction and the incident light direction; Let be the average illuminance value of the region corresponding to the i-th frame image; This represents the average illumination value of the region corresponding to the (i+1)th frame of the image; The maximum illumination value occurring across all image frames is used, with the angle term expressed in radians. The initial displacement is then corrected for perturbation.

[0084] ;in, This is the corrected pixel displacement;

[0085] The set of corrected results for all point pairs is denoted as the corrected displacement dataset. .

[0086] Next, the proportion of overlapping regions in the contour areas is analyzed in the image space. Defined as:

[0087] ;

[0088] in, and These are the target contour regions extracted from the i-th and (i+1)-th frames, respectively. Contour region With outline area The pixel area of ​​the intersection region; The relative overlap ratio of the outlines of the two images; Refers to the outline area The pixel area; Refers to the outline area The pixel area.

[0089] Then, according to the preset mapping rules, which include a difference mapping function, the following will be implemented: and The joint mapping is a pixel-level difference weight matrix. Its weights are defined as follows:

[0090] ;

[0091] in, Image pixel coordinates, This is the corrected displacement vector for the contour points within the pixel region; The ratio of the overlapping area between the i-th frame and the (i+1)-th frame; For normalization parameters, .

[0092] Finally, the pixel-level difference weight matrix With the original image frame and Perform pixel-level spatial fusion to generate overlapping difference images Its color channels reflect the degree of difference.

[0093] Taking a certain data collection as an example, , , and The included angle is 25° (i.e., 0.436 rad), then Used for all After correction, the final overlapping difference image is generated. Regions with abnormal displacement exceeding 5px were identified across multiple areas, indicating occlusion differences.

[0094] This application's solution introduces an illumination perturbation correction mechanism and an overlapping region perception weighting strategy to dynamically correct target contour position errors caused by illumination changes and viewpoint shifts during multi-view image analysis. Based on this, a pixel-level difference weight matrix is ​​constructed to enhance the representation of real structural differences between regions. By spatially mapping and fusing this difference weight matrix with the original multi-view image sequence, the generated overlapping difference map accurately characterizes the contour change features and anomalous structural responses of the target region under adjacent viewpoints, thereby significantly improving the accuracy of occlusion reconstruction and the consistency of image quality after fusion.

[0095] This application further proposes that the specific steps for generating a cross-view structural atlas of seedling outlines include:

[0096] Processed overlap difference atlas Each view frame Canny edge detection and morphological processing were performed to extract the contour boundary point set of the corresponding target seedling. .

[0097] Each set of boundary points is appended with its frame index i and image acquisition timestamp, forming a contour frame relationship set. .

[0098] For each group of adjacent frames Neighborhood matching is performed on boundary points in the data. If a certain boundary point... Its matching point in the next frame The Euclidean distance satisfies:

[0099] ;in, This represents the Euclidean distance threshold for boundary matching between two contour points, in pixels, for example, set between 3 and 10 pixels; express The pixel coordinates of the ns-th contour boundary point in the image; express The pixel coordinates of the ms-th contour boundary point in the image;

[0100] And the boundary offset magnitude Not exceeding the preset offset tolerance threshold Then it is recorded as the cross-frame contour correspondence. .

[0101] If the frame pair view index difference is greater than or corresponding angle of view If the match is not found, it will be removed, resulting in a filtered set of contour matches. .

[0102] For each pair of boundary points after filtering Calculate their Euclidean distances respectively. Image gradient difference Construct the contour connection tensor:

[0103] ;

[0104] in, , Points , In the image , The local boundary gradient values ​​are calculated using the Sobel operator.

[0105] Based on tensors The boundary point connection relationship and the original contour frame relationship Perform node chain generation operations to form a cross-frame contour tracking path, for example:

[0106] ;

[0107] Each node chain This represents a sequence of boundary points that may belong to the same seedling structure.

[0108] Arrange all node chains into a graph structure according to frame order to form a structural graph atlas:

[0109] ;

[0110] in, This represents a set of structure graphs constructed based on the boundary point set and its inter-frame connectivity relationships from multiple frame perspectives; each structure graph... Nodes are boundary points in the image, and edges are inter-frame connections. The occluded structural regions are reconstructed and completed by using the connectivity of nodes in the structural graph.

