A method for organ segmentation from point cloud data of maize plants that integrates dynamic weights and feature enhancement

Abstract

This invention discloses a method for organ segmentation of maize plant point clouds that integrates dynamic weights and feature enhancement, comprising the following steps: (1) inputting maize plant point cloud data, including the spatial coordinates and additional features of the point cloud; (2) sampling and aggregating the input point cloud data to obtain sampled point cloud data, and grouping the point cloud by combining multi-scale clustering, i.e., querying spheres with different radii; (3) using the DWE network for feature extraction; the DWE network is composed of a ScoreNet dynamic weight generation module and a feature enhancement module, generating dynamic weights through ScoreNet, and weighting and aggregating the feature-enhanced point cloud according to the dynamic weights; (4) inputting the point cloud and its corresponding features into a fully connected layer and segmenting the semantic information of the point cloud; this invention provides high-quality three-dimensional point cloud data for phenotypic research of maize plants.

Description

Technical Field

[0001] This invention relates to the field of computer vision and point cloud processing technology, specifically to a method for organ segmentation of maize plant point clouds that integrates dynamic weights and feature enhancement. Background Technology

[0002] Maize is one of the world's most important food crops, and the growth status of its plants directly affects subsequent yield and quality. Therefore, accurate identification and segmentation of maize plant organs are of great significance. In recent years, point cloud data, due to its rich three-dimensional spatial information, has been widely used in crop phenotypic analysis. Acquiring point cloud data of maize plant organs is crucial for agricultural automation, especially in precision agriculture. However, maize plant point cloud data is affected not only by environmental factors and noise during the acquisition process, but also by the overlapping of leaves during the growth of maize plant leaves.

[0003] Traditional point cloud segmentation algorithms struggle to accurately extract structural features from target organs, resulting in low segmentation accuracy. Existing deep learning-based point cloud segmentation methods only provide a general solution to the point cloud segmentation problem and cannot accurately segment organs from maize plant point clouds. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a method for organ segmentation of maize plant point clouds that integrates dynamic weights and feature enhancement, thereby solving the problem of point cloud segmentation and the inability to accurately segment organs from maize plant point clouds.

[0005] Technical solution: The present invention provides a method for organ segmentation of maize plant point clouds that integrates dynamic weights and feature enhancement, comprising the following steps:

[0006] (1) Input maize plant point cloud data, including the spatial coordinates and additional features of the point cloud;

[0007] (2) Perform point cloud sampling and aggregation on the input point cloud data to obtain sampled point cloud data, and combine multi-scale clustering, i.e., perform sphere query by different radii to group the point cloud;

[0008] (3) Feature extraction is performed using the DWE network. The DWE network is composed of the ScoreNet dynamic weight generation module and the feature enhancement module. Dynamic weights are generated through ScoreNet, and the feature-enhanced point cloud is weighted and aggregated according to the dynamic weights.

[0009] (4) Input the point cloud and its corresponding features into the fully connected layer and perform semantic information segmentation of the point cloud.

[0010] Furthermore, in step (2), sampling points are obtained using density-based farthest point sampling, including the following steps:

[0011] (21) Receive point cloud data and set the number of sampling points n. point The number of neighbors k required for density calculation, an array for storing sampling point indices, and an array for recording the distance of each point to the nearest sampling point;

[0012] (22) Calculate the Euclidean distance matrix between any two points in the point cloud, find the k nearest neighbors of each point and their corresponding distances, and use the reciprocal of the average nearest distance of each point as the density weight.

[0013] (23) After obtaining the density weight of each point, a point in the point cloud is randomly selected as the first sampling point. To ensure that all points are likely to be selected when the farthest point is selected for the first time, the distance from all points in the point cloud to the sampling point is initialized to 10. 10 ;

[0014] (24) Among the currently unselected points, select the point with the largest weighted distance to the existing set of sampling points as the new sampling point. Calculate the distance from all points to the current sampling point and update the nearest distance from each point to the set of selected sampling points. Only retain the smallest distance, and perform n... point The sampling is performed in multiple iterations; the distance between the sampling points is obtained based on the density weights.

