Center positioning method and system based on blade skeleton driving geometric analysis
By using a leaf skeleton-driven geometric analysis method, combined with deep learning and morphological skeleton geometric reasoning, the problem of insufficient positioning accuracy of crop growth center in complex field environments is solved, achieving high-precision and robust crop growth center positioning, which is suitable for real-time deployment in embedded devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU UNIV
- Filing Date
- 2026-02-09
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies lack the accuracy and robustness for locating the crop growth center in complex field environments. In particular, under conditions of light variation, shading, weed interference, and crop morphology asymmetry, automated equipment may fail to execute the position, resulting in crop damage or reduced operational effectiveness.
A geometric analysis method based on leaf skeleton is adopted. Through deep learning semantic segmentation and morphological skeleton geometric reasoning, the central skeleton of the leaf is extracted and reliability is judged, and the coordinates of the physiological growth center of the crop are output.
It achieves precise positioning of crop growth centers in complex field environments, improves positioning accuracy and robustness, reduces computational overhead, is suitable for real-time deployment in embedded devices, and has good versatility and noise resistance.
Smart Images
Figure CN122089680A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural machine vision and intelligent equipment technology. Specifically, it relates to a method and system for locating the growth center of crop seedlings based on leaf semantic segmentation and skeleton geometric analysis, which is suitable for precision agriculture operations in complex field environments. Background Technology
[0002] In precision agriculture, the effective implementation of automated equipment such as weeding robots, thinning machines, and precision fertilization machines relies on the rapid, accurate, and stable identification and positioning of the growth center of individual crops in the field. Existing crop center localization methods are mostly based on traditional image processing techniques, estimating the position using color, texture features, or the geometric centroid of segmented contours. However, in actual field environments, the positioning accuracy and robustness of these methods are significantly reduced due to complex lighting changes (strong light, shadow, uneven lighting), occlusion by weeds and soil debris, differences in crop growth stages and varieties, and wind-induced swaying. While the accuracy of crop recognition based on object detection and instance segmentation has improved with the development of deep learning, existing solutions typically still approximate the crop center using the geometric centroid of a segmentation mask. When crop seedlings exhibit morphological asymmetry, uneven leaf distribution, local occlusion, or damage, the geometric centroid easily deviates from the crop's true physiological growth center (e.g., the root collar or tillering node), leading to positional errors in the automated equipment, causing crop damage or reduced operational effectiveness. Furthermore, existing technologies lack a systematic technical solution for inferring crop growth centers based on leaf skeletal morphology and spatial extension patterns from a plant morphology perspective. Given the biological principle that leaf midribs typically converge at the growth point, there is an urgent need to propose a method suitable for complex field scenarios that can accurately and robustly locate the physiological growth center of crops to support high-precision automated agricultural operations. Summary of the Invention
[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for detecting the center of crop seedlings based on leaf skeleton-driven geometric analysis. This method achieves accurate localization of the physiological growth center of crops in complex field environments by integrating deep learning semantic segmentation with morphological skeleton geometric reasoning.
[0004] To achieve the above objectives, the present invention employs the following technical solution:
[0005] A method for detecting the center of crop seedlings based on leaf skeleton-driven geometric analysis, characterized by comprising the following steps:
[0006] S1. Perform leaf detection and segmentation processing on the collected field crop images to obtain a binary image of the target crop leaf mask;
[0007] S2. Extract the central skeleton of the leaf based on the binary image of the leaf mask and generate the leaf center fitting curve, thereby obtaining the centroid clue of a single plant area and the convergence area clue of the extended line; then perform reliability discrimination based on the centroid clue of the single plant area and the convergence area clue of the extended line, and output the center point coordinates of the target crop accordingly.
[0008] Furthermore, step S2 specifically includes the following sub-steps:
[0009] S21. Perform morphological skeletonization processing on the binary image of the leaf mask to obtain a linear skeleton with a single pixel width; count the number of neighbors of the skeleton pixels based on 8 connected neighborhoods, and divide the skeleton points into endpoints, ordinary points and connection points, wherein the number of neighbors of the endpoints is 1, the number of neighbors of the ordinary points is 2, and the number of neighbors of the connection points is not less than 3.