[0111] This application's solution extracts overlapping difference maps from multi-view images, establishes contour frame relationships, and constructs a connectivity tensor. This effectively supplements the missing seedling contour information in a single viewpoint under high-density occlusion environments, improving the integrity and stability of structure recognition. Through viewpoint spacing constraints and boundary offset measurement mechanisms, mismatched frame pairs are eliminated, enhancing the accuracy and continuity of contour matching. Furthermore, the contour connectivity tensor is mapped to the frame order structure to generate a chain of connecting nodes, constructing a cross-view structure map. This effectively strengthens the spatial correlation and topological consistency of target contours between images, exhibiting high occlusion robustness and image domain structure reconstruction capabilities, thus improving recognition accuracy and practical application in complex seedling cultivation scenarios.

[0112] This application further proposes that the specific steps for calculating the boundary stability mean, standard deviation of variation, and contour matching rate of the target seedling structure between adjacent viewpoints include:

[0113] In environments with dense shading of seedlings, the overlapping of leaves and differences in light angles often result in incomplete outlines or unstable boundaries from a single viewpoint, affecting the analysis of seedling outline stability and subsequent individual plant identification. Therefore, this embodiment, based on the previously constructed cross-view structural atlas, further extracts and analyzes the boundary fluctuation characteristics of the target seedlings under different viewpoints to achieve a quantitative assessment of outline stability and separability.

[0114] First, the system obtains the set of contour boundary points of the target seedling in each view frame by using the cross-view connection node chain in the structure graph atlas. This yields the boundary point coordinate sequence for each frame. For these boundary points, the Euclidean distance change between corresponding boundary points in adjacent view frames is calculated according to the view index order of the node chain, forming a boundary fluctuation vector sequence. , where each vector represents the boundary drift between the current viewpoint and the next viewpoint.

[0115] Next, a local statistical analysis is performed on the aforementioned boundary fluctuation vector sequence using a preset sliding window size. For the fluctuation value within each sliding window segment, the boundary stability mean of that segment is calculated. Standard deviation of boundary variation It is used to characterize the consistency and dispersion of boundary fluctuations within the region.

[0116] To eliminate scale bias caused by differences in size and morphology among seedlings, further refinement of each value in the boundary fluctuation vector sequence was performed. Normalization is performed using either extremum normalization or z-score normalization to obtain the standardized fluctuation sequence. .

[0117] Finally, based on the normalization results and stability statistics, the contour matching rate is calculated. Defined as boundary fluctuations below a preset stability threshold across all frame pairs. The frame-to-frame ratio, i.e.:

[0118] ;

[0119] in, This represents the total number of image frames. The total number of image frame pairs; Let be the contour matching stability rate of the i-th target seedling across all its frame pairs;

[0120] This metric reflects the structural stability and matching degree of the seedling outline across multiple viewpoints. A higher outline matching rate indicates that the seedling's boundary is more stable across multiple viewpoints; conversely, a low matching rate suggests that the current target may have severe occlusion, boundary overlap, or other issues, requiring further processing or being marked as an untrusted region. This process can serve as an automatic screening method for seedling outline map quality, improving the reliability of hierarchical recognition.

[0121] The proposed solution constructs a set of boundary points for the target seedling in multiple viewpoint frames based on a cross-view structural atlas and extracts the boundary fluctuation vector sequence. This effectively captures the dynamic changes in the contour boundary of the same seedling under different viewpoints. By combining a sliding window mechanism to calculate the stable mean and standard deviation of the boundary, the system can identify the regularity and abruptness of boundary morphological fluctuations within local viewpoint intervals, improving the detection capability of abnormal contour disturbances. Furthermore, normalization processing eliminates amplitude differences caused by different viewpoints or shooting conditions, ensuring the comparability of boundary fluctuation data in global analysis. The final generated contour matching rate index can be used as a criterion for evaluating structural consistency and contour stability, providing a stable structural basis for subsequent spatial fusion, occlusion processing, and contour repair of individual seedlings, effectively improving the accuracy and robustness of target seedling contour structure reconstruction in multi-view images.

[0122] This application further proposes that the specific steps for generating structural stability labels include:

[0123] Based on the boundary stability mean, the variation standard deviation, and the contour matching rate, a stability index tensor is constructed.

[0124] ;

[0125] Stability index tensor Used to describe the overall stable state of the boundary structure within the current viewpoint interval, where:

[0126] The mean value for boundary stability represents the overall offset strength of the boundary.

[0127] The standard deviation of the boundary variation represents the degree of dispersion of boundary fluctuations.

[0128] The contour matching rate represents the level of contour consistency across different viewpoints.

[0129] Subsequently, a preset stable threshold range is defined, which includes the threshold range for the mean boundary offset, the threshold range for the standard deviation of fluctuation, and the threshold range for the matching rate.

