Automatic detection method and system for soybean mixed plants based on botanical morphological characteristics

Abstract

The application belongs to the technical field of automatic detection, and particularly relates to a soybean hybrid plant automatic detection method and system based on botanical morphological characteristics, which comprises data acquisition, extraction of pixel-level contours of soybean leaflets and flowers and phenotype categories; construction of leaf vein direction vectors and calculation of geometric parameters based on morphological characteristics by using the pixel-level contours; completion of preliminary grouping of the leaflets; weighted calculation of comprehensive leaf types based on the classification confidence of each leaflet in the preliminary grouping, verification of whether there are abnormal leaflets with conflicting phenotype characteristics in the preliminary grouping by using the comprehensive leaf types, and correction of the abnormal leaflets; and judgment of the soybean hybrid plants by comparing the standard feature library of the target breeding variety with the soybean hybrid plants by using the dual-trait cross verification mechanism of the comprehensive leaf types and flower colors. The application overcomes the vulnerability of single morphological characteristics to environmental stress interference, and realizes high-precision screening of field soybean hybrid plants with high robustness and low misjudgment rate without increasing the cost of three-dimensional hardware.

Description

Technical Field

[0001] This application belongs to the field of automated detection technology, specifically relating to an automated detection method and system for soybean hybrid plants based on botanical morphological characteristics. Background Technology

[0002] Soybeans are an important dual-purpose crop for grain and oil, and the purity of seed production directly affects the final yield and quality of soybeans. The purity of soybean primary seeds needs to reach over 99.9%. Even the slightest contamination by a hybrid plant can lead to widespread yield reduction and substandard seed quality. Traditional methods of detecting hybrid plants in soybean breeding rely entirely on manual field inspections. This method is not only labor-intensive and costly, but also prone to missed or false detections due to subjective differences in judgment standards among workers. It completely fails to meet the demands of modern seed industry for large-scale, high-precision screening.

[0003] In recent years, with the development of computer technology, many studies have begun to explore automated methods for detecting hybrid plants in soybeans. Existing methods primarily rely on leaf shape or flower color to identify plant types and remove hybrids. However, when detecting leaf shape, the extremely high planting density of soybeans in open fields or greenhouses often leads to severe overlap and intersection of branches and leaves between adjacent plants. Conventional visual segmentation methods struggle to accurately separate leaves belonging to the same soybean plant from the cluttered field of view, easily mistaking leaves from adjacent hybrid plants that have overstepped the boundary for organs of the target plant, resulting in structural confusion. Furthermore, existing soybean leaf shape detection methods often mechanically rely on two-dimensional geometric parameters (such as aspect ratio) to determine leaf shape. However, in actual top-down imaging, the varying growth angles and postures of each leaflet in three-dimensional space cause severe perspective distortion in the image. Combined with reflections and shadows under natural lighting, normal leaves often appear deformed, "shorter," or "longer" in two-dimensional images, easily leading to misclassification of individual leaflets. The phenotype of a single plant organ is highly susceptible to external environmental stress factors such as drought, pests and diseases, and soil fertility (for example, the leaves of normal soybeans will become narrower and smaller under drought conditions). Relying on a single dimension such as leaf shape or flower color cannot accurately identify soybean hybrids, resulting in extremely poor overall robustness of the system and a high false alarm rate.

[0004] To address the problems of existing methods in densely planted conditions, such as errors in single-plant structure reconstruction due to overlapping and shading of leaves from adjacent plants, distortion of single-leaf classification due to perspective angle distortion, and high misjudgment rate in single-trait detection. Summary of the Invention

[0005] This application provides an automated detection method and system for soybean hybrid plants based on botanical morphological characteristics, which solves the problems of existing methods in dense planting conditions, such as errors in single plant structure reconstruction due to overlapping and occlusion of leaves of adjacent plants, distortion of single leaf classification due to perspective angle distortion, and high misjudgment rate in single trait detection.

[0006] The first aspect of this application provides an automated detection method for soybean hybrid plants based on botanical morphological characteristics, including: Data acquisition includes: controlling the camera to acquire top view and side view images at the top and side of the plant, respectively, to obtain ectopic multi-trait image pairs of the same plant, wherein the top view image is used to capture the structure of soybean compound leaves, and the side view image is used to capture the distribution of flowers. Feature extraction includes: inputting the heterotopic multi-phenotype image into a lightweight YOLO instance segmentation model and a lightweight YOLO pose estimation model to extract pixel-level contours and phenotypic categories of soybean leaves and flowers; using pixel-level contours and constructing leaf midrib orientation vectors and calculating geometric parameters based on morphological features, and simultaneously extracting flower color features and vertical height information; Preliminary grouping, which includes: traversing all leaves within the field of view, calculating the spatial physical distance between adjacent leaves and the angle between the leaf's main vein direction, utilizing the first geometric topological constraint of the radially diverging growth of soybean pinnate trifoliate leaves, removing parallel interfering leaves caused by overlapping and shading of other plants, and completing the preliminary grouping of leaves; The structural reconstruction includes: calculating the comprehensive leaf type based on the classification confidence of each leaflet in the initial group; using the second botanical morphological prior constraint that the terminal leaflet and the lateral leaflet are highly similar in morphology; verifying whether there are abnormal leaves with contradictory phenotypic characteristics in the initial group using the comprehensive leaf type; and correcting the abnormal leaves using the comprehensive leaf type to complete the reconstruction of the characteristics and structure of a single plant. Hybrid identification includes: extracting the vertical height information of flowers in the side view field, binding the color of the side view flowers with the corresponding compound leaf structure through spatial vertical projection, and comparing it with the standard feature library of the target breeding variety using a dual-trait cross-validation mechanism that combines leaf shape and flower color. If any core trait feature does not match, it is directly identified as a soybean hybrid.