[0010] S22. Construct an undirected graph with skeleton points as vertices and 8-connectivity as edges; based on the undirected graph, perform shortest path extraction between skeleton endpoint pairs to obtain a candidate shortest path set, wherein the shortest path extraction adopts breadth-first search; in a preferred embodiment, the edges between adjacent skeleton points are regarded as equally weighted edges, so that the shortest path corresponds to the path with the fewest edges; when there is an equal-length shortest path or when geometric smoothness needs to be considered, for the equal-length shortest path obtained by the breadth-first search, the cumulative path length calculated by the cumulative Euclidean distance (or 8-connectivity fixed step size weight) between adjacent skeleton points is used as the reordering criterion to determine the key path point sequence; and when there are multiple branches in the skeleton, determine the blade main path from the candidate shortest paths according to the preset geometric evaluation criteria.
[0011] S23. The cumulative arc length parameterization of the path point sequence of the main path of the blade is normalized to the [0,1] interval, and interpolation fitting is performed based on the parameterization result to obtain the blade center fitting curve.
[0012] S24. Perform contour calculation on the binary image of each blade mask to obtain the minimum circumcircle curve; take the two endpoints of the fitted curve at the center of the blade as the starting point, search outward along the tangent direction at the endpoints for the intersection point with the minimum circumcircle and / or the mask contour as the ending point, connect the starting point and the ending point to form an extended line segment, and define the blade outward direction with the extended line segment.
[0013] S25. Perform element overlay display: The binary image of the blade mask, the mask contour curve obtained in step S21, the blade center fitting curve obtained in step S23, and the minimum circumcircle curve obtained in step S24 are directly copied and overlaid using pixels; the extended line segment pixels are covered and placed on top of each other to highlight the intersection.
[0014] S26. Then, perform morphological closing operation on the binary image of the leaf mask, merge adjacent or slightly broken single leaf regions to obtain single crop regions; calculate the centroid cues for each single crop region.
[0015] S27. Extract candidate convergence positions from the extended line segments obtained in step S24 to obtain convergence region clues that characterize the degree of convergence in the blade's outward direction.
[0016] S28. Based on the centroid clues of the single crop region obtained in S26 and the convergence region clues obtained in S27, construct at least two candidate centers and perform reliability judgment, and output the coordinates of the target crop center point.
[0017] In an optional embodiment, step S1 may further include: S11, constructing and training a crop leaf detection model; S12, inputting multiple consecutive frames of RGB images to the model to obtain leaf segmentation contour information; S13, normalizing the contour information and converting it into a binary image of a leaf mask. The crop leaf detection model can use YOLOv8 as a baseline framework, replacing some of its standard convolutional layers with ODConv_3rd dynamic convolutions, and using a weighted combination of DFL distributed focal loss and WIoU regression loss as the regression loss function for training.
[0018] In an optional embodiment, the preset geometric evaluation criteria in step S22 include at least the maximum arc length criterion. The interpolation fitting in step S23 preferably uses B-spline interpolation. The candidate convergence location extraction in step S27 may include pixel overlap detection and / or nearest neighbor detection, and the extracted points may undergo aggregation and denoising processing. The reliability judgment in step S28 is based at least on the dispersion of the extended line segment direction, the size of the candidate convergence points, and / or the shape of the convergence envelope, and accordingly adaptively selects the output center point coordinates between the centroid cues and the convergence cues.
[0019] The present invention also provides a system, electronic device, and computer-readable storage medium for implementing the above method.
[0020] Compared with the prior art, the present invention has at least the following beneficial effects:
[0021] (1) Based on the semantic segmentation mask of the leaf, this invention combines skeletonization and main path extraction of the skeleton graph to achieve a stable expression of the central structure of the leaf. Compared with the estimation method based solely on the geometric centroid of the mask, it can obtain a more accurate central positioning result of the crop's actual growth point in cases of uneven leaf distribution, local occlusion, or leaf asymmetry.
[0022] (2) This invention obtains a continuous and smooth leaf center fitting curve by parameterizing the arc length of the main path of the skeleton and using B-spline interpolation, which makes the calculation of the leaf outward direction more continuous and noise resistant, thereby improving the robustness under complex field light changes, leaf defects and boundary noise conditions.