[0130] The stability index tensor is compared dimension-by-dimensionally with the preset stability threshold interval to determine the interval division rules:

[0131] If the boundary stability mean, boundary variation standard deviation, and profile matching rate all fall within the boundary offset mean threshold range, the fluctuation standard deviation threshold range, and the matching rate threshold range, respectively, then the window segment is determined to be a stable range.

[0132] If any of the above conditions are not met, the window segment is determined to be an unstable interval.

[0133] After completing the interval division, a corresponding structural stability label is generated for each window segment:

[0134] ;

[0135] And the set of boundary points covered by all window segments belonging to the "unstable" level:

[0136] ;

[0137] Unified marking of non-stable region boundary points and segments avoids interference from abnormal boundaries caused by severe shading, sudden changes in illumination, or changes in viewpoint on the overall seedling structure modeling and quality grading results.

[0138] This application's solution constructs a stability index tensor and introduces a multi-dimensional threshold interval division mechanism to achieve refined evaluation and region labeling of the structural stability of seedling contours in multi-view sequences. Specifically, the solution integrates the boundary stability mean, variation standard deviation, and contour matching rate to construct a stability index tensor reflecting spatial consistency and the degree of boundary perturbation. This tensor is then classified using preset threshold intervals to generate structural stability labels, thereby accurately identifying unstable boundary regions with significant structural fluctuations. This improves the connectivity and robustness of the contour connection map, effectively solving the problem of contour structural instability caused by occlusion and lighting changes in high-density seedling cultivation scenarios, and enhancing the overall adaptability, accuracy, and interpretability of the system.

[0139] This application further proposes that the specific steps for generating the candidate structure boundary set of a single seedling include:

[0140] In high-density seedling tray monitoring scenarios, due to overlapping seedling leaves, differences in growth direction, and changes in viewing angle, some seedling outlines appear broken or missing in a single-view frame, and have been marked as unstable region boundary segments in previous steps. For these types of regions, this embodiment uses a cross-view structural atlas to perform multi-frame completion and structural filtering of the outlines to restore the continuous boundary structure of a single seedling.

[0141] First, in the cross-view structure atlas, locate the set of node chains corresponding to all boundary segments marked as unstable regions;

[0142] By aggregating the boundary point vectors from different viewpoint frames within the same node chain, a multi-frame boundary candidate vector group is constructed:

[0143] ;

[0144] in, Indicates the first In each view frame, the coordinate vector of the k-th boundary point corresponding to the contour segment; This indicates the total number of boundary points of the contour segment.

[0145] This vector set contains spatial observations of the same contour structure from multiple perspectives, which are used for subsequent temporal alignment and structural reconstruction.

[0146] Since the acquisition time and view index of frames from different viewpoints may have uneven intervals, firstly... Index sequence by viewpoint Sort the frames and construct a unified frame order mapping function:

[0147] ;

[0148] Map all candidate boundary points to a unified time axis.

[0149] Under a unified timeline, for the same contour index, spatial interpolation reconstruction is performed on its point vectors in different viewpoint frames to obtain continuous boundary candidate points across frames:

[0150] ;

[0151] in, This represents the point obtained by linearly interpolating the position of the k-th contour point in different view frames, used to construct a continuous boundary trajectory at time point t; , The indices of the two nearest adjacent view frames to time point t; The linear interpolation weights are calculated from the time offset ratio.

[0152] By performing the above interpolation process on all k, a unified boundary candidate reconstruction map is generated across frames:

[0153] ;

[0154] in, Represents the set of boundary points reconstructed on a unified timeline;

[0155] This atlas describes the continuous boundary structure recovered from occluded or broken contours under multi-view constraints.

[0156] To avoid including interpolated abnormal or erroneous boundaries in the final structure set, a dual screening process is introduced using a structure closure threshold and a boundary continuity index.

[0157] For each candidate boundary curve in the reconstructed map, the ratio of the distance between its first and last points to the total length of the boundary is calculated and defined as the structural closure degree. If the closure degree is greater than or equal to the structural closure degree threshold, the candidate boundary is considered to meet the structural closure requirement.

[0158] Further calculations are made of the local rotation changes formed by adjacent points in the candidate boundary to evaluate the smoothness of the boundary, defined as:

[0159] ;

[0160] in, This is a boundary continuity index, representing the average angle change of the boundary curve. The smaller the value, the more continuous and smooth the boundary. This represents the i-th boundary point; This represents the (i-1)th boundary point; M represents the total number of boundary points contained in the current candidate boundary curve.

[0161] If the structural closure degree is greater than or equal to the preset structural closure degree threshold and the boundary continuity index is less than or equal to the preset boundary continuity threshold, then a set of candidate structural boundaries is selected.