[0007] Furthermore, the construction of the leaf midrib orientation vector based on morphological features includes: Using a lightweight YOLO instance segmentation model to obtain the first instance within the current field of view The bounding box of a soybean leaflet and the coordinates of its center; The key points of soybean leaves were predicted using a lightweight YOLO attitude estimation model. The key points include the petiole base node near the main stem and the leaf tip node extending outward. The center coordinates of the key points were calculated. Set a spatial matching distance threshold, using the principle of minimizing the Euclidean distance between two center coordinates, and perform cross-model fusion matching of bounding box features and keypoints to accurately obtain the first... The petiole base node and leaf tip node corresponding to a single soybean leaflet; Pointing the petiole base node to the leaf tip node, we construct a main vein orientation vector that represents the true growth direction of the leaf.

[0008] Furthermore, the computational geometry parameters include: obtaining the horizontal orthogonal bounding box of soybean leaflets, extracting the absolute width and absolute height, and calculating the diagonal length as a physical dimension parameter of the leaflets.

[0009] Further, preliminary topological grouping specifically includes: Locking in the core leaflet to be grouped Candidate leaflets within the search range Calculate the Euclidean distance between the base nodes of the two blades. : Calculate the absolute angle between the main vein direction vectors of the two blades. : Calculate the average size of the two blades And introduce a dynamic distance tolerance coefficient. Set the threshold for the divergence angle in botany , Core leaflet geometric parameters, For candidate Xiaoye Geometric parameters; Only when Euclidean space distance is satisfied And satisfy the absolute included angle When both leaves conform to the divergent growth pattern of soybean compound leaves, they are placed in the same initial group; if If a leaf is found to be an interfering leaf that is obstructing the boundary of an adjacent plant, it will be removed.

[0010] Furthermore, the comprehensive leaf type is calculated based on the classification confidence scores of each leaflet within the initial group, including: For those in the same initial group For each leaf segment, the classification confidence score output by the lightweight YOLO instance segmentation model is extracted and weighted for voting based on the corresponding category. The category with the highest cumulative weight is taken as the comprehensive leaf shape determination result of the compound leaves in the initial group.

[0011] Furthermore, a comprehensive leaf type verification is used to check whether there are abnormal leaves with contradictory phenotypic characteristics within the initial grouping, including: When the individual phenotypic category of a leaflet within a group contradicts the determined composite leaf type, the leaflet is determined to be a homologous leaflet, and the classification result of the leaflet is corrected based on the overall characteristics using the composite leaf type.

[0012] Furthermore, the classification result of the leaflet based on the comprehensive leaf type is corrected according to the overall characteristics, including: Integrate the standard category template for leaf shape; The original classification results of abnormal leaves are forcibly replaced with the category label corresponding to the comprehensive leaf type.

[0013] Furthermore, the vertical height information of the flower in the side view is extracted, and the color of the side-view flower is positionally bound to the corresponding compound leaf structure through spatial vertical projection, including: The vertical height information of each flower detected in the side view image is extracted. The vertical height information is represented by converting the ground as the reference zero point and pixel coordinates into actual physical distance. Extract the three-dimensional coordinate information of each reconstructed compound leaf structure from the top view image. The planar coordinates are obtained by converting the pixel position of the leaf base node in the top view image through camera calibration, and the vertical coordinates are obtained by the known physical distance difference between the top view camera and the plant canopy. Using the main stem of the plant as a spatial reference axis, a spatial coordinate mapping relationship is established between the top view and the side view images. The vertical height of the side view flower and the vertical coordinate of the top view compound leaf are matched with the nearest neighbor.

[0014] Furthermore, a dual-trait cross-validation mechanism combining leaf type and flower color was employed, comparing the results with the standard characteristic library of the target breeding variety. If any core trait characteristic did not match, the variety was directly identified as a soybean hybrid, including: The leaf type-flower color pairings stored in the standard feature library of the target breeding varieties are used as the sole criterion for judgment. The dual traits of leaf shape and flower color, which are bound together with the current plant, are extracted and compared with the judgment criteria item by item. If both traits match the criteria, the plant is identified as a pure line target plant. If either the overall leaf type matches but the flower color does not, or the flower color matches but the overall leaf type does not, a veto is triggered, and the plant is marked as a soybean hybrid.

[0015] A second aspect of this application provides an automated detection system for soybean hybrids based on plant morphological characteristics, used to execute an automated detection method for soybean hybrids based on plant morphological characteristics, including: The control device is used to control the camera to acquire top view and side view images of the plant at the top and side of the plant, respectively, to obtain ectopic multi-trait image pairs of the same plant. The top view image is used to capture the structure of soybean compound leaves, and the side view image is used to capture the distribution of flowers. The processor receives data collected by the camera and performs feature extraction, preliminary grouping, structural reconstruction, and hybrid identification.

[0016] Compared with the prior art, the advantages of this application are as follows: This application deeply integrates botanical prior laws with computer vision technology. First, it transforms the divergent growth characteristic of soybean's "pinnate trifoliate compound leaf" into a geometric topological constraint of distance and the angle between the midrib, accurately eliminating interfering leaves from adjacent hybrid plants that cross boundaries in high-density planting. Second, it innovatively proposes a comprehensive voting + local prior correction mechanism. Using the comprehensive leaf shape obtained by confidence weighting, it forcibly corrects homologous single leaves within the group that are misclassified due to perspective distortion caused by overhead photography, strongly correcting local visual errors and achieving accurate reconstruction of single plant features. Compared with independent classification relying on single leaf features, the feature recognition accuracy of this method is significantly improved. Finally, by constructing a physical spatial binding between top-view leaf shape and side-view flower color, and introducing a dual-trait cross-validation and veto mechanism, it completely overcomes the vulnerability of single morphological features to environmental stress interference. Without increasing the cost of 3D hardware, it achieves highly robust and low-misclassification-rate automated high-precision screening of hybrid soybean plants in the field. Attached Figure Description

