Plant structure pest and disease three-dimensional detection system based on three-dimensional point cloud reconstruction

Abstract

The invention, which belongs to the technical field of image processing and plant disease and insect pest detection, discloses a three-dimensional point cloud reconstruction-based plant structure disease and insect pest three-dimensional detection system comprising a neural radiation field three-dimensional reconstruction module, a plant organ semantic segmentation module, a hidden disease three-dimensional detection module and a disease three-dimensional quantitative evaluation module. A three-dimensional Gaussian sputtering technology is adopted to generate a high-density point cloud model, organ-level semantic segmentation is realized, geometric features and texture features are combined to carry out multi-dimensional disease analysis, diseases and insect pests of hidden parts such as leaf backs, stem base inner sides and fruit interiors are effectively detected, the total detection rate of the diseases and insect pests is improved by 38.5%, and the measurement precision is within 5%. And a scientific basis is provided for health assessment and precise prevention and control of garden plants.

Description

Technical Field

[0001] This invention relates to the fields of image processing and plant disease and pest detection technology, specifically to a three-dimensional point cloud reconstruction system for plant structural disease and pest detection. This system employs neural radiation field technology for three-dimensional plant reconstruction, combined with point cloud semantic segmentation and multi-dimensional feature analysis, to achieve comprehensive three-dimensional detection of diseases and pests in concealed plant areas. It is particularly suitable for refined health monitoring and control decision-making for garden plants. Background Technology

[0002] Plant diseases and pests are a significant factor affecting the quality of landscaping projects. Timely and accurate detection of diseases and pests is crucial for ensuring the aesthetic appeal of the landscape and improving plant survival rates. Traditional methods of disease and pest detection rely primarily on manual inspections and experience-based judgment, which are not only inefficient and costly but also fail to detect early-stage diseases and pests in hidden parts of plants, leading to missed opportunities for optimal prevention and control.

[0003] In recent years, with the rapid development of computer vision and deep learning technologies, image-based automatic pest and disease detection methods have been widely researched and applied. Two-dimensional image analysis methods automatically identify pests and diseases by taking photos of plant leaves, fruits, and other parts, and using image segmentation, feature extraction, and classification techniques. However, two-dimensional image methods have obvious limitations: First, they are limited by the shooting angle, only observing the visible parts of the plant surface, making it difficult to detect pests and diseases in hidden parts such as the back of leaves, the inner side of the stem base, and the inside of fruits; second, they lack depth information, making it impossible to accurately assess the depth of pest and disease infestation and the degree of infection; third, two-dimensional images are easily affected by changes in lighting, occlusion, and background interference, resulting in unstable detection accuracy.

[0004] To overcome the shortcomings of two-dimensional image methods, three-dimensional reconstruction technology has been gradually introduced into the field of plant disease and pest detection. Existing technologies, such as Chinese patent application CN119417787A, disclose a three-dimensional measurement method for multiple diseases on structural surfaces. This method uses a drone to collect image sequences carrying GPS coordinate information, employs three-dimensional reconstruction technology to perform depth calculations on the image sequences and generate an initial three-dimensional model, uses a semantic segmentation network to identify semantic maps of each disease in the image sequences, integrates these semantic maps into a three-dimensional point cloud model, and finally uses a point cloud clustering algorithm to extract disease instances, performing three-dimensional measurements on three typical diseases: cracks, corrosion, and peeling. While this method achieves 3D localization and measurement of plant diseases, it still has the following shortcomings: First, this method mainly targets cracks, corrosion, and peeling diseases on building structures. Its technical solutions and algorithm design are not suitable for the specific needs of plant disease and pest detection. Plants have complex 3D structures, flexible deformations, and multi-level organs, which are fundamentally different from rigid building structures. Second, this method uses traditional SfM / MVS 3D reconstruction technology, which suffers from insufficient reconstruction accuracy and completeness when dealing with complex scenes such as small plant branches and leaves, transparent fruits, and highly occluded areas, making it difficult to generate high-density plant point cloud models. Third, this method lacks semantic segmentation capabilities at the plant organ level and cannot distinguish between leaves. The method is limited by several factors: First, different organs such as leaves, stems, and fruits make it impossible to differentiate the detection of pests and diseases based on the characteristics of different organs. Second, the method relies solely on the geometric information of point clouds for disease identification, ignoring the important role of texture features in disease and pest diagnosis. In particular, for early-stage diseases and diseases in hidden areas, geometric changes are often not obvious, and relying solely on geometric features can easily lead to missed detections. Third, the method lacks a dedicated detection mechanism for pests and diseases in hidden areas, making it unable to effectively identify pests and diseases in areas that are difficult to observe from conventional perspectives, such as the back of leaves, the inner side of the stem base, and the inside of fruits. These hidden areas are often the initial locations and high-incidence areas of pests and diseases, and missing early detection will lead to the rapid spread of pests and diseases.

[0005] In addition, plant disease and pest detection also faces some unique challenges: the diversity and complexity of plant morphology require three-dimensional reconstruction methods to be highly adaptable; plants undergo flexible deformation during growth, which increases the difficulty of tracking diseases across time points; changes in lighting, wind swaying, and cluttered backgrounds in the garden environment place higher demands on image acquisition and feature extraction; different types of diseases and pests exhibit differentiated characteristics in different organs and at different developmental stages, requiring the establishment of multi-level and multi-dimensional detection mechanisms.

[0006] Therefore, there is an urgent need for a three-dimensional pest and disease detection system designed specifically for the structural characteristics of plants. This system can generate high-precision plant point cloud models, achieve organ-level semantic segmentation, and perform multi-dimensional analysis by combining geometric and texture features. In particular, it can effectively detect pests and diseases in hidden parts, providing reliable technical support for plant health assessment and precise prevention and control. Summary of the Invention

[0007] The purpose of this invention is to provide a three-dimensional point cloud reconstruction system for detecting plant diseases and pests, in order to solve the technical problems of low detection accuracy, inability to effectively detect diseases and pests in hidden parts, and lack of organ-level analysis capabilities in the existing technology.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is: a three-dimensional point cloud reconstruction system for plant structure pest and disease detection, including a three-dimensional reconstruction module for neural radiation fields, a semantic segmentation module for plant organs, a three-dimensional detection module for hidden diseases, and a three-dimensional quantitative assessment module for diseases.

[0009] The neural radiation field 3D reconstruction module is used to generate high-density point cloud models of plants based on multi-view RGB image sequences using 3D Gaussian sputtering technology, achieving high-fidelity reconstruction of complex 3D plant structures. This module adaptively adjusts the density and size parameters of the Gaussian ellipsoid, providing refined reconstruction capabilities for small branches and leaves and highly occluded areas, generating high-density point cloud models with a spatial resolution of no less than 0.5 mm. This provides a high-quality 3D data foundation for subsequent organ segmentation and disease detection.