[0162] This application's solution extracts and reconstructs boundaries of contour segments in unstable regions from a cross-view structural map set, effectively overcoming boundary loss issues caused by occlusion, lighting interference, or viewpoint shift in single-view images, thus improving the completeness and continuity of boundary reconstruction. Employing temporal alignment and spatial interpolation mechanisms, dynamic fusion of boundary data across multiple frames ensures geometric consistency and spatial coherence of the contour structure over time, significantly enhancing the ability to reconstruct real seedling structures. Simultaneously, a dual screening criterion of structural closure and boundary continuity effectively filters candidate boundaries with abnormal shapes or abrupt curvature changes, ensuring the geometric stability and biological rationality of the generated candidate structural boundary set. This provides accurate structural input for subsequent seedling quality grading and individual identification, improving the robustness and discrimination accuracy of the overall recognition system.

[0163] This application further proposes that the specific steps for generating the verification structure image set include:

[0164] In this embodiment, to address the technical challenges of occlusion, intersection, or uneven light reflection at the boundaries of seedling structures in high-density seedling trays, a boundary verification and image reconstruction mechanism based on multi-view reflection direction features is proposed. This mechanism is used to further screen stable and reliable seedling contour regions from candidate structural boundaries and generate a set of visual verification images.

[0165] Based on the candidate structure boundary set obtained above, the reflection direction distribution parameter set corresponding to each candidate boundary segment is extracted sequentially in the multi-view image frames.

[0166] In each image frame, the reflection direction vector is calculated using the local illumination vector of the candidate boundary region in the image and the corresponding pixel normal vector, and is defined as:

[0167] ;

[0168] in, It refers to the reflection direction vector of the i-th candidate boundary segment from the v-th viewpoint; It refers to the local illumination vector of the i-th candidate boundary segment from the v-th viewpoint; This refers to the pixel normal vector of the i-th candidate boundary segment from the v-th viewpoint;

[0169] By concatenating all reflection direction vectors in each frame, a sequence of candidate boundary reflection directions is obtained:

[0170] Select the leaf surface reflection vector in the current scene and normalize all candidate boundary reflection vectors;

[0171] The similarity between the candidate boundary reflection vector and the leaf surface standard vector is calculated using the cosine similarity formula. The maximum similarity is taken as the reflection direction similarity of the candidate boundary segment in the current frame, and a reflection direction similarity sequence is constructed.

[0172] The reflection direction similarity sequence is denoised and smoothed using a sliding window filtering method. The denoised and smoothed reflection direction similarity sequence is then segmented and clustered (e.g., using DBSCAN or a custom sliding window clustering) to identify segments that are continuously higher than a preset similarity threshold, which are then recorded as candidate segments for verification boundaries.

[0173] For each candidate segment of the verification boundary, mark its spatial region in the original image frame, extract the corresponding region from the original image frame, and stitch together all the corresponding regions of the verification boundary to generate a set of verification structure images.

[0174] This application's scheme constructs a sequence of reflection directions for candidate structural boundaries from multiple perspectives and performs cosine similarity analysis with the normalized leaf reflection vector. This effectively identifies boundary segments in multi-angle images that are highly consistent with the optical properties of the real leaf surface. The method integrates geometric structural features and light reflection direction characteristics, significantly enhancing the verification capability of boundary authenticity and the robustness of recognition under occlusion interference. By denoising and clustering high-similarity segments and locating and reconstructing the verification image region in the original image, it can achieve structural credibility judgment and region-level reconstruction of candidate boundary segments. This provides strong data support and structural closed-loop assurance for the accurate extraction of single seedling boundaries and health status identification in dense seedling cultivation scenarios.

[0175] This application further proposes that the specific steps for generating seedling quality grades include:

[0176] First, key geometric parameters of the target seedlings in image frames from various viewpoints are extracted from the verification structure image set, including structural projection length, leaf width-to-thickness ratio, edge curvature change rate, and contour integrity. These are defined as follows:

[0177] ;

[0178] in, This represents the length of the structure projection from the v-th viewpoint; and This represents the coordinates of the two endpoints of the seedling outline along the principal axis from the perspective v.

[0179] ;

[0180] in, Indicates the width-to-thickness ratio of the blade; Indicates the average width of the leaf surface; This represents the average thickness of the blade surface, obtained by reconstructing the profile section.

[0181] Edge curvature change rate:

[0182] ;

[0183] in, is the rate of change of edge curvature of the contour region corresponding to the image frame at the v-th viewpoint; Let be the local curvature value of the i-th boundary point; Let be the local curvature value of the (i+1)th boundary point.

[0184] ;

[0185] in, The integrity of the outline c; This represents the area enclosed by the actual outline; This is an estimated value for the area of ​​the ideal elliptical profile.