[0023] (3) This invention introduces the fusion of multiple clues, namely “single plant region centroid clue” and “extension line convergence region clue”, and achieves adaptive center decision by reliability discrimination based on stability indicators such as directional dispersion, convergence scale and envelope shape. It can automatically select a more reliable center output under different working conditions such as leaves tending to be parallel, convergence not obvious or adjacent seedlings sticking together, thereby reducing the risk of mispositioning.
[0024] (4) This invention combines lightweight segmentation model improvement with efficient morphological / geometric calculation process, which reduces computational overhead while ensuring positioning accuracy and stability. It is suitable for real-time deployment on embedded edge devices and has low dependence on crop color and texture. It has good versatility and promotion value. Attached Figure Description
[0025] Figure 1 This is a flowchart illustrating a center positioning method based on blade skeleton driven geometric analysis provided by the present invention.
[0026] Figure 2 This is a schematic diagram of the structure of the crop leaf detection model provided by the present invention;
[0027] Figure 3 This is a schematic diagram illustrating the center identification of the present invention under three different conditions, wherein:
[0028] (a) is a schematic diagram of the output centroid candidates when the blade direction tends to be parallel;
[0029] (b) A schematic diagram of outputting convergence candidates when the convergence region is reliable;
[0030] (c) is a schematic diagram of convergence candidates when seedlings are stuck together;
[0031] Figure 4 This is a schematic diagram of a center positioning system based on blade skeleton driven geometric analysis provided by the present invention;
[0032] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention.
[0033] Figure 6 These are schematic diagrams of three conventional crop center detection methods, in which:
[0034] (a) is a schematic diagram of the mask centroid method;
[0035] (b) is a schematic diagram of the target detection center method;
[0036] (c) Schematic diagram of the target box centering method;
[0037] Figure 7 This is a schematic diagram of the visualization results provided by the present invention; Detailed Implementation
[0038] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of this invention.
[0039] like Figure 1 As shown, the present invention provides a method for detecting the center of crop seedlings based on leaf skeleton driven geometric analysis, including steps S1 to S2, wherein step S2 may further include S21 to S28.
[0040] S1. Obtain the binary image of the leaf mask.
[0041] In practice, a crop leaf detection model can be built and trained first. In a preferred embodiment, YOLOv8 is used as the baseline framework, and some standard convolutional layers in its backbone network and / or feature fusion network are replaced with ODConv_3rd dynamic convolutional units to enhance the model's adaptability to complex field conditions. The model's regression loss function adopts a weighted combination of DFL distributed focal loss and WIoU regression loss. The improved model grid structure is as follows: Figure 2 As shown, the model was trained using a dataset with polygonal annotations of crop leaves.
[0042] In application, RGB images of crops in the field are acquired through image acquisition equipment and input into the trained model to obtain semantic segmentation contour information of the leaves. Subsequently, this contour information is normalized and converted into a binary image of the leaf mask, which is used as input for subsequent steps.
[0043] S2. Based on the binary image of the leaf mask obtained in step S1, extract the central skeleton of the leaf, and determine and output the coordinates of the center point of the target crop through skeleton-driven geometric morphology analysis. Step S2 may include the following sub-steps.
[0044] S21. Contour extraction and morphological skeletonization are performed on the binary image of the leaf mask to obtain a thinned single-pixel-width linear skeleton. Further, based on 8-connected neighborhoods, the number of neighbors for each skeleton pixel is counted, and the skeleton points are classified into endpoints, ordinary points, and connectors, where endpoints have 1 neighbor, ordinary points have 2 neighbors, and connectors have at least 3 neighbors. The skeletonization algorithm can be implemented using a thinning algorithm, such as the Zhang-Suen thinning algorithm or other equivalent thinning algorithms in this field; this invention is not limited to these methods.
[0045] S22. Construct an undirected graph with skeleton points as vertices and 8-connectivity as edges. Based on the undirected graph, perform shortest path extraction between skeleton endpoint pairs to obtain a candidate shortest path set. The shortest path extraction adopts breadth-first search. In a preferred embodiment, the edges between adjacent skeleton points are considered as equally weighted edges, so that the shortest path corresponds to the path with the fewest edges. When there are equal-length shortest paths or when geometric smoothness needs to be considered, the cumulative path length calculated by the Euclidean distance (or 8-connectivity fixed step-size weight) between adjacent skeleton points is used as a reordering criterion to determine the key path point sequence. When the skeleton has multiple branches, the main path of the leaf is determined from the candidate shortest paths according to a preset geometric evaluation criterion. It should be noted that the cumulative length of the above-mentioned Euclidean distance or fixed step-size weight is used to reorder the candidate paths geometrically and select the main path, and does not change the solution method of breadth-first search to determine the shortest path by the number of edges.