[0010] The plant organ semantic segmentation module is connected to the neural radiation field 3D reconstruction module to perform organ-level semantic segmentation on high-density point cloud models, identifying and labeling leaf point clouds, stem point clouds, and fruit point clouds. This module employs a combination of multi-scale feature extraction and organ classification prediction to accurately distinguish different organ types. Boundary optimization techniques further enhance segmentation accuracy, generating organ-segmented point cloud models that provide organ-level data support for targeted disease detection.

[0011] The hidden disease 3D detection module is connected to the plant organ semantic segmentation module. Based on the organ segmentation point cloud model, it extracts the geometric and textural features of each organ's point cloud, identifying diseases and pests in hidden areas that are difficult to observe from a conventional perspective. This module targets hidden areas such as the underside of leaves, the inner side of the stem base, and the interior of fruits. It comprehensively analyzes geometric anomalies and textural changes to establish a multi-dimensional disease identification mechanism, generating a 3D disease distribution map and effectively improving the detection rate of hidden diseases.

[0012] The three-dimensional quantitative assessment module for plant diseases is connected to the three-dimensional detection module for hidden diseases. Based on a three-dimensional distribution map of diseases, it calculates the three-dimensional spatial distribution characteristics and infection severity scores of pests and diseases. This module measures the three-dimensional dimensional parameters of each diseased area, analyzes the distribution patterns of pests and diseases in the three-dimensional space of the plant, comprehensively assesses the overall health status of the plant, and generates a comprehensive disease assessment report, providing quantitative basis for prevention and control decisions.

[0013] The beneficial effects of this invention are as follows:

[0014] First, this invention uses neural radiation field technology and three-dimensional Gaussian sputtering method for three-dimensional reconstruction of plants. Compared with the traditional SfM / MVS method, it can generate plant point cloud models with higher density and higher precision, with a spatial resolution of less than 0.5 mm. It effectively solves the problem of reconstructing small branches and leaves and highly occluded areas of plants, and provides a high-quality three-dimensional data foundation for accurate disease detection.

[0015] Second, this invention achieves semantic segmentation at the plant organ level, which can accurately distinguish different organs such as leaves, stems and fruits, making it possible to adopt differentiated disease detection strategies for different organ types, and significantly improving the accuracy and targeting of disease and pest identification.

[0016] Third, this invention innovatively designs a three-dimensional detection mechanism for hidden diseases, which comprehensively utilizes geometric and textural features for multi-dimensional analysis. It can effectively detect diseases and pests in areas that are difficult to observe from conventional perspectives, such as the back of leaves, the inner side of the stem base, and the inside of fruits. Compared with two-dimensional methods that can only detect visible surface diseases, the overall detection rate of diseases and pests is increased by 38.5%, especially in the detection of early hidden diseases, which has a significant advantage and provides key information for timely prevention and control.

[0017] Fourth, this invention establishes a deep coupling and closed-loop feedback mechanism between the modules. The evaluation results of the disease three-dimensional quantitative assessment module can provide feedback to guide the neural radiation field three-dimensional reconstruction module to improve the reconstruction accuracy of disease-prone areas. The plant organ semantic segmentation module and the hidden disease three-dimensional detection module work together to promote each other, thereby achieving overall optimization of detection performance.

[0018] Fifth, the comprehensive disease assessment report generated by this invention not only includes the type and location information of pests and diseases, but also provides quantitative indicators such as three-dimensional size parameters, spatial distribution patterns and infection degree scores, providing a scientific basis for the comprehensive assessment of plant health status and the formulation of precise prevention and control strategies, and is particularly suitable for the refined management of garden plants. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall architecture of the three-dimensional point cloud reconstruction system for three-dimensional detection of plant structure diseases and pests of the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of the three-dimensional reconstruction module of the neural radiation field of the present invention;

[0021] Figure 3 This is a schematic diagram of the structure of the plant organ semantic segmentation module of the present invention;

[0022] Figure 4 This is a schematic diagram of the structure of the concealed disease three-dimensional detection module of the present invention;

[0023] Figure 5This is a schematic diagram of the structure of the three-dimensional quantitative assessment module for diseases of the present invention. Detailed Implementation

[0024] Please refer to the attached document. Figures 1-5 The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0025] like Figure 1 As shown, the three-dimensional point cloud reconstruction system for plant structural pest and disease detection provided by this invention includes a neural radiation field three-dimensional reconstruction module 1, a plant organ semantic segmentation module 2, a hidden disease three-dimensional detection module 3, and a disease three-dimensional quantitative assessment module 4. These modules form a close data flow and collaborative working relationship, jointly achieving comprehensive three-dimensional detection of plant pests and diseases.

[0026] like Figure 2 As shown, the neural radiation field 3D reconstruction module 1 includes a multi-view image acquisition unit, a camera pose estimation unit, and an adaptive Gaussian sputtering unit. The core function of this module is to generate a high-density, high-precision plant point cloud model based on multi-view RGB image sequences.

[0027] The multi-view image acquisition unit is used to acquire multi-view RGB image sequences of plants. In a preferred embodiment, a surround shooting method is adopted, with 36-72 shooting positions evenly set within a 360° range around the plant. Each position acquires one RGB image, with an image resolution of no less than 4096×3072 pixels. During shooting, the distance between the camera and the center of the plant is maintained between 0.5m and 2.0m, and the angle between two adjacent shooting positions is 5°-10°, ensuring a 50%-70% overlap between adjacent images. In practical applications, the number of shooting positions and the shooting distance can be adjusted according to the size and complexity of the plant. For small garden plants such as roses and azaleas, the shooting distance can be set to 0.5m-1.0m, and the number of shooting positions can be 36-48; for medium to large garden plants such as crabapples, magnolias, and ginkgoes, the shooting distance can be set to 1.0m-2.0m, and the number of shooting positions can be 48-72. In addition, to better capture structural information of the top and bottom of the plant, in addition to horizontal panoramic shooting, you can also add shooting from top and bottom angles. The top angle is usually set to 15°-30°, and the bottom angle is set to -15° to -30°.

[0028] The camera pose estimation unit is connected to the multi-view image acquisition unit to determine the camera extrinsic and intrinsic parameters corresponding to each view image. The camera extrinsic parameters include the camera's rotation matrix and translation vector, describing the camera's position and pose in the world coordinate system; the camera intrinsic parameters include focal length, principal point coordinates, and radial distortion coefficients, describing the camera's internal geometric characteristics. In a preferred embodiment, an incremental SfM method is used for camera pose estimation. First, the SIFT algorithm is used to extract feature points from each image, typically 2000-5000 feature points; then, a feature matching algorithm is used to calculate the correspondence between feature points in adjacent images, with a matching threshold set to 0.7-0.8; next, the RANSAC algorithm and the five-point method are used to estimate the essential matrix between adjacent image pairs, with the outlier elimination ratio typically controlled within 10%-20%; finally, triangulation and bundle adjustment optimization are used to simultaneously estimate the extrinsic and intrinsic parameters of all cameras, as well as the spatial coordinates of sparse 3D points. The bundle adjustment optimization uses the Levenberg-Marquardt algorithm, with 50-100 iterations and a convergence threshold set to... After optimization, the estimation error of camera pose is typically less than 0.5°, and the estimation error of camera position is less than 5mm.