[0015] (25) Output the final selected n point The index of each sampling point is used to complete the sampling.

[0016] Furthermore, in step (3), the ScoreNet module constituting the DWE network generates dynamic weights based on the input point cloud data features and additional features, which are used for weighted processing of local point cloud features. The specific steps are as follows:

[0017] (31) The ScoreNet module generates dynamic weights for each point based on the local features of the point cloud input, representing the importance of the local region in the segmentation process. The values ​​are between 0 and 1.

[0018] (32) The generated dynamic weights are used to perform weighted aggregation on the sampled and grouped point cloud data to amplify the contribution of important features to the results, and vice versa.

[0019] Furthermore, in step (3), the feature enhancement module constituting the DWE network enhances the features of the point cloud data through the following steps:

[0020] (S31) Obtain local aggregated point cloud data from the sampled point cloud data;

[0021] (S32) Extract neighborhood point features from the global point cloud;

[0022] (S33) Using the broadcast mechanism, the features of the sampling points are broadcast and transmitted to construct the point cloud feature matrix, forming a broadcast feature with the same dimension as the features of the neighboring points, and calculating the difference gcn between the features of the neighboring points and the broadcast feature, where gcn = groupedpoints - pointstile;

[0023] (S34) The local point feature grouped_points, the multiple copy point feature points_tile, and the difference feature gcn are concatenated to obtain the enhanced local point cloud feature grouped_points_agg;

[0024] (S35) The enhanced local point cloud features grouped_points_agg are concatenated with the local geometric information grouped_xyz to obtain the final feature representation.

[0025] The present invention discloses a maize plant point cloud organ segmentation system that integrates dynamic weights and feature enhancement, comprising:

[0026] Point cloud data module: used to input corn plant point cloud data, including the spatial coordinates and additional features of the point cloud;

[0027] Sampling module: Used to sample and aggregate the input point cloud data to obtain sampled point cloud data, and combine it with multi-scale clustering, i.e., grouping the point cloud by querying spheres with different radii;

[0028] DWE Network Module: Used for feature extraction using the DWE network; The DWE network is composed of the ScoreNet dynamic weight generation module and the feature enhancement module. It generates dynamic weights through ScoreNet and performs weighted aggregation of the feature-enhanced point cloud based on the dynamic weights.

[0029] Segmentation module: Used to input point clouds and their corresponding features into the fully connected layer and segment the semantic information of the point clouds.

[0030] Furthermore, in the sampling module, sampling points are obtained using density-based farthest point sampling, as detailed below:

[0031] Receive point cloud data and set the number of sampling points n. point The number of neighbors k required for density calculation, an array for storing sampling point indices, and an array for recording the distance of each point to the nearest sampling point;

[0032] Calculate the Euclidean distance matrix between any two points in the point cloud, find the k nearest neighbors of each point and their corresponding distances, and use the reciprocal of the average nearest distance of each point as the density weight.

[0033] After obtaining the density weight of each point, a point in the point cloud is randomly selected as the first sampling point. To ensure that all points have a chance of being selected when choosing the farthest point for the first time, the distance from all points in the point cloud to the sampling point is initialized to 10. 10 ;

[0034] Among the currently unselected points, select the point with the largest weighted distance to the existing set of sampled points as the new sampled point. Calculate the distance from all points to the current sampled point and update the nearest distance from each point to the set of selected sampled points. Only retain the smallest distance, and repeat the process n times. point The sampling is performed in multiple iterations; the distance between the sampling points is obtained based on the density weights.

[0035] Output the final selected n point The index of each sampling point is used to complete the sampling.

[0036] In the further DWE network module, the ScoreNet module that constitutes the DWE network generates dynamic weights based on the input point cloud data features and additional features, which are used for weighted processing of local point cloud features. The specific steps are as follows:

[0037] The ScoreNet module generates dynamic weights for each point based on the local features of the point cloud input, representing the importance of the local region in the segmentation process. The values ​​range from 0 to 1.

[0038] The generated dynamic weights are used to perform weighted aggregation on the sampled and grouped point cloud data, amplifying the contribution of important features to the results and conversely reducing the contribution of less important features.