[0046] S23. The cumulative arc length parameterization is performed on the main path point sequence determined in S22, and the parameter is normalized to the [0,1] interval. Interpolation fitting is then performed on the path point sequence over the parameter domain to obtain a smooth blade center fitting curve. The interpolation fitting can be implemented using B-spline interpolation; in a preferred embodiment, the B-spline order is 3 to 5, and endpoint repetition nodes are used to enhance endpoint fit. To suppress local noise, the path point sequence can be smoothed and filtered before fitting, for example, using Savitzky-Golay filtering or moving average; this invention is not limited to these methods.
[0047] S24. For each blade mask's binary image, calculate the minimum circumcircle curve of the mask profile. Taking the two endpoints of the center-fitted curve obtained in S23 as starting points, extend outwards along the tangent direction of the curve at those endpoints. Search for the intersection points of this extension line with the minimum circumcircle of the blade and / or the mask profile, and take the first intersection point as the ending point. Connect the starting point and the ending point to form an extended line segment representing the blade's outward direction.
[0048] S25. To facilitate debugging and analysis, the binary image of the blade mask, the mask contour curve, the blade center fitting curve, and the minimum circumscribed circle curve can be overlaid and displayed. To highlight possible intersections, the extended line segments are displayed in a top-overlay manner.
[0049] S26. Next, perform morphological closing operations (e.g., using 3×3 or 5×5 rectangular structuring elements) on the binary image of the leaf mask to merge leaf regions belonging to the same crop but possibly slightly broken or adjacent, thus obtaining complete single-crop regions. Calculate the geometric centroid of each single-crop region as centroid cues.
[0050] S27. Analyze all extended line segments generated in S24 and extract their potential convergence regions. Specifically, this can be achieved through pixel overlap detection (directly finding pixels where different line segments intersect) and / or nearest neighbor detection (finding neighboring pixels after slightly dilating the line segments). To further refine the results, this point set can be aggregated for noise reduction (e.g., using the DBSCAN clustering algorithm). Based on the processed stable point clusters, construct their spatial envelopes (e.g., axis-aligned minimum bounding rectangles), called convergence envelope boxes. These boxes and their internal regions constitute the convergence region cues.
[0051] S28. This step integrates the centroid cues from S26 (denoted as centroid candidates) and the convergence region cues from S27 (using representative points of their convergence envelopes, such as the geometric center, as convergence candidates). The reliability of the convergence region cues is determined by evaluating the following stability metrics:
[0052] 1. Extend the dispersion of line segment directions: If all line segment directions tend to be parallel (the variance is too small), then the convergence is unreliable.
[0053] 2. Size of candidate convergence points: If the number of convergence points or the size of the clusters after clustering is too small, the convergence will be unstable.
[0054] 3. Convergence Envelope Shape: If the area of the envelope is abnormal, the aspect ratio is unbalanced, or the overlap rate with the single plant area is too low, the judgment is unreliable.
[0055] The output center point coordinates are adaptively selected based on the discrimination results: (a) When the leaf direction tends to be parallel, output centroid candidates; (b) or when the convergence area is determined to be reliable, output convergence candidates first; (c) When the centroid of a single plant area is inaccurate due to the adhesion of adjacent seedlings, convergence candidates are also used first. Figure 3 The study demonstrates center identification under three different scenarios.
[0056] In addition, the present invention also provides a crop seedling center detection system based on leaf skeleton driven geometric analysis, such as Figure 4As shown, the system includes a crop center positioning processor 701 and a program or instruction stored thereon. When the program or instruction is executed by the crop center positioning processor 701, it is used to implement any of the aforementioned method steps (including steps S1 to S2 and their sub-steps).