[0017] Figure 1 A schematic diagram of the overall process of an automated detection method for soybean hybrids based on botanical morphological characteristics provided in this application embodiment; Figure 2 A schematic diagram of the spatial physical location of the collected ectopic multi-trait images of soybean plants provided in the embodiments of this application; Figure 3 A schematic diagram of the main vein orientation vector and bounding box of soybean leaflets provided in the embodiments of this application; Figure 4 This is a schematic diagram of geometric topological constraints grouping of soybean pinnate trifoliate leaves based on Euclidean spatial distance and absolute angle provided in the embodiments of this application, wherein (a) is geometric topology based on Euclidean spatial distance and (b) is geometric topology based on absolute angle. Figure 5 Images of soybean leaflets collected on-site for embodiments of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] This application provides an automated detection method and system for soybean hybrid plants based on botanical morphological characteristics. Its core lies in transforming the divergent growth pattern of soybean pinnate trifoliate leaves into a strict geometric topological constraint, and combining overall confidence voting and local distortion correction mechanisms to completely solve the problems of leaf overlap and occlusion and visual distortion under high-density planting.

[0020] It should be noted that each compound leaf of soybean consists of three leaflets, and the positional distribution of the three leaflets follows a fixed pattern: one leaflet is located at the top of the leaf axis and is called the terminal leaflet; the other two leaflets are located on both sides of the leaf axis and are called lateral leaflets. In this article, the leaflets refer to the leaflets in the compound leaf, and the leaf blade is used to represent the leaf-like targets detected in the image.

[0021] This application embodiment provides an automated detection system for soybean hybrid plants based on botanical morphological characteristics, including: a control device for controlling a camera to successively acquire top view images and side view images at the top and side of the plant, thereby obtaining ectopic multi-trait image pairs of the same plant, wherein the top view image is used to capture the structure of soybean compound leaves, and the side view image is used to capture the distribution of flowers. The processor receives data collected by the camera and performs feature extraction, preliminary grouping, structural reconstruction, and hybrid identification.

[0022] The control device is used to achieve fully automated control of the image acquisition process, ensuring the acquisition of heterotopic multi-trait image pairs of the same plant from two different perspectives: top and side. Specifically, the control device includes a motion control unit, a timing control unit, a trigger control unit, and a communication unit.

[0023] The motion control unit is used to control the trajectory movement of the camera in three-dimensional space. This unit includes at least two control states: In the first control state, the motion control unit drives the camera to a preset height directly above the target plant, keeping the camera's optical axis vertically downwards. The image acquired in this state is a top-view image, used to capture the compound leaf structure of the soybean plant. In the second control state, the motion control unit drives the camera to a preset distance laterally to the target plant, keeping the camera's optical axis horizontally towards the plant. The image acquired in this state is a side-view image, used to capture the flower distribution and vertical height information of the plant.

[0024] The timing control unit is used to set the time interval between the acquisition of top-view and side-view images. In actual operation, after the camera completes the acquisition of the top-view image, it waits for the camera to complete the spatial position change before triggering the acquisition of the side-view image, according to the delay parameters preset by the timing control unit (usually in the millisecond to second range, depending on the speed of mechanical movement and system response time), to ensure that the images from the top and side views belong to the same plant in the same growth state.

[0025] The trigger control unit sends precise acquisition trigger signals to the camera. Once the motion control unit confirms that the camera has reached the predetermined acquisition position and its attitude is stable, the trigger control unit outputs a level signal or software command to drive the camera to perform exposure and image acquisition operations. Simultaneously, the trigger control unit is also responsible for synchronously recording the spatial coordinates and acquisition timestamp of each frame, forming traceable metadata.

[0026] The communication unit is responsible for packaging and transmitting the acquired top-view and side-view images, along with their associated metadata (including spatial coordinates, timestamps, camera pose angles, etc.), to the processor for subsequent steps such as feature extraction, topology grouping, structure reconstruction, and heterogeneous identification. The communication unit can use wired interfaces (such as USB, GigE, Camera Link) or wireless interfaces (such as Wi-Fi, 5G) to achieve efficient data transmission.

[0027] In one specific embodiment, the control device can also integrate position feedback and closed-loop adjustment functions. The motion control unit dynamically adjusts the drive parameters based on the real-time feedback of the camera position information to ensure that the positioning accuracy of the camera between the top and side positions is better than a preset error tolerance (e.g., ±2 mm), laying the physical foundation for the subsequent spatial projection association between the top view image and the side view image.

[0028] The processor receives image data and metadata transmitted from the control device and executes the core algorithm flow for soybean hybrid identification. The processor can be a computing unit in an embedded processor (such as ARM or DSP), an industrial computer, an edge computing device, or a cloud server, and its internal components are loaded with computer program instructions that implement the method of this application.

[0029] The processor first preprocesses the received top-view and side-view images, including image denoising, illumination normalization, and distortion correction. Then, the processor calls a lightweight YOLO instance segmentation model to extract pixel-level contours and phenotypic category information (terminal or lateral leaflets) of soybean leaves from the top-view image, and pixel-level contours and color category information (purple or white flowers) of flowers from the side-view image. The processor then calls a lightweight YOLO pose estimation model to predict keypoints of the soybean leaves. These keypoints include the petiole base node near the main stem and the leaf tip node extending outwards, and the processor calculates the center coordinates of these keypoints.

[0030] Based on this, the processor performs the following computational tasks: First, for each extracted leaflet, the midrib is extracted based on morphological features, the basal and apical nodes are located, a midrib orientation vector representing the leaf's growth direction is constructed, and geometric parameters reflecting the absolute size of the leaf (such as the diagonal length of the bounding box) are calculated. Simultaneously, the flower's color features and vertical height information are extracted from the side view image.

[0031] Second, the processor iterates through all leaflets within the field of view, calculating the spatial physical distance between adjacent leaflets and the angle between the midrib direction. The processor introduces the first layer of geometric topological constraints, namely the prior knowledge that soybean pinnate trifoliate leaves grow radially, and removes parallel interfering leaves caused by overlapping and shading of adjacent plants, thus completing the initial topological grouping of the leaves.