[0029] An adaptive Gaussian sputtering unit, connected to a camera pose estimation unit, adaptively adjusts the density and size parameters of the Gaussian ellipsoids based on the local complexity of the plant structure, generating a high-density point cloud model. This unit employs an improved 3D Gaussian Splatting technique, representing the scene as a collection of three-dimensional Gaussian ellipsoids, each with position, covariance matrix, opacity, and color attributes. Compared to the traditional NeRF method, 3D Gaussian Splatting utilizes explicit representation and differentiable rendering techniques, resulting in a 10-50x improvement in training speed, a 100-1000x improvement in rendering speed, and sub-millimeter level reconstruction accuracy.

[0030] The core innovation of the adaptive Gaussian sputtering unit in this invention lies in the proposed adaptive density control algorithm. This algorithm dynamically adjusts the density and size of the Gaussian ellipsoid based on the local complexity of the plant structure. For complex regions such as small branches and flowers, the Gaussian ellipsoid density is increased and the ellipsoid size is decreased. For relatively simple regions such as stems and fruits, the Gaussian ellipsoid density is decreased and the ellipsoid size is increased, thus optimizing computational efficiency while ensuring reconstruction accuracy.

[0031] Specifically, the adaptive density control algorithm is implemented as follows:

[0032] First, define the local complexity metric. , used to measure the The structural complexity of each spatial region:

[0033] ,

[0034] in, For the first Geometric complexity of a spatial region For texture complexity, For depth change rate, , , The weighting coefficients of the three are respectively, satisfying In a preferred embodiment, , , .

[0035] Geometric complexity It is measured by calculating the rate of change of the normal direction of the point cloud within the region:

[0036] ,

[0037] in, For the first The number of point clouds in a spatial region. For the first Normal vectors at points This is the average of the normal vectors of all points within the region. This represents the L2 norm. Geometric complexity reflects the consistency of surface normals within a region; regions with drastic changes in normals (such as where branches intersect) have higher geometric complexity.

[0038] Texture complexity It is measured by calculating the gradient magnitude of the image within the region:

[0039] ,

[0040] in, For the first The number of pixels projected onto the image plane from a spatial region. and The first The gradient values ​​of each pixel in the horizontal and vertical directions. Texture complexity reflects the degree of variation in color and texture within a region; regions with rich textures (such as lesions or insect damage) have higher texture complexity.

[0041] Depth change rate Measured by calculating the standard deviation of the point cloud depth within the region:

[0042] ,

[0043] in, For the first The depth value of each point, This is the average depth value of all points within the region. The depth variation rate reflects the degree of surface undulation within the region; areas with uneven surfaces (such as wrinkled blades) have a higher depth variation rate.

[0044] Based on the local complexity index, adaptively adjust the... Gaussian ellipsoid density of a spatial region and average size :

[0045] ,

[0046] ,

[0047] in, and These are the minimum and maximum values ​​of the Gaussian ellipsoid density, respectively. and These are the minimum and maximum values ​​of the average size of the Gaussian ellipsoid, respectively. This represents the maximum local complexity across all spatial regions. In a preferred embodiment, Units / m³ Units / m³ mm, mm. This adaptive adjustment strategy ensures that structurally complex regions have a higher Gaussian ellipsoid density and a smaller ellipsoid size, thus achieving higher reconstruction accuracy; while structurally simple regions have a lower Gaussian ellipsoid density and a larger ellipsoid size, reducing unnecessary computational overhead.

[0048] Through the aforementioned adaptive density control algorithm, this invention can generate high-density point cloud models with a spatial resolution of no less than 0.5 mm. In actual tests, for the reconstruction of rose plants, the traditional SfM / MVS method generates point cloud densities of approximately 50-100 points / cm², with a reconstruction error of 2-3 mm; while the adaptive Gaussian sputtering method used in this invention generates point cloud densities of 500-1000 points / cm², reducing the reconstruction error to 0.5-0.8 mm, significantly improving the reconstruction quality.

[0049] like Figure 3 As shown, the plant organ semantic segmentation module 2 includes a multi-scale feature extraction unit, an organ classification prediction unit, and a boundary optimization unit. The core function of this module is to perform organ-level semantic segmentation on a high-density point cloud model, identifying and labeling leaf point clouds, stem point clouds, and fruit point clouds.

[0050] A multi-scale feature extraction unit is used to extract multi-scale geometric features from a high-density point cloud model. The geometric features of a point cloud include local surface shape, spatial distribution patterns, and neighborhood relationships, which play a crucial role in distinguishing different organ types. In this invention, an improved PointNet++ network architecture is employed for multi-scale feature extraction. PointNet++ extracts local features of the point cloud at different spatial scales through multi-level sampling and grouping operations, and then generates a multi-scale feature vector for each point through feature propagation and fusion.

[0051] Specifically, the multi-scale feature extraction process includes the following steps:

[0052] The first step is downsampling and grouping. The furthest-point sampling algorithm is used to select points from the high-density point cloud. Each representative point is used as the first layer of sampling points. Then, with each sampling point as the center, a sphere query method is used to search within a radius... Search for neighboring points within the range, with the number of each neighboring point limited to a certain limit. The searched neighborhood points are normalized, and their coordinates are transformed into a local coordinate system with the sampling point as the origin.

[0053] The second step is local feature learning. For the neighborhood point set of each sampling point, a multilayer perceptron is used to extract local features. The multilayer perceptron contains three fully connected layers with 64, 128, and 256 hidden neurons, respectively, and ReLU is used as the activation function. Through max pooling, the local features of the neighborhood points are aggregated into a feature vector of the sampling point, with a feature vector dimension of 256.

[0054] The third step is iterative sampling and feature extraction. Based on the first layer of sampling points, further downsampling and grouping are performed to generate the second layer of sampling points. One, sphere query radius Number of neighboring points Repeat the local feature learning process to extract feature vectors from the second layer of sampling points. The feature vector dimension is 512. Third and fourth layers of sampling and feature extraction can be performed as needed to form a multi-level feature pyramid.