[0186] The above indicators are summarized to construct a morphological feature dataset: ;

[0187] To improve the feasibility of comparisons between indicators, each indicator is normalized based on its spatial continuity and lateral symmetry in multi-view image sequences.

[0188] Meanwhile, based on the previously generated structural stability labels, a weight factor is assigned to each indicator to generate a structural weighted feature set; for example, if the label level is "high stability", then the weight factor is 1.0; if the label level is "medium stability", then the weight factor is 0.7; if it is "low stability", then the weight factor is 0.4.

[0189] The weighting factors of the structurally weighted feature set are weighted together with the previously calculated reflection direction similarity vector to generate a feature verification value. If the feature verification value is greater than or equal to a preset feature verification threshold, the structurally weighted feature set corresponding to the feature verification value is denoted as the verification feature vector set. .

[0190] Define a mapping function Ф: This function maps validation feature vectors to growth state scoring intervals. It can be constructed using a weighted linear model or a non-linear kernel function model. The following is the form of a linear weighted scoring function: ;

[0191] in, To output a growth status score for the target seedling, the range is... ; To verify the i-th normalized feature in the feature vector set; These are the weighting factors for the corresponding features;

[0192] To enhance the scoring model's responsiveness to structural features, each feature is categorized into low, medium, and high stability levels based on previously generated structural stability labels using a preset stability threshold range. This results in structural stability level labels. Set feature adjustment factor , so that: ;

[0193] This generates a structure-sensitive weighted scoring model:

[0194] ;

[0195] Finally, the quality grade of the target seedling is determined based on its position in the preset quality grade threshold range [T1, T2, T3] according to its growth status score.

[0196] If the growth status score is greater than T3, the quality grade is excellent;

[0197] If the growth status score is greater than T2 and less than or equal to T3, the quality grade is good.

[0198] If the growth status score is greater than T1 and less than or equal to T2, the quality grade is medium.

[0199] If the growth status score is less than or equal to T1, the quality grade is poor.

[0200] This application's solution extracts multi-dimensional morphological features of target seedlings from a verification structural image set, including structural projection length, leaf width-to-thickness ratio, edge curvature change rate, and contour integrity from multiple perspectives. A morphological feature dataset is constructed, and each indicator is normalized by combining spatial continuity and lateral symmetry. A structural stability level label is introduced to set a weighting factor, generating a structure-weighted feature set. This feature set is then jointly encoded with the reflection direction similarity of candidate boundaries to construct a verification feature vector set, which is projected onto a preset growth state evaluation interval mapping function to output a growth state score. Finally, the seedling quality level is automatically determined based on the score's position within a multi-level threshold interval. This process achieves a closed-loop processing from structural perception and feature fusion to level evaluation, significantly improving the accuracy and stability of seedling growth state determination and reducing the uncertainty of manual intervention.

[0201] Through the above design, this application realizes an integrated processing flow for seedling structural stability recognition and quality level assessment in high-density, multi-occlusion scenarios. It can accurately extract seedling boundary structure and morphological features from multi-view image sequences, dynamically identify unstable regions, and complete structural verification based on boundary reconstruction and reflection direction similarity. Combining multi-source feature indicators such as structural projection length, leaf width-to-thickness ratio, edge curvature change rate, and contour integrity, a weighted feature set with fused stability weights is constructed. Growth status scoring and grade classification are completed through joint encoding and mapping functions, forming a closed-loop process from boundary perception, stability judgment, structural verification to quality level output. This technical solution effectively solves the problems of occlusion interference, multi-angle mismatch, and strong subjectivity in quality assessment in existing methods, significantly improving the accuracy and robustness of seedling image recognition.

[0202] In another embodiment, this application also provides an intelligent monitoring and quality grading system for seedling growth status images, such as... Figure 2 As shown, it includes the following modules:

[0203] The image monitoring module is used to simultaneously acquire multi-view image sequences, light parameter vectors, and leaf reflectance vectors of the target seedling area;

[0204] The image difference module is used to identify the proportion of overlapping areas and the displacement vector of contour points of seedling outlines under different perspectives between adjacent multi-view image frames based on the multi-view image sequence and the illumination parameter vector, and to construct an overlapping difference map set.

[0205] The image contour module is used to continuously track the contours of target seedlings in multi-view images across different perspectives using the overlapping difference map set, and generate a cross-view structure map set of seedling contours.

[0206] The image contour analysis module is used to calculate the boundary stability mean, variation standard deviation and contour matching rate of the target seedling structure between adjacent views based on the cross-view structure atlas, and generate a structure stability label based on the boundary stability mean, variation standard deviation and contour matching rate within a preset stability interval.