[0032] Third, the classification confidence scores of multiple leaflets within the initial group are extracted and weighted to determine the comprehensive leaf type of the compound leaves in that group. Subsequently, a second botanical morphological prior constraint is introduced—the terminal and lateral leaflets of soybean pinnate trifoliate leaves are highly similar in morphology—to verify whether there are any abnormal leaves with contradictory phenotypic characteristics within the initial group. For contradictory leaves, the processor does not simply discard them, but identifies them as homologous leaves whose classification is distorted due to factors such as camera angle distortion, occlusion, or reflection. The comprehensive leaf type is then used to correct the classification results, thereby completing the accurate reconstruction of the characteristics and structure of individual plants.

[0033] Fourth, the processor combines spatial vertical projection relationships to spatially bind the flower color detected in the side view image with the reconstructed compound leaf structure in the top view image. Specifically, using the main stem of the plant as the spatial reference axis, the vertical height of the side view flower is matched with the vertical coordinate of the top view compound leaf node using nearest neighbor matching. When the vertical distance between the two does not exceed a preset tolerance threshold, it is determined that the flower and the corresponding compound leaf belong to the same node, thus completing the physical association between flower color and leaf shape.

[0034] Fifth, the processor extracts two core traits bound to the target variety—comprehensive leaf type (divided into long and short leaf categories) and flower color (divided into purple and white flower categories)—and performs targeted cross-validation with the standard feature library of the target variety. The processor follows a veto mechanism: if any deviation occurs in either "normal leaf type but mismatched flower color" or "normal flower color but mismatched leaf type," the plant is directly identified as a soybean hybrid, and a detection alarm signal is output, or the subsequent mechanical execution device (such as a pneumatic rejection mechanism or a robotic arm sorting device) is driven to complete the automated rejection action.

[0035] In a preferred embodiment, the processor also has result storage and log recording functions. The judgment result, feature data, original image and intermediate processing result of each detection are written to a local database or cloud storage for subsequent variety purity statistical analysis, model iteration training and traceability query.

[0036] The control device and processor work in concert according to the following sequence: After system startup, the control device receives a detection start command from the host computer or operation panel; the control device drives the camera to the top position, triggering the acquisition of a top view image and transmitting it to the processor; the control device drives the camera to change its spatial position to the side position, waits for a preset time interval, and then triggers the acquisition of a side view image and transmits it to the processor; after receiving a complete pair of images (top view image + side view image), the processor sequentially performs feature extraction, topological grouping, structural reconstruction, spatial binding, dual-trait cross-validation, and veto determination; the processor outputs the determination result (qualified / hybrid) and the corresponding control signal (such as a rejection command); if it is a hybrid, the control device can further trigger the rejection mechanism; if it is a qualified plant, the system continues to process the next crop. The above process is repeated until the automated detection and sorting of all plants to be tested is completed.

[0037] Based on the above system, an automated detection method for soybean hybrids based on plant morphological characteristics is described in [reference needed]. Figure 1 As shown, it includes: S101, Data acquisition, the data acquisition includes: controlling the camera to successively acquire top view images and side view images at the top and side of the plant, and obtaining ectopic multi-trait image pairs of the same plant, wherein the top view image is used to capture the structure of soybean compound leaves, and the side view image is used to capture the distribution of flowers. S102, Feature extraction, the feature extraction includes: inputting the heterotopic multi-phenotype image into a lightweight YOLO instance segmentation model and a lightweight YOLO pose estimation model, extracting pixel-level contours and phenotypic categories of soybean leaves and flowers; using pixel-level contours and constructing leaf midrib orientation vectors and calculating geometric parameters based on morphological features, simultaneously extracting flower color features and vertical height information; S103, preliminary grouping, the preliminary grouping includes: traversing all leaves in the field of view, calculating the spatial physical distance between adjacent leaves and the angle between the direction of the leaf's main vein, using the first geometric topological constraint of the radially diverging growth of the soybean's pinnate trifoliate compound leaves, removing parallel interfering leaves caused by overlapping and shading of other plants, and completing the preliminary grouping of leaves. S104, Structural reconstruction, the structural reconstruction includes: calculating the comprehensive leaf type based on the classification confidence of each leaflet in the preliminary group, then using the second botanical morphological prior constraint that the terminal leaflet and the lateral leaflet are highly similar in morphology, using the comprehensive leaf type to check whether there are abnormal leaves with contradictory phenotypic characteristics in the preliminary group, and using the comprehensive leaf type to correct the abnormal leaves, thus completing the reconstruction of the characteristics and structure of a single plant. S105, Hybrid plant determination, the hybrid plant determination includes: extracting the vertical height information of flowers in the side view field, binding the color of the side view flower with the corresponding compound leaf structure through spatial vertical projection, and comparing it with the standard feature library of the target breeding variety using a dual-trait cross-validation mechanism that combines leaf shape and flower color. If any core trait feature does not match, it is directly determined to be a soybean hybrid plant.

[0038] In step S101, the camera acquires ectopic multi-morphological image pairs through time-division displacement acquisition: (Refer to...) Figure 2 As shown, after the control device wakes up the host through the automated detection equipment, the host first supplies power to the single-channel camera. The camera is then controlled to capture a top-view image containing the structure and shape characteristics of soybean compound leaves directly above the plant. Subsequently, the camera position is moved to a designated location to the side of the plant. Once the image stabilizes, the camera is triggered again to capture a side-view image containing the distribution and color characteristics of soybean flowers. After both acquisitions are completed, the system extracts the spatial displacement of the camera and the time difference between captures for temporal verification, ensuring that the two acquisitions strictly correspond to the same plant. After successful verification, the top-view and side-view images of the same plant are spatiotemporally matched and bound, merging them to form ectopic multi-trait image pairs.