[0039] Furthermore, within the DWE network module, the feature enhancement module, which constitutes the DWE network, enhances the features of the point cloud data through the following steps:

[0040] Obtain locally aggregated point cloud data from sampled point cloud data;

[0041] Extract neighborhood point features from the global point cloud;

[0042] Using a broadcast mechanism, the features of the sampling points are broadcast and transmitted to construct a point cloud feature matrix, forming a broadcast feature with the same dimension as the features of the neighboring points, and calculating the difference gcn between the features of the neighboring points and the broadcast feature, where gcn = groupedpoints - pointstile;

[0043] The local point cloud feature grouped_points, the multiple copy point feature points_tile, and the difference feature gcn are concatenated to obtain the enhanced local point cloud feature grouped_points_agg;

[0044] The enhanced local point cloud features grouped_points_agg are concatenated with the local geometric information grouped_xyz to obtain the final feature representation.

[0045] The present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements any of the above-described methods for organ segmentation of maize plant point clouds by fusing dynamic weights and feature enhancement.

[0046] The present invention provides a storage medium storing a computer program, which, when executed by a processor, implements any one of the methods for organ segmentation of maize plant point clouds that integrates dynamic weights and feature enhancement.

[0047] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: It uses density-weighted farthest point sampling in point cloud sampling, reducing the impact of noise and density inhomogeneity, and effectively improving the coverage of sampling points. By fusing dynamic weights and feature enhancement, a DWE feature extraction network more suitable for maize plant point cloud segmentation is constructed, greatly improving the segmentation accuracy of maize plant point clouds. Using this method for maize plant point cloud segmentation can yield more accurate maize plant point cloud segmentation examples, providing high-quality 3D point cloud data for maize plant phenotypic studies. Attached Figure Description

[0048] Figure 1 This is a flowchart of the present invention;

[0049] Figure 2 This is a comparison diagram of the sampling methods of the present invention;

[0050] Figure 3 This is a structural diagram of the feature enhancement module of the present invention;

[0051] Figure 4 This is a structural diagram of the ScoreNet weighted module of the present invention;

[0052] Figure 5 This is a comparison diagram of the segmentation results of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0054] like Figure 1As shown, this embodiment of the invention provides a 3D point cloud segmentation method that integrates dynamic weights and feature enhancement. Based on the input point cloud data, a feature learning network performs multi-scale group learning. After feature propagation, the learned features are used to obtain the score of each point, thus achieving the point cloud segmentation task. The difference between this invention and the original PointNet++ network lies in the construction of the feature learning network. By using density-based farthest point sampling and a DWE feature extraction network, a feature learning network more suitable for corn plant point cloud segmentation is constructed.

[0055] a1. Density-based farthest point sampling

[0056] The original PointNet++ network uses the farthest point sampling method in its feature learning network, focusing only on the farthest distance between points for point cloud downsampling. In complex scenes, this method is affected by uneven point cloud density distribution and outliers, leading to a decrease in sampling quality. This invention adopts density-based farthest point sampling, and the specific operation is as follows:

[0057] a1.1 Based on a given point cloud P = {p1, p2, ..., p...} N}, the coordinates of each point are p i =(x i ,y i ,z i Set the number of sampling points n point The algorithm calculates the number of neighbors k for density calculation, and initializes the array controids to store sampled points and the array distance to record the nearest distance of each point.

[0058] a1.2 The Euclidean distance between two points is calculated using the following formula:

[0059]

[0060] For each point, find its k nearest neighbors and calculate the corresponding Euclidean distance set as: {d i1 ,d i2 ,···,d ik The reciprocal of the average nearest distance of each point is used as the density weight for that point, where the density weight ρ(p i ) is defined as Density weight ρ(p) i The larger the value of p, the greater the value of p. i The higher the density of the area.