[0055] The fourth step is feature propagation and fusion. Interpolation and skip connections are used to propagate feature vectors from different levels back to each point in the original point cloud. For the first step... The sampling point features of layer 1 are propagated to layer 2 using k-nearest neighbor interpolation. The sampling points of the layer are weighted based on spatial distance. Simultaneously, the sampling points of the [layer name missing] are [number missing], and the interpolation weights are calculated based on spatial distance. The original features of the layer sampling points are concatenated with the features obtained through propagation to form fused features. Through layer-by-layer propagation and fusion, a multi-scale fused feature vector is finally generated for each point in the original point cloud, with a feature vector dimension of 1024.

[0056] The multi-scale feature extraction process described above can effectively capture the geometric characteristics of point clouds at different spatial scales, providing rich feature representations for subsequent organ classification.

[0057] The organ classification prediction unit is connected to the multi-scale feature extraction unit to predict the organ category label of each point cloud point based on multi-scale geometric features. In this invention, organ categories include four types: leaves, stems, fruits, and background. Organ classification prediction is implemented using a multilayer perceptron. The input is a multi-scale fused feature vector (dimension 1024) for each point, and the output is the probability distribution (dimension 4) of the point belonging to each organ category. The multilayer perceptron contains three fully connected layers with 512, 256, and 128 hidden layer neurons, respectively. The ReLU activation function is used, and the last layer uses the Softmax activation function to convert the output into a probability distribution.

[0058] During the training phase, the cross-entropy loss function was used to optimize the network parameters. The training data consisted of 1000-2000 plant point cloud samples with organ annotations, covering various garden plants such as roses, crabapples, magnolias, and ginkgo. The optimizer was Adam, with an initial learning rate of 0.001, which decayed to 0.5 times the original rate every 30 epochs, for a total of 150 epochs. To improve the model's generalization ability, data augmentation techniques were employed during training, including random rotation, random translation, random scaling, and random point cloud discarding.

[0059] During the inference phase, organ category prediction is performed for each point in the high-density point cloud model, and the category with the highest probability is selected as the predicted label for that point. To improve the stability of the prediction, a Conditional Random Field (CRF) is used for post-processing optimization. The CRF considers spatial consistency constraints between adjacent points, ensuring that points within the same organ have the same category label and that the boundaries between different organs are clearer. The energy function of the CRF includes unary and binary terms. The unary term is based on the output probability of the organ classification prediction unit, and the binary term is based on the spatial distance and feature similarity between adjacent points. By minimizing the energy function, the organ category label for each point is optimized.

[0060] The boundary optimization unit is connected to the organ classification prediction unit to optimize the segmentation boundaries between point clouds of different organs, generating an organ segmentation point cloud model. Since the boundaries between plant organs often contain occlusion, contact, and transition regions, simple classification prediction can easily lead to incorrect labels or blurred boundaries. The boundary optimization unit employs a combination of region growing and boundary refinement to improve the accuracy of boundary segmentation.

[0061] Specifically, the boundary optimization process includes the following steps:

[0062] First, identify boundary regions. For each point in the point cloud, if there are points of different categories among its k nearest neighbors (k=10), then mark that point as a boundary point. Count all boundary points to form a boundary region point set.

[0063] Second, boundary regions are reclassified. For each point in the boundary region, the confidence score for its classification into each organ category is recalculated. The confidence score considers not only the point's own classification probability but also the category distribution of its neighboring points. A weighted voting method is used, with the voting weights of neighboring points calculated based on spatial distance and feature similarity. The reclassified category labels are more accurate and stable.

[0064] Third, region growing and smoothing. For points in non-boundary regions, a region growing algorithm is used for label propagation. Starting from the seed point, adjacent unlabeled points or boundary points are included in the same organ region until the boundary of different organs or the growing stops. During the region growing process, feature similarity thresholds and spatial distance thresholds are set to ensure the rationality of the growing. Finally, median filtering is used to smooth the organ regions and eliminate isolated erroneous labeled points.

[0065] Using the boundary optimization method described above, the organ segmentation point cloud model generated by this invention has clear organ boundaries and accurate organ labels. In actual tests, for the segmentation of rose plants, the segmentation accuracy rates for leaves, stems, and fruits reached 92.3%, 89.7%, and 94.1%, respectively, with an average crossover-union ratio of 0.87, which is significantly better than the results without boundary optimization (average crossover-union ratio of 0.79).

[0066] like Figure 4 As shown, the hidden disease 3D detection module includes a geometric anomaly detection unit, a texture feature analysis unit, and a hidden disease identification unit. The core function of this module is to extract the geometric and texture features of the point cloud of each organ based on the organ segmentation point cloud model, and identify diseases and pests in hidden areas that are difficult to observe from a conventional perspective.

[0067] The geometric anomaly detection unit is used to detect geometrically abnormal regions in organ segmentation point cloud models. Plant diseases and pests often cause geometrical changes on the surface of plant organs, such as leaf shrinkage, fruit depression, and stem swelling. These geometrical anomalies are important indicators of diseases and pests, especially for diseases and pests in hidden areas, where geometrical anomalies are often the most direct detection clues.

[0068] The implementation of geometric anomaly detection includes the following steps:

[0069] First, calculate the point cloud normals and curvature. For each point in the organ segmentation point cloud model, principal component analysis is used to calculate its local normal direction. Specifically, with that point as the center, within a radius... The system searches for neighboring points within a range of mm and constructs a covariance matrix for these points. The eigenvector corresponding to the smallest eigenvalue of the covariance matrix is ​​the normal direction of that point. Simultaneously, the curvature value of that point is calculated. The curvature value is defined as the ratio of the smallest eigenvalue to the sum of all eigenvalues, reflecting the degree of curvature of the local surface.

[0070] Second, identify regions with abnormal curvature. Statistically analyze the curvature values ​​of all points in the organ point cloud and calculate the mean of these curvature values. and standard deviation Define the curvature anomaly threshold. Points with curvature values ​​exceeding this threshold are marked as candidate anomalies. Candidate anomalies are often concentrated in diseased areas, wrinkled areas, and edge areas.

[0071] Third, neighborhood consistency test. To distinguish geometric anomalies caused by disease from normal folds and edges, a neighborhood consistency test is performed on candidate anomaly points. Specifically, for each candidate anomaly point, its radius is calculated. The density of candidate anomalies within a mm neighborhood. If the density of candidate anomalies within the neighborhood exceeds a threshold of 30%, the candidate anomaly is considered to belong to a geometrically abnormal region; otherwise, it is considered an isolated curvature abrupt change point and not a disease anomaly. Through neighborhood consistency testing, normal wrinkles and edges can be effectively filtered out, focusing on geometrically abnormal regions caused by diseases.

[0072] In this invention, geometric anomalies include types such as local depressions, surface wrinkles, and thickness variations. Local depressions are typically characterized by a significantly increased curvature value and a normal direction pointing inwards, commonly seen in fruit and stem diseases; surface wrinkles are characterized by rapid changes in curvature values ​​within a local area and disordered normal directions, commonly seen in leaf diseases; thickness variations are characterized by a significant depth difference between the organ surface and the normal area, commonly seen in insect damage and disease erosion. By comprehensively analyzing curvature, normal direction, and depth information, different types of geometric anomalies can be accurately identified.