[0039] In step S102, the ectopic multi-phenomenon image obtained in step S101 is input into a lightweight YOLO instance segmentation model (e.g., YOLOv8 or MobileNet architecture) to quickly extract pixel-level contours (including bounding boxes and center coordinates) and phenotypic categories (e.g., long leaves / short leaves, purple flowers / white flowers) of soybean leaves and flowers. The pixel-level contours (including bounding boxes and center coordinates) and phenotypic categories (e.g., long leaves / short leaves, purple flowers / white flowers) of soybean leaves and flowers are then extracted by the lightweight YOLO instance segmentation model. The bounding box and center coordinates of a soybean leaf, combined with Figure 3 Based on morphological features, a leaf midrib orientation vector is constructed, including: Using a lightweight YOLO instance segmentation model to obtain the first instance within the current field of view The bounding box of a soybean leaflet and the center coordinates D of the bounding box; A lightweight YOLO attitude estimation model was used to predict key points of soybean leaves. These key points include node A at the base of the petiole near the main stem and node B extending outwards from the leaf tip. The center coordinates E of these key points were calculated. Figure 3 In this context, A represents the basal node of the petiole. Figure 3 In this context, B represents the leaf tip node, and E represents the center coordinates of the key points. A spatial matching distance threshold is set, and a cross-model feature center point distance matrix is ​​constructed. The two-dimensional Euclidean spatial distance between the center coordinates D of the bounding box of the i-th soybean leaflet and the center coordinates E of all key points in the current field of view is calculated. A dynamic spatial matching distance threshold is set, which is adaptively correlated with the feature scale of the bounding box. The optimal binary matching algorithm is used to traverse the distance matrix. If and only if the Euclidean distance between the center coordinates E of a certain key point and the center coordinates D of the bounding box is less than the set spatial matching distance threshold, and the distance is the smallest among all candidate groups, it is determined that the feature of the bounding box and the key points of that group belong to the same entity leaf, and the cross-model fusion matching is completed. Thus, the coordinates of the corresponding key points are physically assigned to the soybean leaflet, and the precise petiole base node and precise leaf tip node corresponding to the i-th soybean leaflet are obtained. By pointing the precise petiole base node to the precise leaf tip node, a main vein orientation vector C representing the true growth direction of the leaf is constructed. Figure 3 The Chinese character is represented by C.

[0040] The computational geometry parameters include: obtaining the horizontal orthogonal bounding box of soybean leaflets, extracting the absolute width and absolute height, and calculating the diagonal length as the physical size parameter of the leaflets.

[0041] The specific process is as follows: First, the collected heterotopic multimorphic image pairs are input into the lightweight YOLO instance segmentation model and the lightweight YOLO pose estimation model, respectively; Using a lightweight YOLO instance segmentation model to obtain the first instance within the current field of view The bounding box and center coordinates D of a soybean leaflet are given. The lightweight YOLO instance segmentation model includes a backbone network, a neck network, and a head. The backbone network uses a lightweight CSPDarknet architecture to extract multi-scale basic features of the image through cross-stage local network layers. The neck network uses a path aggregation network (PANet) to achieve bidirectional fusion of high-level semantic information and low-level spatial information. The head contains two parallel branches that output the detection box regression parameters and the prototype coefficients of the pixel-level mask, respectively, which are then multiplied by matrix to generate the bounding box and the binarized mask. Keypoints of soybean leaves, including node A at the base of the petiole near the main stem, were directly predicted and extracted using a lightweight YOLO pose estimation model. This is denoted as... And the outwardly extending leaf tip node B, denoted as The center coordinates E of the key points in the set are calculated. The backbone and neck networks of the lightweight YOLO pose estimation model are consistent with the aforementioned instance segmentation model to achieve lightweight deployment. Its detection head is a key point prediction decoupling head (Pose Head), which directly outputs the two-dimensional pixel coordinates and visibility confidence of the topological key points contained in the target object through an additional regression branch.

[0042] A spatial matching distance threshold is set, and based on the principle of minimizing the Euclidean distance between the two center coordinates, the bounding box features extracted by the lightweight YOLO instance segmentation model and the keypoints extracted by the lightweight YOLO pose estimation model are fused and matched across models to obtain the first... Precise petiole base node corresponding to a single soybean leaflet With precise leaf tip node ; Subsequently, the precise petiole base node With precise leaf tip node The main vein orientation vector C is constructed to represent the actual growth direction of the leaf: , This represents the vector indicating the direction of the main vein.

[0043] See Figure 5 As shown, the bounding boxes of the three leaflets were collected and processed, along with the obtained main vein direction vector. The main vein direction vector is consistent with the actual main vein direction vector.

[0044] Finally, obtain the horizontal orthogonal bounding box of the leaf output by the lightweight YOLO instance segmentation model, and extract its absolute width. with absolute height Calculate the diagonal length of the horizontal orthogonal circumbound bounding box. As a geometric parameter reflecting the physical dimensions of the blade: , Simultaneously extract the color features and vertical height information of the flowers in the side field of view.

[0045] In step S103, the first-level topological preliminary grouping is performed using the radially diverging growth of soybean pinnate trifoliate leaves: to eliminate interference caused by overlapping of adjacent hybrid plants, the system traverses all leaves within the field of view to perform topological connections. The specific grouping principle is as follows: Figure 4 As shown. Locking the core leaflets to be grouped. Candidate leaflets within the search range Calculate the Euclidean distance between the base nodes of the two blades. : , Among them, the core leaflet The coordinates are Candidate Xiaoye The coordinates are .

[0046] Simultaneously calculate the absolute angle between the main vein direction vectors of the two blades. : , in, Indicates the first The vector of the main vein of a soybean leaf. Indicates the first The vector of the main vein of a soybean leaf.

[0047] The system calculates the average dimensions of the two blades. ,in Core leaflet geometric parameters, For candidate Xiaoye The geometric parameters are determined, and a dynamic distance tolerance coefficient is introduced. (Usually ranging from 0.15 to 0.3, depending on the specific soybean breeding variety), while simultaneously setting a botanical divergence angle threshold. The decision logic is as follows: The distance constraint is satisfied if and only if... And simultaneously satisfy the included angle constraint When both leaves conform to the divergent growth pattern of soybean compound leaves, they are placed in the same initial group. This indicates that the main vein direction vectors are highly parallel, and even if they are close, they are identified as interfering leaves that are obstructing the boundary of adjacent hybrid plants and are clearly removed, thus completing the initial grouping of high purity.