[0207] The image structure module is used to compensate the boundary region outside the preset stability interval by using the cross-view structure map and the structure stability label to generate a candidate structure boundary set for a single seedling.

[0208] The image verification module is used to filter each candidate structure boundary by comparing the similarity between the reflection direction distribution of the candidate structure boundary and the reflection vector of the leaf surface, and generate a set of verification structure images.

[0209] The quality grading module is used to extract the morphological feature dataset of the target seedling based on the verification structure image set, and combine the structure stability label and the similarity to map it to a preset growth state evaluation interval to generate the seedling quality grade.

[0210] The following is a complete implementation description based on the template you provided, which elaborates on the engineering scenarios and reconstructs the technical loop for the technical modules you listed, highlighting the advantages of systematicity, dynamic control, and practicality:

[0211] This method acquires multi-view images of the seedling tray area using an image monitoring module, combining ambient light parameters and leaf reflectance vector information to simultaneously obtain multi-dimensional visual information of the target area. The module acquires image sequences from different viewpoints and associates each frame with the light parameter vector and leaf reflectance characteristic vector under the current illumination conditions, constructing a ternary dataset of light-viewpoint-reflectance. This significantly improves the adaptability to changes in lighting conditions and reflection interference during subsequent structure recognition, providing a comprehensive data foundation for structural stability analysis.

[0212] Through the image difference module, this method, based on multi-view image sequences and illumination parameter vectors, jointly calculates the proportion of overlapping contour regions and boundary point displacement vectors between adjacent view frames to construct an overlap difference atlas. This atlas is used to accurately measure the deformation characteristics of seedling contours in multi-view images under illumination interference, solving the problem that existing single-frame images are unable to effectively capture contour inconsistencies caused by factors such as occlusion and viewpoint shifts, and providing dynamic difference indicators for subsequent contour fusion and stability analysis.

[0213] The subsequent image contour module utilizes an overlapping difference atlas to achieve continuous cross-view tracking of the target seedling contour in the image sequence, generating a structurally complete cross-view contour atlas. This atlas preserves the correspondence and deformation trajectory of contour information under each viewpoint, constructs a spatially continuous structural representation, enhances the ability to fully model the seedling boundary morphology, and effectively improves the spatial robustness of structural boundary recognition.

[0214] In the image contour analysis module, the system quantitatively models the boundary characteristics of the target seedling structure based on the aforementioned image set, extracts key indicators such as the boundary stability mean, standard deviation of variation, and contour matching rate, and sets a structural stability reference interval to generate a structural stability label set. This module dynamically evaluates the stability of the seedling structure from multiple angles by statistically analyzing contour offset and matching degrees between different viewpoints, effectively identifying unstable regions affected by illumination interference, occlusion disturbances, or morphological variations, and constructing a foundation for stability classification.

[0215] Guided by structural stability labels, the image structure module performs path compensation on boundary segments corresponding to unstable regions, constructing a candidate structure boundary set based on the spatial alignment relationship and boundary continuity parameters of the cross-view contour atlas. This compensation mechanism ensures the closure and coherence of the final contour structure under multiple views by filling in missing boundary regions and adjusting contour connectivity, significantly improving the realism and completeness of the candidate structure boundaries.

[0216] Next, the image verification module performs physical consistency screening of candidate structure boundaries based on leaf reflection vectors. It compares the similarity between the reflection direction distribution of candidate structure boundaries and leaf reflection vectors, and constructs a verification structure image set based on a similarity threshold. This module combines the cosine function of the vector angle to achieve physical constraint verification of structural reflection features, effectively eliminating erroneous boundary structures caused by strong reflections, ghosting, or occlusion, ensuring that the final extracted structure possesses true physical consistency under illumination conditions.

[0217] Finally, through the quality grading module, the system extracts key morphological features of the target seedlings from the verification structural image set, including structural projection length, leaf width-to-thickness ratio, edge curvature change rate, and contour integrity, to construct a morphological feature dataset. This dataset is then combined with structural stability labels and reflection similarity to construct a verification feature vector set. This vector set is projected onto a pre-defined growth state mapping function, outputting a corresponding growth state score. Based on the score's placement within multiple grade intervals, the final quality grade label of the target seedling is determined. This mechanism integrates morphology, stability, and reflection consistency—three dimensions—to achieve multi-faceted and high-precision evaluation of seedling quality.