[0061] a1.3 Randomly select a point as the first sampling point c1, and set the initial distance D(p) of all points. i Set it to a large value, for example: D(p) i )=∞, In each iteration (a total of n) point (Next time), select the point c with the largest (weighted) distance between the currently unsampled point and all sampled points. k Its formula is:

[0062]

[0063] Where C = (C1C2, ..., C k-1 ) is the set of selected sampling points. For each unsampled point p i Update its nearest distance D(p) to the sample set C. i The formula is:

[0064]

[0065] a1.4 Output the selected n point The array of indices for each sampling point is called controids.

[0066] like Figure 2 As shown, comparing the two different sampling methods, density-based farthest point sampling not only effectively avoids the influence of outliers by taking density into account, but also optimizes the uniformity of sampling, effectively ensuring that more representative points are selected within the region.

[0067] a2.DWE Feature Extraction Network

[0068] In the feature learning network, the feature extraction module, as described in the original PointNet++ network, uses PointNet layers to extract and aggregate features from the space composed of sampling points and neighboring points. However, this only considers the relative positions between sampling points and their corresponding neighbors, resulting in insufficient extraction of local feature information. This invention employs a DWE (Dynamic Weighting and Enhancement) module for feature extraction. The DWE module is a fusion of a ScoreNet dynamic weighting network and a feature enhancement network, where the feature enhancement module is as follows... Figure 2 As shown.

[0069] For the input point cloud data, considering the relationship between the features of neighboring points and the features of the sampled points, the features of the sampled points are transmitted through a broadcast mechanism to obtain broadcast features. The difference between the features of neighboring points and the broadcast features is calculated to obtain geometric features. By concatenating the features of neighboring points, broadcast features, and geometric features, a fused feature is obtained, which expands the feature dimension and increases the feature expressive power. Finally, the fused feature is concatenated with relative coordinates to obtain the final feature representation.

[0070] SoreNet is a fully connected network (MLP) used to compute dynamic weights. These dynamic weights are used to adjust the influence of different neighboring points, thereby better aggregating neighborhood features. Specifically, the structure of ScoreNet is as follows:

[0071] Input layer: The input point cloud has a shape of [B, N, K, D], representing the features of each sampling point and its neighboring points. Here, B is the batch size, N is the number of sampling points (representing N representative points selected from the point cloud), K is the number of neighboring points for each sampling point, and D is the feature dimension of each point.

[0072] Fully Connected Layers: Feature extraction is performed through ScoreNet's fully connected layers (MLPs). These include multiple fully connected layers, with the number of layers and neurons per layer determined by `score_mlp`, which is the number of units in the MLP layer specified during network definition. In SA1, `score_mlp` = [64, 128, 256]. In SA2, `score_mlp` = [128, 256, 512]. Specifically, `score_mlp` is a list specifying the number of neurons in each layer. For example, if `score_mlp = ...`

[0073] With a feature size of [64, 128, 256], ScoreNet will have three fully connected layers, containing 64, 128, and 256 neurons respectively. In each layer, the fully connected layer performs a linear transformation on the input features and a non-linear transformation using the ReLU activation function. Output layer: After multiple fully connected layers, the output of ScoreNet is typically the weights of each neighboring point, which influence the contribution of that neighboring point to feature aggregation. The output shape is [B, N, K, 1], meaning each neighboring point of each sampled point has a weight value. After feature enhancement, the feature dimension of the input point cloud is increased from 3+D. in Expanded to 3+3*D in The corresponding dynamic weight generation process is as follows: Figure 3 As shown.

[0074] The weights calculated using ScoreNet represent the contribution of each neighboring point to its corresponding sampled point. Each sampled point corresponds to a set of weights, indicating the importance of its neighbors to the features of that sampled point. The resulting weights (of shape [B, N, K, 1]) are then weighted with the features of each neighboring point in the local features. Specifically, for each sampled point and its neighbors, the local features are multiplied by the calculated weight values ​​to obtain the weighted features.

[0075] Assuming that after ScoreNet computation, a weight tensor of shape [B, N, K, 1] is obtained, the weighted features of each sampling point and its neighborhood points can be represented as:

[0076] Weighted features = Local features × Weight tensor. (4)

[0077] The shape of the local features is [B,N,K,D], and the shape of the weight tensor is [B,N,K,1], which weights the features of each neighboring point.