[0048] See Figure 4 The dimension of the dashed circle in (a) is It satisfies the distance constraint, i.e., Euclidean space distance. They belong to the same initial group, and the two are valid topologically connected edges. When this condition is not met, an edge rejection and truncation is triggered. See Figure 4 As shown in (b) in the figure, When this occurs, it indicates that the main vein's direction vector is highly parallel, triggering the rejection and truncation of edges, and the two do not become related.

[0049] In step S104, distortion correction and structural reconstruction based on comprehensive leaf type voting and prior assumptions are performed: In actual field photography, overhead shots easily cause perspective distortion in normally growing leaves, leading to single-leaf classification distortion. Therefore, this application innovatively introduces a "comprehensive voting + local prior correction" mechanism, calculating the comprehensive leaf type based on the classification confidence of each leaflet within the initial group. This includes: for leaves within the same initial group... For each leaf, the classification confidence score output by the lightweight YOLO instance segmentation model is extracted and weighted for each category through a weighted voting process. The leaf with the highest cumulative weight among all categories is taken as the comprehensive leaf shape determination result within the initial group. For leaves within the same initial group... Extracting classification confidence from the output of a lightweight YOLO instance segmentation model using leaflets. Perform corresponding categories Weighted cumulative voting based on (long-leaf / short-leaf) characteristics: , in This is Xiaoye's number.

[0050] The system selects the category with the highest cumulative weight as the comprehensive leaf type determination result for that group of compound leaves, thus strongly correcting classification misclassifications caused by reflections or shadows on individual leaflets. Subsequently, a second layer of botanical morphological prior constraints is introduced: a standard soybean pinnate trifoliate compound leaf consists of one terminal leaflet and two lateral leaflets, and "within the same compound leaf, the terminal leaflet and lateral leaflets are highly similar in morphology." The system checks whether there are any abnormal leaves within the initial group whose individual phenotypic category severely contradicts the determined comprehensive leaf type. If such leaves exist (i.e., abnormal leaves), considering that physical homology screening has already been completed, the system does not crudely remove them. Instead, it classifies them as homologous leaves whose classification is distorted due to factors such as camera angle distortion, occlusion, or reflections, and corrects the classification result of the leaflet based on the overall characteristics using the comprehensive leaf type. This mechanism effectively overcomes local visual errors and achieves accurate reconstruction of the core features of a single plant and the compound leaf structure.

[0051] The classification result of the leaflet using the comprehensive leaf type is corrected based on overall characteristics. This means that the original classification result of the abnormal leaf is forcibly replaced with the category label corresponding to the comprehensive leaf type by using the comprehensive leaf type standard category template. The meaning of the standard category template is that the actual biological category of all leaflets in this group of compound leaves should be consistent with the standard category template, rather than being independent and unrelated.

[0052] The composite leaf type is obtained by statistically analyzing the classification confidence of each leaflet within a group, weighting the terminal and lateral leaflets separately, and selecting the one with the highest cumulative weight. In other words, the composite leaf type represents the high-confidence classification results of most leaves in that group of compound leaves, and is the most reliable judgment obtained by the system based on a collective voting mechanism. Using it as a standard category template reflects the technical idea of ​​using overall information to suppress local noise.

[0053] Once the standard category template is established, the system performs a forced replacement operation. Specifically, the system locates abnormal leaves that contradict the comprehensive leaf type, reads the original classification result of the leaf output by the lightweight YOLO instance segmentation model (e.g., "lateral leaflet"), and then directly changes the original classification result to the category label corresponding to the comprehensive leaf type (e.g., "terminal leaflet").

[0054] This replacement operation is mandatory and does not depend on the visual characteristics or confidence level of the abnormal leaf itself. Even if the original classification confidence of the abnormal leaf is low (e.g., due to blurred features caused by reflection) or high (e.g., due to misjudgment caused by camera angle distortion but the model itself has high confidence), the system performs label replacement in a uniform manner. The logical basis for this mandatory replacement is that the system fully trusts the physical homology judgment and the high-confidence vote of the majority of leaves, believing that the original misjudgment of the abnormal leaf was caused by local external interference, and its own characteristics are unreliable, therefore it should not be used as a basis for judgment.

[0055] After the above correction operations, the classification label of the abnormal leaf was unified to the type consistent with the comprehensive leaf type. The final grouping results show that all leaflets within the compound leaf exhibit a unified category attribute—either all are labeled as terminal leaflets or all are labeled as lateral leaflets. This result aligns with the principle of high morphological similarity between terminal and lateral leaflets in the second layer of botanical morphology prior constraints; that is, the system no longer strictly distinguishes between terminal and lateral leaflets (because they are indeed highly similar in actual biology), but rather eliminates classification noise through unified labeling.

[0056] It is important to note that this correction does not contradict biological facts. In soybean pinnate trifoliate leaves, although the terminal and lateral leaflets differ in location, they are indeed highly similar in morphology, especially in leaf size, shape, and color. Therefore, labeling all leaflets of the entire compound leaf as the same category (regardless of whether they are terminal or lateral) does not substantially affect subsequent cross-validation based on leaf type and flower color. The system's use of a comprehensive leaf type is itself a coarse-grained category label (such as long or short leaves), rather than strictly distinguishing between terminal and lateral leaves.