[0057] Optionally, the system further includes a power supply module 702, an image acquisition device 703, and a display device 704. The power supply module 702 supplies power to the crop center positioning processor 701 and the display device 704; the image acquisition device 703 acquires images or video streams of crops in the field and sends them to the crop center positioning processor 701; the crop center positioning processor 701 performs leaf detection and segmentation, binary mask generation, skeleton extraction, geometric reasoning, and center determination on the received images, and outputs the coordinates of the crop seedling center point; the display device 704 displays the crop center positioning results, including the center point coordinates and / or overlaid visual images. The system can also send the center point coordinates to the work execution mechanism or a host computer system via a wired or wireless communication interface.
[0058] It should be understood that the crop center positioning processor 701 can be implemented using an embedded computing platform, edge computing device, server, or general-purpose processor; the deep learning inference framework and operating environment can also be selected according to deployment requirements, and this invention is not limited thereto. For example, the crop center positioning processor 701 can be implemented using an NVIDIA Jetson series platform or an equivalent computing power platform.
[0059] Figure 5 A schematic diagram of an electronic device according to an embodiment of the present invention is shown. The electronic device may include a processor 810, a communication interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communication interface 820, and the memory 830 are connected via the communication bus 840. The memory 830 stores a computer program. When the processor 810 executes the computer program, it implements any of the aforementioned crop seedling center detection methods based on leaf skeleton-driven geometric analysis, including: acquiring multiple consecutive frames of images acquired by an image acquisition device; inputting each frame of images into a crop leaf detection model to obtain leaf segmentation information and generate a binary image of a leaf mask; extracting the leaf skeleton based on the binary image of the leaf mask and performing geometric reasoning analysis, and outputting the coordinates of the target crop seedling center point.
[0060] To further verify the effectiveness and robustness of the crop seedling center detection method based on leaf skeleton driven geometric analysis proposed in this invention in complex field scenarios, this embodiment uses maize seedling field images as the object for comparative experimental analysis.
[0061] (1) Sample and comparison method settings
[0062] To ensure the breadth and representativeness of the verification samples, 100 RGB images of corn seedlings with different postures, different occlusions, and weed interference were selected as test samples. The total number of manually labeled corn seedlings in the 100 images was 872.
[0063] Under the same test set and the same center determination criteria, the method of the present invention is compared with three conventional crop center detection methods, namely:
[0064] A) Masked centroid method;
[0065] B) Target detection center method;
[0066] C) Target box centering method;
[0067] Its schematic diagram is as follows Figure 6 As shown.
[0068] (2) Comparison of quantitative detection capabilities (number of centers)
[0069] Quantitative analysis was used to evaluate the degree of similarity between the number of centroids detected by different methods and the total number of human-labeled true centroids. The statistical results of the number of detected centroids, the actual number of centroids, and the accuracy of different methods are shown in Table 1.
[0070] As shown in Table 1, the method of the present invention detected 759 center points, which is closer to the actual number of 872 centers, achieving an accuracy of 87.04%, significantly higher than the mask centroid method, target detection center method, and target box center method. This indicates that the method of the present invention has a significant advantage in detecting the center of corn seedlings in complex field environments, effectively reducing the operational risks caused by missed or false detections.
[0071] Table 1. Results of central quantitative analysis using different methods: method Number of testing centers actual center number <![CDATA[Accuracy P1 / %]]> Mask centroid method 509 872 58.92 Target detection center method 399 872 45.76 Target box centering method 704 872 80.73 This patented method 759 872 87.04
[0072] (3) Comparison of center positioning accuracy (distance deviation)
[0073] Center accuracy measures the spatial distance deviation between the detected center point and the manually labeled true center. In this embodiment, the conversion is based on the pixel calibration relationship of the row spacing of seedlings in the test image (400mm corresponds to 720 pixels), with each pixel corresponding to an actual distance of approximately 0.56mm. Combining this with the conclusion in existing research that the diameter range of maize crop protection zones is 60–80mm, 30 pixels and 40 pixels are selected as error thresholds, corresponding to actual distance deviations of approximately 16.8mm and 22.4mm respectively, meeting the safety range requirements for crop protection zones.