[0073] The texture feature analysis unit is connected to the geometric anomaly detection unit to analyze the texture features of areas with geometric anomalies. Although geometric anomalies are important clues to pests and diseases, relying solely on geometric information can easily lead to misjudgments, especially for early-stage diseases and certain disease types with obvious texture changes but inconspicuous geometric changes. Texture features can provide supplementary information such as color, grayscale, and texture roughness, forming a multi-dimensional analysis with geometric features, significantly improving the accuracy of pest and disease identification.

[0074] The core innovation of texture feature analysis lies in the proposed multi-view texture fusion algorithm. This algorithm projects the 3D point cloud of geometrically abnormal regions onto multiple 2D view planes, extracts the texture features of the images from each view plane, and then fuses the multi-view texture features to generate a comprehensive texture feature descriptor.

[0075] Specifically, the implementation of the multi-view texture fusion algorithm is as follows:

[0076] First, determine the set of projection viewpoints. For each geometrically abnormal region, select six projection viewpoints: front view, back view, left view, right view, top view, and bottom view. The selection of projection viewpoints ensures that the abnormal region can be observed from multiple angles, capturing texture information from different directions.

[0077] Then, for each projection viewpoint, the 3D point cloud of the abnormal region is projected onto the corresponding 2D plane. The projection process uses an orthogonal projection model, and the projected 2D coordinates are calculated as follows:

[0078] ,

[0079] in, For the world coordinates of a 3D point, The center coordinates of the projection plane, These are the projected two-dimensional coordinates. and Let be two orthogonal basis vectors of the projection plane. For each 2D plane, generate a projection image with a resolution of 256×256 pixels, where the color value of each pixel is determined by the RGB color of the 3D point projected onto that pixel location.

[0080] Next, for each projected image, a texture feature vector is extracted. Texture features include color histograms, gray-level co-occurrence matrices (GLCMs), and local binary patterns. The color histogram statistically analyzes the distribution of each color channel in the image, dividing it into red, green, and blue channels, each further subdivided into 32 intervals, generating a 96-dimensional color histogram feature. The GLCM statistically analyzes the co-occurrence frequency of gray values ​​between adjacent pixels in the image, extracting four statistical measures as texture features: contrast, homogeneity, energy, and correlation. The local binary pattern compares the gray values ​​of the center pixel with those of its neighbors, generating a binary code. The histogram of this code is used as a texture roughness feature, with a dimension of 59. These three types of texture features are concatenated to form a texture feature vector for each projection viewpoint, with a dimension of 159.

[0081] Finally, the multi-view texture feature vectors are fused to generate a comprehensive texture feature descriptor. The fusion is performed using a weighted average method.

[0082] ,

[0083] in, For the first The fusion weights are assigned to each viewpoint, and these weights are determined based on the quality of the projected image at that viewpoint. Quality evaluation metrics include projected point cloud density and image sharpness. Viewpoints with higher projected point cloud density and higher image sharpness are assigned greater weights. The fusion weights satisfy normalization constraints. By using multi-view texture fusion, the comprehensive texture feature descriptor can fully reflect the texture characteristics of abnormal areas, overcoming the limitations of single-view observation.

[0084] The hidden disease identification unit is connected to the texture feature analysis unit to identify the types of pests and diseases in hidden areas based on geometric anomalies and texture features, and generate a three-dimensional distribution map of the diseases. The hidden disease identification is implemented using a deep learning classifier. The input is geometric features (curvature, normal direction, depth) and texture features (comprehensive texture feature descriptor). The output is the type of pest or disease (healthy, fungal disease, bacterial disease, insect pest, physiological disease) and a confidence score.

[0085] Specifically, the process for identifying hidden diseases employs a multi-layer convolutional neural network and a fully connected layer architecture. First, geometric and texture features are input into two independent feature encoding branches. The geometric feature encoding branch uses a 3-layer fully connected network with 128, 256, and 512 neurons in the hidden layers, using ReLU activation, and outputting a geometric feature encoding vector (dimension 512). The texture feature encoding branch also uses a 3-layer fully connected network with 256, 512, and 1024 neurons in the hidden layers, using ReLU activation, and outputting a texture feature encoding vector (dimension 1024). Then, the geometric and texture feature encoding vectors are concatenated to form a joint feature vector (dimension 1536). Finally, the joint feature vector is used to classify disease types through a 2-layer fully connected network with 512 neurons in the hidden layers and 5 neurons in the output layer (corresponding to 5 disease types), using Softmax activation, and outputting the probability distribution for each disease type.

[0086] To improve the accuracy of identifying hidden diseases, this invention proposes a confidence assessment algorithm. This algorithm outputs not only the disease type but also a confidence score for the identification, used to measure the reliability of the identification results. The confidence score comprehensively considers three factors: classification probability, feature quality, and consistency test.

[0087] ,

[0088] in, Score the confidence level. The classification probability score is defined as the difference between the probability of the largest class and the probability of the second largest class, reflecting the degree of certainty of the classification result. The feature quality score is defined as the signal-to-noise ratio of the input feature vector, reflecting the reliability of feature extraction. The consistency test score is defined as the consistency ratio of disease types within the neighborhood, reflecting spatial continuity. , , These are the weighting coefficients for the three components. In a preferred embodiment, , , The confidence score ranges from 0 to 1, with a higher score indicating a more reliable recognition result.

[0089] Using the aforementioned method for identifying hidden diseases, this invention can accurately identify pests and diseases in hidden areas such as the underside of leaves, the inner side of the stem base, and the inside of fruits. In actual tests, for detecting hidden diseases in rose plants, the accuracy rate for fungal diseases was 89.5%, bacterial diseases 86.3%, insect pests 91.2%, and physiological diseases 84.7%, with an average accuracy rate of 87.9%. Compared to methods using only geometric features or only texture features (accuracies of 78.2% and 81.4%, respectively), multi-dimensional feature fusion significantly improves the identification performance.

[0090] Finally, the hidden disease identification unit generates a three-dimensional distribution map of diseases. This distribution map is presented in the form of a three-dimensional point cloud. Each disease point cloud point is labeled with the disease type and confidence score, which intuitively shows the distribution of diseases and pests in the three-dimensional space of the plant.

[0091] like Figure 5 As shown, the three-dimensional quantitative assessment module 4 for diseases includes a disease area measurement unit, a spatial distribution analysis unit, and an infection degree scoring unit. The core function of this module is to calculate the three-dimensional spatial distribution characteristics and infection degree scores of pests and diseases based on the three-dimensional distribution map of diseases, and generate a comprehensive disease assessment report.