[0057] In step S105, a dual-trait cross-validation and veto mechanism are employed: After accurately reconstructing the compound leaf structure, the system extracts the vertical height information of the flower in the side view and spatially projects it onto the three-dimensional coordinates of the top-view compound leaf node, achieving a physical binding between flower color characteristics and compound leaf structure. Finally, the system extracts the two core traits bound as "comprehensive leaf type (long / short leaf)" and "flower color (purple / white flower)" and performs targeted cross-validation with the standard feature library of the target breeding variety. The system follows an extremely strict veto mechanism: if any trait deviation occurs, such as "normal leaf type but mismatched flower color" or "normal flower color but mismatched leaf type," the mechanism is directly triggered, clearly marking the plant as a soybean hybrid and outputting a detection alarm or driving subsequent mechanical automated removal actions.

[0058] This includes extracting the vertical height information of the flower in the side view field, and binding the color of the side-view flower to the corresponding compound leaf structure through spatial vertical projection, including: The vertical height information of each flower detected in the side view image is extracted. The vertical height information is represented by converting the actual physical distance with the ground as the reference zero point and pixel coordinates. Extract the three-dimensional coordinate information of each reconstructed compound leaf structure from the top view image. The planar coordinates are obtained by converting the pixel position of the leaf base node in the top view image through camera calibration, and the vertical coordinates are obtained by the known physical distance difference between the top view camera and the plant canopy. Using the main stem of the plant as a spatial reference axis, a spatial coordinate mapping relationship between the top view and the side view images is established, and the vertical height of the side view flower and the vertical coordinate of the top view compound leaf are matched with the nearest neighbor.

[0059] In one embodiment, a dual-trait cross-validation mechanism combining leaf type and flower color is used to compare the results with the standard feature library of the target breeding variety. If any core trait does not match, the variety is directly identified as a soybean hybrid, including: The leaf type-flower color pairings stored in the standard feature library of the target breeding varieties are used as the sole criterion for judgment. The dual traits of leaf shape and flower color, which are bound together with the current plant, are extracted and compared with the judgment criteria item by item. If both traits match the criteria, the plant is identified as a pure line target plant. If either the overall leaf type matches but the flower color does not, or the flower color matches but the overall leaf type does not, a veto is triggered, and the plant is marked as a soybean hybrid.

[0060] Specifically, the system pre-establishes a standard feature library for the target breeding variety, which stores the "leaf type-flower color" pairings of the variety. The so-called pairings refer to the comprehensive leaf type and flower color characteristics that the target variety should possess simultaneously under normal growth and development conditions.

[0061] Taking typical soybean varieties as an example: the standard pairing combination for some varieties is "long leaves - purple flowers," meaning that normal plants of this variety should exhibit long, slender leaves and purple flowers; the standard pairing combination for other varieties is "short leaves - white flowers," meaning that normal plants should exhibit short, round leaves and white flowers; and the standard pairing combination for still others is either "long leaves - white flowers" or "short leaves - purple flowers." The pairing combinations in the standard feature library are pre-defined and stored based on the pureline standards of the target breeding varieties, serving as the sole criterion for subsequent judgment.

[0062] It should be noted that the comprehensive leaf type here refers to the overall leaf type category of the compound leaf structure of the plant after the above corrections, including two basic types: long leaves and short leaves; flower color refers to the flower color category after extraction and spatial binding, including two basic types: purple flowers and white flowers. For a complete soybean plant, the system may detect multiple sets of compound leaves and multiple flowers. In actual judgment, the system takes the main characteristics of the plant as the standard, that is, it extracts the most frequently occurring comprehensive leaf type as the representative leaf type of the plant, and extracts the most frequently occurring flower color as the representative flower color of the plant, forming the "leaf type-flower color" bitrait combination of the plant.

[0063] After extracting the two traits, the system compares the current plant's "comprehensive leaf type" and "flower color" item by item with the target variety benchmark stored in the standard feature library. The comparison process consists of two independent steps: The first comparison step involves comparing the current plant's overall leaf shape with the standard leaf shape in the baseline. If the current plant's leaf shape matches the standard leaf shape, the plant passes the leaf shape comparison; otherwise, the comparison fails.

[0064] The second comparison involves comparing the flower color of the current plant with the standard flower color in the baseline. If the flower color of the current plant matches the standard flower color, the plant passes the flower color comparison; otherwise, the flower color comparison fails.

[0065] The two comparisons are independent of each other and do not interfere with each other, each producing a pass or fail result.

[0066] A veto mechanism is implemented: a plant is considered a qualified pureline only when both traits simultaneously meet the baseline requirements; however, if either trait deviates from the baseline, regardless of the degree of deviation or whether the other trait matches, the plant is immediately deemed a hybrid. The strictness of this mechanism stems from the stringent requirements for varietal purity in soybean breeding—any difference in the core phenotypic trait indicates possible natural hybridization, mechanical mixing, or decreased seed purity, and the plant should be discarded.

[0067] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An automated detection method for soybean hybrid plants based on botanical morphological characteristics, characterized in that, include: Data acquisition includes: controlling the camera to acquire top view and side view images at the top and side of the plant, respectively, to obtain ectopic multi-trait image pairs of the same plant, wherein the top view image is used to capture the structure of soybean compound leaves, and the side view image is used to capture the distribution of flowers. Feature extraction includes: inputting the heterotopic multi-phenotype image into a lightweight YOLO instance segmentation model and a lightweight YOLO pose estimation model to extract pixel-level contours and phenotypic categories of soybean leaves and flowers; using pixel-level contours and constructing leaf midrib orientation vectors and calculating geometric parameters based on morphological features, and simultaneously extracting flower color features and vertical height information; Preliminary grouping, which includes: traversing all leaves within the field of view, calculating the spatial physical distance between adjacent leaves and the angle between the leaf's main vein direction, utilizing the first geometric topological constraint of the radially diverging growth of soybean pinnate trifoliate leaves, removing parallel interfering leaves caused by overlapping and shading of other plants, and completing the preliminary grouping of leaves; The structural reconstruction includes: calculating the comprehensive leaf type based on the classification confidence of each leaflet in the initial group; using the second botanical morphological prior constraint that the terminal leaflet and the lateral leaflet are highly similar in morphology; verifying whether there are abnormal leaves with contradictory phenotypic characteristics in the initial group using the comprehensive leaf type; and correcting the abnormal leaves using the comprehensive leaf type to complete the reconstruction of the characteristics and structure of a single plant. Hybrid identification includes: extracting the vertical height information of flowers in the side view field, binding the color of the side view flowers with the corresponding compound leaf structure through spatial vertical projection, and comparing it with the standard feature library of the target breeding variety using a dual-trait cross-validation mechanism that combines leaf shape and flower color. If any core trait feature does not match, it is directly identified as a soybean hybrid.