[0074] Table 4 shows the statistical results of center point errors for different methods under two thresholds. As can be seen from Table 4, when the distance deviation is less than 30 pixels, the number of center points using the method of this invention is 606, accounting for 79.84%; when the distance deviation is less than 40 pixels, the number of center points is 719, accounting for 94.72%. In comparison, the three conventional methods have varying degrees of shortcomings in center localization accuracy. Among them, the target bounding box centering method has a larger error, while the mask centroid method and the target detection centering method are significantly affected by occlusion and shape asymmetry in complex scenes.
[0075] Table 2. Results of center accuracy analysis using different methods: method Mask centroid method (quantity / percentage) Target detection center method (number / percentage) Target box centering method (number / percentage) Our method (quantity / percentage) Within 30 pixels (16.8mm) of the actual center 331 / 43.61% 345 / 86.65% 129 / 18.32% 606 / 79.84% Within 40 pixels (22.4mm) of the actual center 460 / 60.61% 360 / 90.23% 226 / 32.10% 719 / 94.72%
[0076] like Figure 7 As shown, to further verify the effectiveness of the proposed center identification method and demonstrate its application effect, a traditional target box detection center method was selected for comparison. The test was conducted on actual crop seedling images taken in the field. In the image, the red dots represent the reference crop seedling centers, the blue dots represent the detection centers of the proposed method, and the yellow dots represent the detection centers of the traditional target box detection center method. Through visualization of the results, the proposed method is significantly superior to the traditional target box detection center method.
[0077] Based on the above understanding, the present invention may also provide a computer-readable storage medium (e.g., a non-transitory storage medium such as ROM, RAM, disk or optical disk) storing a computer program thereon, wherein the computer program, when executed by a processor, implements any of the aforementioned method steps.
[0078] Finally, it should be noted that the above specific embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; those skilled in the art can modify the above embodiments or make equivalent substitutions for some of the technical features without departing from the spirit and scope of the present invention, and all such modifications and substitutions should fall within the protection scope of the present invention.
Claims
1. A center positioning method based on blade skeleton driven geometric analysis, characterized in that, include: S1. Perform leaf detection and segmentation processing on the collected field crop images to obtain a binary image of the target crop leaf mask; S2. Extract the central skeleton of the leaf based on the binary image of the leaf mask and generate the leaf center fitting curve, thereby obtaining the centroid clues of a single plant area and the convergence area clues of the extension line. Then, based on the centroid clues of the single-plant region and the convergence region clues of the extended lines, a reliability judgment is made, and the center point coordinates of the target crop are output accordingly, specifically including: S21. Perform morphological skeletonization processing on the binary image of the leaf mask to obtain a linear skeleton with a single pixel width; at the same time, extract the binary image of the leaf mask to obtain the mask contour curve; count the number of neighbors of the skeleton pixels based on 8-connected neighborhood, and divide the skeleton points into endpoints, ordinary points and connection points, wherein the number of neighbors of the endpoints is 1, the number of neighbors of the ordinary points is 2, and the number of neighbors of the connection points is not less than 3. S22. Construct an undirected graph with skeleton points as vertices and 8-connectivity relationships as edges; based on the undirected graph, perform shortest path extraction between skeleton endpoint pairs to obtain a candidate shortest path set, wherein the shortest path extraction adopts breadth-first search; when there is an equal-length shortest path or when geometric smoothness needs to be considered, for the equal-length shortest path obtained by the breadth-first search, use the cumulative path length calculated by the Euclidean distance between adjacent skeleton points as the reordering criterion to determine the key path point sequence; and when the skeleton has multiple branches, determine the blade main path from the candidate shortest paths according to the preset geometric evaluation criteria; S23. The cumulative arc length parameterization of the path point sequence of the main path of the blade is normalized to the [0,1] interval, and interpolation fitting is performed based on the parameterization result to obtain the blade center fitting curve. S24. Perform contour calculation on the binary image of each blade mask to obtain the minimum circumcircle curve; take the two endpoints of the fitted curve at the center of the blade as the starting point, search outward along the tangent direction at the endpoints for the intersection point with the minimum circumcircle and / or the mask contour as the ending point, connect the starting point and the ending point to form an extended line segment, and define the blade outward direction with the extended line segment. S25. Perform element overlay display: The binary image of the blade mask, the mask contour curve obtained in step S21, the blade center fitting curve obtained in step S23, and the minimum circumcircle curve mentioned in step S24 are directly copied and overlaid using pixels; the extended line segment pixels are covered and placed on top of each other to highlight the intersection point. S26. Then, perform morphological closing operation on the binary image of the leaf mask, merge adjacent or slightly broken single leaf regions to obtain single crop regions; calculate the centroid cues for each single crop region. S27. Extract candidate convergence positions from the extended line segments obtained in step S24 to obtain convergence region clues that characterize the degree of convergence in the blade's outward direction. S28. Based on the centroid clues of the single crop region obtained in S26 and the convergence region clues obtained in S27, construct at least two candidate centers and perform reliability judgment, and output the coordinates of the target crop center point.