[0092] The disease area measurement unit is used to measure the three-dimensional dimensional parameters of each disease area in the three-dimensional distribution map of the disease. The three-dimensional dimensional parameters include lesion area, erosion depth, and infection volume. These parameters can quantify the severity of pests and diseases and provide a scientific basis for prevention and control decisions.

[0093] The area of ​​lesions was measured using a point cloud surface area estimation method. For each affected area, a triangular mesh reconstruction was first performed on the point cloud to generate a triangular mesh model of the affected surface. The triangular mesh reconstruction employed a Poisson surface reconstruction algorithm, which recovers a smooth implicit surface from the point cloud and its normal information by solving the Poisson equation, and then extracts isosurfaces to generate the triangular mesh. The reconstructed triangular mesh model accurately represents the surface morphology of the affected area. The lesion area is defined as the sum of the areas of all triangles.

[0094] ,

[0095] in, The number of triangles, For the first The area of ​​a triangle is calculated using the coordinates of its three vertices. In practical applications, lesion area is usually statistically analyzed and reported in cm².

[0096] The erosion depth was measured using a normal surface fitting and depth deviation calculation method. For each affected area, the point cloud of the healthy surface surrounding the affected area was first identified, and a reference plane or surface of the healthy surface was fitted using the RANSAC algorithm. Then, the vertical distance from each point within the affected area to the reference surface was calculated, and the maximum vertical distance was defined as the erosion depth.

[0097] ,

[0098] in, This represents the number of point clouds within the affected area. For the first The vertical distance from each diseased point to the reference surface. Erosion depth reflects the extent to which pests and diseases erode plant tissue; the greater the erosion depth, the more severe the damage. In practical applications, erosion depth is usually statistically analyzed and reported in mm.

[0099] The infected volume was measured using a volume integration method between the point cloud of the affected area and a reference surface. First, a closed 3D mesh model was constructed based on the point cloud of the affected area and the reference surface of the healthy surface. This mesh model encloses the volume of the cavity formed by the erosion. Then, the volume of the closed mesh was calculated using the divergence theorem.

[0100] ,

[0101] in, , , For the first The three vertices of a triangle are represented by their coordinate vectors. Infection volume reflects the amount of tissue loss caused by pests and diseases; a larger infection volume indicates more severe tissue damage. In practical applications, infection volume is usually statistically analyzed and reported in cm³.

[0102] The spatial distribution analysis unit is connected to the disease area measurement unit to analyze the distribution patterns of pests and diseases in the three-dimensional space of plants. The spatial distribution patterns of pests and diseases can reflect the spread characteristics and development trends of diseases, which is of great significance for formulating control strategies. In this invention, the spatial distribution patterns include three types: clustered distribution, striped distribution, and scattered distribution.

[0103] Clustered distribution indicates that pests and diseases are concentrated in certain localized areas of a plant, forming several independent disease clusters. The disease density is high within each cluster, while the intervals between clusters are relatively large. Clustered distribution is common in fungal and bacterial diseases, which often start from a single infection source and gradually spread outwards, forming a clustered distribution pattern. Clustered distribution identification uses the DBSCAN clustering algorithm, which can automatically discover high-density clusters in a point cloud without pre-specifying the number of clusters. Key parameters of the DBSCAN algorithm include the neighborhood radius. mm and minimum number of points In the clustering results, clusters containing more than a threshold (such as 100 points) are identified as diseased clusters, and the number, average size, and spatial distribution of the clusters are statistically analyzed.

[0104] Banded distribution indicates that pests and diseases are distributed linearly or in bands along a certain direction on a plant, such as along leaf veins, stems, or cracks on the surface of fruits. Banded distribution is common in physiological diseases and some insect pests, which often spread along vulnerable parts of the plant or nutrient transport channels, forming a banded distribution pattern. Banded distribution is identified using principal component analysis to calculate the principal and secondary directions of the disease point cloud. If the eigenvalue of the principal direction is significantly greater than the eigenvalues ​​of the secondary and tertiary directions (ratio exceeding 5), the disease is considered to have a banded distribution. The length, width, and directional angle of the bands, as well as the number of bands and the average spacing, are statistically analyzed.

[0105] Scattered distribution indicates that pests and diseases are randomly distributed in the three-dimensional space of the plant, with relatively independent disease points that do not form obvious clusters or bands. Scattered distribution is common in early-stage diseases and mild pest infestations, when the disease is in its initial stage and has not yet spread on a large scale. Scattered distribution is identified using nearest neighbor distance analysis, calculating the distance from each disease point to its nearest neighbor and statistically analyzing the mean and variance of these distances. If the mean distance is large (e.g., exceeding 50 mm) and the variance is small (coefficient of variation less than 0.5), the disease is considered to have a scattered distribution. Quantitative indicators of scattered distribution include the total number of disease points, average spacing, and spatial density.

[0106] The infection severity scoring unit is connected to the spatial distribution analysis unit to calculate the overall infection severity score of the plant based on three-dimensional size parameters and distribution patterns, generating a comprehensive disease assessment report. The infection severity score comprehensively considers multiple factors such as the quantity, size, distribution, and type of the disease, providing a quantitative assessment of the overall health status of the plant.

[0107] The infection severity score was calculated using a multi-factor weighted fusion method:

[0108] ,

[0109] in, This is the ratio of the total area of ​​the diseased region to the plant surface area. This is the normalized value of the maximum erosion depth. This is the normalized value of the total infection volume. This represents the normalized value for the number of diseased areas. The severity coefficients for disease types are: fungal diseases 1.0, bacterial diseases 0.9, insect pests 0.8, and physiological diseases 0.7. , , , , These are the weight coefficients for each factor. In a preferred embodiment, , , , , The infection severity score ranges from 0 to 1, where 0 indicates complete health and 1 indicates severe infection. Based on the score, the plant health status is divided into five levels: excellent (0-0.2), good (0.2-0.4), fair (0.4-0.6), poor (0.6-0.8), and severe (0.8-1.0).

[0110] The comprehensive disease assessment report includes the following: overall plant infection severity score and health level; quantity, area, volume, and distribution patterns of various diseases and pests; spatial location and 3D visualization of high-incidence areas; disease development trend prediction and control recommendations. The report is presented in a combination of charts and graphs, intuitively displaying the detection results and quantitative analysis of diseases and pests, providing decision support for landscape engineering technicians and maintenance personnel.

[0111] The modules of the system of this invention are not simply linearly connected, but form a collaborative working mechanism with deep coupling and closed-loop feedback, which is an important innovation of this invention.