2. The automated detection method for soybean hybrids based on botanical morphological characteristics according to claim 1, characterized in that, The method of constructing the leaf midrib orientation vector based on morphological features includes: Using a lightweight YOLO instance segmentation model to obtain the first instance within the current field of view The bounding box of a soybean leaflet and the coordinates of its center; The key points of soybean leaves were predicted using a lightweight YOLO attitude estimation model. The key points include the petiole base node near the main stem and the leaf tip node extending outward. The center coordinates of the key points were calculated. Set a spatial matching distance threshold, using the principle of minimizing the Euclidean distance between two center coordinates, and perform cross-model fusion matching of bounding box features and keypoints to accurately obtain the first... The petiole base node and leaf tip node corresponding to a single soybean leaflet; Pointing the petiole base node to the leaf tip node, we construct a main vein orientation vector that represents the true growth direction of the leaf.

3. The automated detection method for soybean hybrids based on botanical morphological characteristics according to claim 2, characterized in that, The computational geometry parameters include: obtaining the horizontal orthogonal bounding box of soybean leaflets, extracting the absolute width and absolute height, and calculating the diagonal length as the physical size parameter of the leaflets.

4. The automated detection method for soybean hybrids based on botanical morphological characteristics according to claim 1, characterized in that, Preliminary grouping, specifically including: Locking in the core leaflet to be grouped Candidate leaflets within the search range Calculate the Euclidean distance between the base nodes of the two blades. : Calculate the absolute angle between the main vein direction vectors of the two blades. : Calculate the average size of the two blades And introduce a dynamic distance tolerance coefficient. Set the threshold for the divergence angle in botany , Core leaflet geometric parameters, For candidate Xiaoye Geometric parameters; Only when Euclidean space distance is satisfied And satisfy the absolute included angle When both leaves conform to the divergent growth pattern of soybean compound leaves, they are placed in the same initial group; if If a leaf is found to be an interfering leaf that is obstructing the boundary of an adjacent plant, it will be removed.

5. The automated detection method for soybean hybrids based on botanical morphological characteristics according to claim 1, characterized in that, The comprehensive leaf type is calculated based on the classification confidence scores of each leaflet within the initial group, including: For those in the same initial group For each leaf segment, the classification confidence score output by the lightweight YOLO instance segmentation model is extracted and weighted for voting based on the corresponding category. The category with the highest cumulative weight is taken as the comprehensive leaf shape determination result of the compound leaves in the initial group.

6. The automated detection method for soybean hybrids based on botanical morphological characteristics according to claim 1, characterized in that, The comprehensive blade shape verification method is used to check whether there are abnormal blades with contradictory phenotypic characteristics within the initial grouping, including: When the individual phenotypic category of a leaflet within a group contradicts the determined composite leaf type, the leaflet is determined to be a homologous leaflet, and the classification result of the leaflet is corrected based on the overall characteristics using the composite leaf type.

7. The automated detection method for soybean hybrids based on botanical morphological characteristics according to claim 6, characterized in that, The classification result of the leaflet based on the comprehensive leaf type is corrected according to overall characteristics, including: Integrate the standard category template for leaf shape; The original classification results of abnormal leaves are forcibly replaced with the category label corresponding to the comprehensive leaf type.

8. The automated detection method for soybean hybrids based on botanical morphological characteristics according to claim 1, characterized in that, Extract the vertical height information of the flower in the side view field, and use spatial vertical projection to bind the color of the side-view flower with the corresponding compound leaf structure, including: The vertical height information of each flower detected in the side view image is extracted. The vertical height information is represented by converting the actual physical distance with the ground as the reference zero point and pixel coordinates. Extract the three-dimensional coordinate information of each reconstructed compound leaf structure from the top view image. The planar coordinates are obtained by converting the pixel position of the leaf base node in the top view image through camera calibration, and the vertical coordinates are obtained by the known physical distance difference between the top view camera and the plant canopy. Using the main stem of the plant as a spatial reference axis, a spatial coordinate mapping relationship between the top view and the side view images is established, and the vertical height of the side view flower and the vertical coordinate of the top view compound leaf are matched with the nearest neighbor.

9. The automated detection method for soybean hybrids based on botanical morphological characteristics according to claim 1, characterized in that, A dual-trait cross-validation mechanism combining leaf type and flower color was used to compare the results with the standard characteristic library of the target breeding variety. If any core trait characteristic did not match, the variety was directly identified as a soybean hybrid, including: The leaf type-flower color pairing combination stored in the standard feature library of the target breeding variety is used as the sole criterion for judgment. The dual traits of leaf shape and flower color, which are bound together with the current plant, are extracted and compared with the judgment criteria item by item. If both traits match the criteria, the plant is identified as a pure line target plant. If either the overall leaf type matches but the flower color does not, or the flower color matches but the overall leaf type does not, a veto is triggered, and the plant is marked as a soybean hybrid.

10. An automated detection system for soybean hybrids based on botanical morphological characteristics, used to execute the automated detection method for soybean hybrids based on botanical morphological characteristics according to any one of claims 1-8, characterized in that, include: The control device is used to control the camera to acquire top view and side view images of the plant at the top and side of the plant, respectively, to obtain ectopic multi-trait image pairs of the same plant. The top view image is used to capture the structure of soybean compound leaves, and the side view image is used to capture the distribution of flowers. The processor receives data collected by the camera and performs feature extraction, preliminary grouping, structural reconstruction, and hybrid identification.

Images