[0218] Through the aforementioned technical process, this method constructs a closed-loop intelligent evaluation mechanism encompassing multimodal image recognition, structural stability analysis, boundary verification and reconstruction, and growth level assessment. This overcomes the core challenges of traditional seedling identification, such as limitations imposed by single-view images, structural occlusion interference, and subjective quality assessment. The system integrates morphological contour information and reflection direction physical properties under multi-view illumination environments, achieving highly reliable and adaptive seedling boundary extraction and quality grading. This significantly improves the automation level of refined management and early screening in the seedling cultivation process.

[0219] The core innovation of this embodiment lies in the first-ever construction of an integrated evaluation system for seedling trays under the complex scenarios of occlusion, viewing angle deviation, and reflection interference. This system integrates cross-view contour continuous recognition, reflection consistency verification, and quality level mapping. By fusing key modules such as image difference modeling, structural stability classification, boundary physical verification, and feature weighted mapping, a dynamic evaluation framework driven by multi-source data is formed, enabling accurate identification and intelligent classification of seedlings under multi-interference environments.

[0220] The above embodiments are only used to illustrate the technical methods of this application and are not intended to limit it. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of this application without departing from the spirit and scope of the technical methods of this application.

Claims

1. A method for intelligent monitoring and quality grading of seedling growth status images, characterized in that, Includes the following steps: Simultaneously acquire multi-view image sequences, light parameter vectors, and leaf reflectance vectors of the target seedling area; Based on the multi-view image sequence and the illumination parameter vector, the overlapping area ratio and contour point displacement vector of the seedling outline between adjacent multi-view image frames under different views are identified, and an overlap difference map set is constructed. The overlapping difference map is used to continuously track the outline of the target seedling in the multi-view image across different perspectives, and generate a cross-view structure map of the seedling outline. Based on the cross-view structure atlas, the boundary stability mean, standard deviation of variation and contour matching rate of the target seedling structure between adjacent views are calculated. Based on the boundary stability mean, standard deviation of variation and contour matching rate within the preset stability interval, a structure stability label is generated. By using the cross-view structure diagram and the structure stability label, compensation processing is performed on the boundary region outside the preset stability interval to generate a candidate structure boundary set for a single seedling. By comparing the similarity between the reflection direction distribution of the candidate structure boundary and the leaf surface reflection vector, each candidate structure boundary is filtered to generate a set of verification structure images. Based on the verification structure image set, extract the morphological feature dataset of the target seedling, and combine it with the structural stability label and the similarity to map it to a preset growth state evaluation interval to generate the seedling quality grade.

2. The intelligent monitoring and quality grading method for seedling growth status images according to claim 1, characterized in that, The process of identifying the proportion of overlapping areas and the displacement vectors of contour points of seedling outlines between adjacent multi-view image frames includes: Based on the multi-view image sequence and the illumination parameter vector, timestamp alignment and view index matching are performed on adjacent view image frames. The contour point set of the target seedling is extracted in the corresponding view image frame through the matching result, and the contour point set is grouped based on a preset spatial neighborhood radius to generate a view contour subset. Based on the aforementioned viewpoint contour subset, the proportion of overlapping areas of contour regions between adjacent viewpoints and the displacement vector of contour points are calculated.

3. The intelligent monitoring and quality grading method for seedling growth status images according to claim 1, characterized in that, The process of generating an overlapping difference atlas includes: Based on the illumination parameter vector, the contour point displacement vector is corrected for illumination perturbation, and a corrected displacement dataset is generated. Based on the ratio of the corrected displacement dataset to the overlapping region, a pixel-level difference weight matrix is ​​generated according to a preset mapping rule. The pixel-level difference weight matrix is ​​then spatially mapped and fused with the original multi-view image sequence to generate the overlapping difference map set.

4. The intelligent monitoring and quality grading method for seedling growth status images according to claim 1, characterized in that, The process of generating a cross-view structural atlas of seedling outlines includes: Based on the overlapping difference map set, the boundary point set of the target seedling outline under each view is extracted, and the outline frame relationship is constructed. Based on the boundary point positions and boundary offset magnitudes between adjacent frames in the aforementioned contour frame relationship, a cross-frame contour correspondence is constructed, and frame pairs with excessive view spans in the cross-frame contour correspondence are removed according to a preset view spacing threshold. Based on the filtered cross-frame contour correspondence, the Euclidean distance and boundary gradient difference between each pair of boundary points are calculated to generate the contour connectivity tensor. The contour connection tensor is jointly mapped to the contour frame relationship to generate a cross-view connection node chain arranged according to the contour frame relationship, and a cross-view structure atlas is constructed based on the cross-view connection node chain.