[0112] First, a closed-loop feedback mechanism is established between the neural radiation field 3D reconstruction module 1 and the disease 3D quantitative assessment module 4. During the generation of a comprehensive disease assessment report, the disease 3D quantitative assessment module 4 can identify high-incidence areas and key areas of concern. This area information is transmitted to the neural radiation field 3D reconstruction module 1 in the form of reconstruction accuracy feedback signals. Upon receiving the feedback signals, the neural radiation field 3D reconstruction module 1 responsively increases the Gaussian ellipsoid density in high-incidence areas from 0.5 mm to 0.3 mm or even higher, ensuring more refined point cloud reconstruction in these areas and providing more accurate 3D data for disease detection. This closed-loop feedback mechanism achieves adaptive optimization of reconstruction accuracy, concentrating limited computing resources on the areas most requiring refined reconstruction, improving the detection performance of key areas while ensuring overall efficiency.

[0113] Second, a bidirectional collaborative mechanism is established between the plant organ semantic segmentation module 2 and the hidden disease stereoscopic detection module 3. On the one hand, the organ segmentation results output by the plant organ semantic segmentation module 2 provide organ type information to the hidden disease stereoscopic detection module 3. The hidden disease stereoscopic detection module 3 adopts differentiated disease detection strategies according to different organ types. For example, for leaf organs, it focuses on detecting fungal and bacterial diseases on the underside of the leaves; for stem organs, it focuses on detecting insect pests and physiological diseases on the inner side of the stem base; for fruit organs, it focuses on detecting internal rot and insect infestation. The differentiated detection strategy can be optimized for the disease characteristics of different organs, significantly improving the targeting and accuracy of detection. On the other hand, the disease features detected by the hidden disease stereoscopic detection module 3 are fed back to the plant organ semantic segmentation module 2 to optimize the boundary segmentation accuracy between organs and diseased areas. Diseased areas are often the difficult part of organ segmentation because diseases can cause changes in the appearance and morphology of organs, which can easily lead to segmentation errors. By incorporating disease characteristic information, the plant organ semantic segmentation module 2 can more accurately identify organ deformation caused by diseases, optimize boundary segmentation, and avoid misclassifying diseased areas into neighboring organs. This bidirectional collaborative mechanism achieves mutual promotion between organ segmentation and disease detection, resulting in a synergistic effect where 1+1>2.

[0114] Third, parameter-level coupling between modules ensures the efficiency and consistency of data flow. The high-density point cloud model output by the neural radiation field 3D reconstruction module 1 is directly used as input to the plant organ semantic segmentation module 2. The point cloud data contains complete information such as 3D coordinates, RGB colors, and normal directions, requiring no additional data conversion or preprocessing. The organ segmentation point cloud model output by the plant organ semantic segmentation module 2 adds organ category labels to the original point cloud data, maintaining data structure consistency, and is directly passed to the hidden disease 3D detection module 3. The disease 3D distribution map output by the hidden disease 3D detection module 3 further adds disease type and confidence score labels, and continues to be passed to the disease 3D quantitative assessment module 4 for quantitative analysis. This parameter-level coupling avoids data format conversion and information loss between modules, improving the overall efficiency and accuracy of the system.

[0115] Through the aforementioned deep coupling and closed-loop feedback mechanism, the various modules of the system of this invention form an organic and unified whole, realizing the full-process collaborative optimization from three-dimensional reconstruction, organ segmentation, disease detection to quantitative assessment, and significantly improving the performance and practicality of three-dimensional detection of plant diseases and pests.

[0116] To verify the practical effectiveness of the system of this invention, an application test was conducted in a horticultural nursery. The test subjects were 50 rose bushes in their growth period, including healthy plants and those infected with different types of pests and diseases. The system of this invention was used to perform three-dimensional reconstruction and pest and disease detection on each rose bush, and the results were compared with those obtained through manual inspection and two-dimensional image detection methods.

[0117] Test results show that the point cloud model of rose plants generated by the system of this invention achieves a spatial resolution of 0.6 mm, significantly better than the 2.3 mm of the traditional SfM / MVS method. The accuracy rate of organ semantic segmentation is 91.8%, accurately distinguishing leaves, stems, and fruits. Regarding pest and disease detection, the system of this invention identified a total of 327 pests and diseases, while manual inspection found 352, resulting in a detection rate of 92.9%. In contrast, the two-dimensional image detection method identified only 215 pests and diseases, with a detection rate of 61.1%. Particularly for pests and diseases in concealed areas, the system of this invention identified 152, while the two-dimensional image detection method only identified 37, improving the detection rate by 310%.

[0118] Specific Case Analysis: A rose bush had early powdery mildew infection on the underside of its leaves, with lesions covering approximately 1.2 cm² and an erosion depth of about 0.8 mm. Because the lesions were located on the underside of the leaves, they were difficult to observe from a conventional perspective and easily missed during manual inspection; two-dimensional image detection methods failed to detect them at all. However, the system of this invention, through three-dimensional reconstruction and hidden disease detection, accurately identified the lesions, measuring an area of ​​1.18 cm² and an erosion depth of 0.76 mm, which highly matched the actual values ​​with an error of less than 5%. The disease assessment report generated by the system showed that the plant's infection severity score was 0.35, belonging to the good level, and recommended preventative pesticide spraying.

[0119] Another case: A rose bush developed root rot on the inner base of its stem, with an erosion depth of 5.2 mm and an infected volume of approximately 2.8 cm³. This disease is insidious and difficult to detect visually, but it severely impacts plant growth. The system of this invention accurately identified the root rot, measuring an erosion depth of 5.1 mm and an infected volume of 2.7 cm³, both with an error of less than 5%. The system assessed the plant's infection severity at a score of 0.72, classifying it as poor. Immediate isolation of the infected plant and soil disinfection are recommended to prevent the disease from spreading to neighboring plants.

[0120] Application tests in garden nurseries have proven that the system of this invention can effectively detect pests and diseases in hidden parts of plants. The overall detection rate of pests and diseases is 38.5% higher than that of two-dimensional methods, and the measurement accuracy reaches within 5%, providing reliable technical support for the health assessment and precise control of garden plants.

[0121] Comprehensive Mechanism Analysis: This invention achieves comprehensive three-dimensional detection of plant diseases and pests through the synergistic effect of four modules: 3D reconstruction of neural radiation fields, semantic segmentation of plant organs, three-dimensional detection of hidden diseases, and three-dimensional quantitative assessment of diseases. The 3D reconstruction module employs 3D Gaussian sputtering technology and dynamically adjusts the reconstruction accuracy based on the local complexity of the plant using an adaptive density control algorithm, solving the problem of insufficient accuracy in traditional methods when dealing with small branches and leaves and highly occluded areas. The semantic segmentation module uses an improved PointNet++ architecture and boundary optimization technology to achieve accurate organ-level segmentation, providing a foundation for differentiated disease detection. The three-dimensional detection module for hidden diseases innovatively proposes a multi-dimensional analysis method combining geometric and texture features, as well as a multi-view texture fusion algorithm, effectively solving the problem of detecting diseases in hidden areas. The three-dimensional quantitative assessment module achieves quantitative assessment of diseases and pests through 3D size measurement and spatial distribution analysis. The closed-loop feedback and bidirectional collaborative mechanism between the modules ensures the optimization of the overall system performance. The synergistic effect of the above-mentioned technical features makes the present invention significantly superior to the prior art in terms of pest and disease detection rate, measurement accuracy, and ability to identify hidden diseases.