2. The method as described in claim 1, characterized in that, S1 includes: S11. Construct and train a crop leaf detection model; S12. Input the acquired consecutive multi-frame RGB images into the crop leaf detection model to obtain leaf segmentation contour information; S13. Normalize the blade segmentation contour information and convert it into a binary image of the blade mask.
3. The method as described in claim 2, characterized in that, The crop leaf detection model in S11 includes an image input module, a feature extraction module, a feature fusion module, and a prediction output module; wherein the feature extraction module and / or the feature fusion module includes ODConv_3rd convolutional units; specifically, using YOLOv8 as the baseline framework, at least some of the standard convolutional layers are replaced with ODConv_3rd dynamic convolutions, and a weighted combination of DFL distributed focal loss and WIoU regression loss is used as the regression loss function during the training phase, and the crop leaf detection model is trained based on a dataset with crop leaf polygon annotations.
4. The method as described in claim 3, characterized in that, The dataset with polygonal annotations on crop leaves was obtained by: collecting field images of crop leaves under complex lighting, shading, and weed interference scenarios, and annotating the crop leaves in the images with polygons; and dividing the dataset into training set, validation set, and test set according to a preset ratio. The ODConv_3rd convolutional unit includes a dynamic convolutional layer, a normalization layer, and an activation layer; wherein the dynamic convolutional layer is used to generate multidimensional attention coefficients based on input features to dynamically adjust the convolution weights and / or biases to enhance feature adaptability, taking into account the characteristics of simple leaf textures, clear edges, but susceptibility to light effects.
5. The method as described in claim 1, characterized in that, The preset geometric evaluation criteria in S22 include at least the maximum arc length criterion, wherein the blade main path is the path with the maximum arc length in the candidate shortest path set, and the arc length is calculated based on the cumulative Euclidean distance between adjacent skeleton points.
6. The method as described in claim 1, characterized in that, The interpolation fitting described in S23 uses B-spline interpolation; and the B-spline order is 3 to 5, with endpoints using repeated endpoint nodes.
7. The method as described in claim 1, characterized in that, The convergence region clues of the extended line segments mentioned in S27 are obtained by extracting candidate convergence points from the extended line segments. The extraction of candidate convergence points includes pixel overlap detection and / or nearest neighbor detection. A convergence envelope is generated based on the set of candidate convergence points. The convergence envelope is an axis-aligned minimum bounding rectangle and / or an oriented minimum bounding rectangle. After extracting the set of candidate convergence points, aggregation denoising processing is performed on it. The aggregation denoising processing adopts a density clustering algorithm, connected component analysis, or a merging strategy based on a distance threshold. The density clustering algorithm is the DBSCAN algorithm.
8. The method as described in claim 1, characterized in that, The reliability determination described in S28 is based at least on the dispersion of the extended line segment direction and / or the size of the candidate convergence point and / or the shape of the convergence envelope; when the convergence region is determined to be reliable, the representative point of the convergence envelope is output as the center point coordinates, and the representative point is the geometric center of the convergence envelope or the centroid of the overlapping area between the convergence envelope and the single plant region; when the convergence region is determined to be unreliable, the geometric centroid of the single plant region is output as the center point coordinates.
9. A crop seedling center positioning system based on leaf skeleton-driven geometric analysis, characterized in that, The system includes a crop center positioning processor and a program or instructions stored on the crop center positioning processor, wherein the program or instructions, when executed, are used to implement the method described in any one of claims 1 to 8. The system also includes an image acquisition device for acquiring field crop images or video streams and sending them to the crop center positioning processor; and a visualization device for displaying the center positioning results, including center point coordinates and / or overlaid visualization images.
10. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method according to any one of claims 1 to 8.