5. The intelligent monitoring and quality grading method for seedling growth status images according to claim 1, characterized in that, The process of calculating the boundary stability mean, standard deviation of variation, and contour matching rate of the target seedling structure between adjacent viewpoints is as follows: Based on the cross-view structure atlas, a set of boundary points for the seedling outline in each view frame is constructed, and a sequence of boundary fluctuation vectors is established according to the view sequence. A sliding window is preset, and the boundary stability mean and standard deviation of variation within the view segment are calculated; The fluctuation values ​​in the boundary fluctuation vector sequence are normalized, and the contour matching rate is calculated.

6. The intelligent monitoring and quality grading method for seedling growth status images according to claim 5, characterized in that, The process of generating structural stability labels includes: Based on the boundary stability mean, the variation standard deviation and the contour matching rate, a stability index tensor is constructed, and the stability index tensor is divided into intervals according to a preset stability threshold interval. Based on the interval division results, structural stability labels are generated, and boundary point segments that do not belong to the preset stability threshold interval are marked as unstable regions.

7. The intelligent monitoring and quality grading method for seedling growth status images according to claim 6, characterized in that, The process of generating the candidate structure boundary set for individual seedlings includes: Based on the contour segments of the unstable region, the multi-view contour trajectory corresponding to the unstable region is extracted from the cross-view structure map set, and a multi-frame boundary candidate vector group is constructed. Temporal alignment and spatial interpolation are performed on the multi-frame boundary candidate vector group to generate a unified cross-frame boundary candidate reconstruction map; Based on the preset structural closure threshold, the degree of closure of each candidate boundary in the reconstructed map is screened, and a set of candidate structural boundaries is generated by combining the preset boundary continuity index.

8. The intelligent monitoring and quality grading method for seedling growth status images according to claim 1, characterized in that, The process of generating a verification structure image set includes: Based on the candidate structure boundary set, the reflection direction distribution parameters of each candidate structure boundary in multi-view image frames are extracted to construct the candidate boundary reflection direction sequence. The leaf surface reflection vector is normalized, and the similarity between the candidate boundary reflection direction sequence and the leaf surface reflection vector is calculated based on the cosine function of the vector angle. The similarity is denoised and segmented into clusters. High similarity segments are identified by a preset similarity threshold and marked as candidate segments for verification boundaries. The image regions corresponding to the candidate verification boundary segments in the original image frame are extracted and reconstructed to generate a set of verification structure images.

9. The intelligent monitoring and quality grading method for seedling growth status images according to claim 1, characterized in that, The process of generating seedling quality grades includes: Based on the verification structure image set, the structural projection length, leaf width-to-thickness ratio, edge curvature change rate and contour integrity of the target seedling under multiple views are extracted to construct a morphological feature dataset. The indicators in the morphological feature dataset are normalized according to their spatial continuity and lateral symmetry, and weight factors are set according to the structural stability label to generate a structural weighted feature set. The structurally weighted feature set and the similarity are jointly encoded to construct a verification feature vector set; The verification feature vector set is projected onto a mapping function of a preset growth state evaluation interval to output a growth state score; The seedling quality grade is determined based on the position of the growth status score within multiple quality grade threshold ranges.

10. A seedling growth status image intelligent monitoring and quality grading system, characterized in that: The application includes the intelligent monitoring and quality grading method for seedling growth status images as described in any one of claims 1 to 9, comprising: The image monitoring module is used to simultaneously acquire multi-view image sequences, light parameter vectors, and leaf reflectance vectors of the target seedling area; The image difference module is used to identify the proportion of overlapping areas and the displacement vector of contour points of seedling outlines under different perspectives between adjacent multi-view image frames based on the multi-view image sequence and the illumination parameter vector, and to construct an overlapping difference map set. The image contour module is used to continuously track the contours of target seedlings in multi-view images across different perspectives using the overlapping difference map set, and generate a cross-view structure map set of seedling contours. The image contour analysis module is used to calculate the boundary stability mean, variation standard deviation and contour matching rate of the target seedling structure between adjacent views based on the cross-view structure atlas, and generate a structure stability label based on the boundary stability mean, variation standard deviation and contour matching rate within a preset stability interval. The image structure module is used to compensate the boundary region outside the preset stability interval by using the cross-view structure map and the structure stability label to generate a candidate structure boundary set for a single seedling. The image verification module is used to filter each candidate structure boundary by comparing the similarity between the reflection direction distribution of the candidate structure boundary and the reflection vector of the leaf surface, and generate a set of verification structure images. The quality grading module is used to extract the morphological feature dataset of the target seedling based on the verification structure image set, and combine the structure stability label and the similarity to map it to a preset growth state evaluation interval to generate the seedling quality grade.

Images