[0122] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A three-dimensional point cloud reconstruction system for detecting plant structural diseases and pests, characterized in that, include: A neural radiation field three-dimensional reconstruction module is used to generate a high-density point cloud model of plants based on a multi-view RGB image sequence using three-dimensional Gaussian sputtering technology. The spatial resolution of the high-density point cloud model is not less than 0.5 mm. The plant organ semantic segmentation module is connected to the neural radiation field three-dimensional reconstruction module and is used to perform organ-level semantic segmentation on the high-density point cloud model, identify and label leaf point clouds, stem point clouds and fruit point clouds, and generate organ segmentation point cloud models. The hidden disease three-dimensional detection module is connected to the plant organ semantic segmentation module. It is used to extract the geometric and texture features of each organ point cloud based on the organ segmentation point cloud model, identify hidden diseases and pests that are difficult to observe from a conventional perspective, including the back of the leaf, the inner side of the stem base and the inside of the fruit, and generate a three-dimensional distribution map of the disease. The disease three-dimensional quantitative assessment module is connected to the hidden disease three-dimensional detection module. It is used to calculate the three-dimensional spatial distribution characteristics and infection degree score of the disease based on the disease three-dimensional distribution map, and generate a comprehensive disease assessment report.

2. The three-dimensional point cloud reconstruction system for plant structure pest and disease detection according to claim 1, characterized in that, The three-dimensional reconstruction module for the neural radiation field includes: A multi-view image acquisition unit is used to acquire multi-view RGB image sequences of plants, wherein the image sequence contains no fewer than 36 images from different perspectives; A camera pose estimation unit, connected to the multi-view image acquisition unit, is used to determine the camera extrinsic and intrinsic parameters corresponding to each view image. An adaptive Gaussian sputtering unit, connected to the camera attitude estimation unit, is used to adaptively adjust the density and size parameters of the Gaussian ellipsoid according to the local complexity of the plant structure, thereby generating the high-density point cloud model.

3. The three-dimensional point cloud reconstruction system for plant structure pest and disease detection according to claim 1, characterized in that, The plant organ semantic segmentation module includes: A multi-scale feature extraction unit is used to extract multi-scale geometric features of the high-density point cloud model. The organ classification prediction unit is connected to the multi-scale feature extraction unit and is used to predict the organ category label of each point cloud point based on the multi-scale geometric features. The boundary optimization unit, connected to the organ classification prediction unit, is used to optimize the segmentation boundary between different organ point clouds and generate the organ segmentation point cloud model.

4. The three-dimensional point cloud reconstruction system for plant structural pest and disease detection according to claim 1, characterized in that, The hidden disease three-dimensional detection module includes: A geometric anomaly detection unit is used to detect geometric anomaly regions in the organ segmentation point cloud model, wherein the geometric anomalies include local depressions, surface wrinkles, and thickness variations. A texture feature analysis unit, connected to the geometric anomaly detection unit, is used to analyze the texture features of the geometric anomaly region, the texture features including color distribution, grayscale variation and texture roughness; The hidden disease identification unit is connected to the texture feature analysis unit and is used to identify the types of diseases and pests in hidden areas based on the geometric anomalies and texture features, and generate the three-dimensional distribution map of the diseases.

5. The three-dimensional point cloud reconstruction system for plant structure pest and disease detection according to claim 4, characterized in that, The geometric anomaly detection unit is further used for: Calculate the normal direction and curvature value of each point in the organ segmentation point cloud model; Based on the normal direction and curvature value, point cloud regions with abnormal curvature values ​​are identified as candidate abnormal regions. The candidate abnormal regions are subjected to a neighborhood consistency test to determine the geometrically abnormal regions.

6. The three-dimensional point cloud reconstruction system for plant structure pest and disease detection according to claim 4, characterized in that, The texture feature analysis unit is further used for: The three-dimensional point cloud of the geometrically abnormal region is projected onto multiple two-dimensional view planes; Extract texture feature vectors from the two-dimensional planar images; Multi-view texture feature vectors are fused to generate a comprehensive texture feature descriptor.

7. The three-dimensional point cloud reconstruction system for plant structure pest and disease detection according to claim 1, characterized in that, The three-dimensional quantitative assessment module for diseases includes: The disease area measurement unit is used to measure the three-dimensional dimensional parameters of each disease area in the three-dimensional distribution map of the disease, including the lesion area, erosion depth and infection volume. The spatial distribution analysis unit, connected to the disease area measurement unit, is used to analyze the distribution patterns of pests and diseases in the three-dimensional space of plants. The distribution patterns include clustered distribution, striped distribution, and scattered distribution. The infection severity scoring unit, connected to the spatial distribution analysis unit, is used to calculate the overall infection severity score of the plant based on the three-dimensional size parameters and the distribution pattern, and generate the comprehensive disease assessment report.

8. The three-dimensional point cloud reconstruction system for plant structure pest and disease detection according to claim 1, characterized in that, The neural radiation field three-dimensional reconstruction module is further used for: Receive reconstruction accuracy feedback signal from the three-dimensional quantitative assessment module of the disease; In response to the high-incidence disease area indicated by the reconstruction accuracy feedback signal, the Gaussian ellipsoid density of the corresponding area is increased to enhance the point cloud reconstruction accuracy of the high-incidence disease area.

9. The three-dimensional point cloud reconstruction system for plant structure pest and disease detection according to claim 1, characterized in that, The plant organ semantic segmentation module and the hidden disease three-dimensional detection module work together, specifically as follows: The organ segmentation results output by the plant organ semantic segmentation module guide the hidden disease three-dimensional detection module to adopt differentiated disease detection strategies for different organ types. The disease features detected by the hidden disease three-dimensional detection module are fed back to the plant organ semantic segmentation module to optimize the boundary segmentation accuracy between organs and diseased areas.

10. The three-dimensional point cloud reconstruction system for plant structural pest and disease detection according to claim 1, characterized in that, The system further includes: A multimodal data fusion module, connected to the neural radiation field three-dimensional reconstruction module, is used to fuse RGB images, depth images, and near-infrared images to enhance the detection capability of early-stage hidden diseases. The real-time monitoring and early warning module is connected to the three-dimensional quantitative assessment module for diseases, and is used to generate a disease early warning signal when the infection degree score exceeds a preset threshold.

Images