[0078] The weighted features are further processed through convolution operations. Finally, after max pooling aggregation, a feature representation with a higher level of abstraction is obtained for each sampling point.

[0079] Segmentation experiments were conducted on the syau single maize 3D point cloud instance segmentation dataset and compared with PointNet and PointNet++, such as... Figure 5 As shown in Table 1, (a), (b), (c), and (d) are the segmentation results of PointNet, PointNet++_SSG, PointNet++_MSG, and the present invention, respectively. It can be seen that PointNet performs the worst, with segmentation errors present in all three leaf regions. The two versions of PointNet++ show deficiencies in detail in overlapping leaf regions and areas of uneven density. The method of the present invention provides the most complete handling of the boundaries between leaves and between leaves and stems. Specific segmentation accuracy is shown in Table 1. Experimental results demonstrate that, compared to PointNet and PointNet++, the method proposed in this invention exhibits significantly superior performance in organ segmentation of maize plant point clouds.

[0080] Table 1 shows the segmentation results on the Syau Single Maize dataset.

[0081]

[0082] In summary, this invention improves upon the original PointNet++ model by using density-based farthest point sampling and a DWE feature extraction module that integrates dynamic weights and feature enhancement in the feature learning network, thereby improving network performance and making it more suitable for organ segmentation of maize plant point clouds.

[0083] The present invention discloses a maize plant point cloud organ segmentation system that integrates dynamic weights and feature enhancement, comprising:

[0084] Point cloud data module: used to input corn plant point cloud data, including the spatial coordinates and additional features of the point cloud;

[0085] The sampling module is used to sample and aggregate the input point cloud data to obtain sampled point cloud data. It then combines multi-scale clustering (i.e., grouping the point cloud by querying spheres with different radii) to form point cloud groups. Sampling points are obtained using density-based farthest point sampling, as detailed below:

[0086] Receive point cloud data and set the number of sampling points n. point The number of neighbors k required for density calculation, an array for storing sampling point indices, and an array for recording the distance of each point to the nearest sampling point;

[0087] Calculate the Euclidean distance matrix between any two points in the point cloud, find the k nearest neighbors of each point and their corresponding distances, and use the reciprocal of the average nearest distance of each point as the density weight.

[0088] After obtaining the density weight of each point, a point in the point cloud is randomly selected as the first sampling point. To ensure that all points have a chance of being selected when choosing the farthest point for the first time, the distance from all points in the point cloud to the sampling point is initialized to 10. 10 ;

[0089] Among the currently unselected points, select the point with the largest weighted distance to the existing set of sampled points as the new sampled point. Calculate the distance from all points to the current sampled point and update the nearest distance from each point to the set of selected sampled points. Only retain the smallest distance, and repeat the process n times. point The sampling is performed in multiple iterations; the distance between the sampling points is obtained based on the density weights.

[0090] Output the final selected n point The index of each sampling point is used to complete the sampling.

[0091] DWE Network Module: Used for feature extraction using the DWE network; the DWE network is composed of a ScoreNet dynamic weight generation module and a feature enhancement module. Dynamic weights are generated through ScoreNet, and the feature-enhanced point cloud is weighted and aggregated based on these dynamic weights; the specific steps are as follows:

[0092] The ScoreNet module generates dynamic weights for each point based on the local features of the point cloud input, representing the importance of the local region in the segmentation process. The values ​​range from 0 to 1.

[0093] The generated dynamic weights are used to perform weighted aggregation on the sampled and grouped point cloud data, amplifying the contribution of important features to the results and conversely reducing the contribution of less important features.

[0094] Furthermore, within the DWE network module, the feature enhancement module, which constitutes the DWE network, enhances the features of the point cloud data through the following steps:

[0095] Obtain locally aggregated point cloud data from sampled point cloud data;

[0096] Extracting neighborhood point features from the global point cloud;

[0097] Using a broadcast mechanism, the features of the sampling points are broadcast and transmitted to construct a point cloud feature matrix, forming a broadcast feature with the same dimension as the features of the neighboring points, and calculating the difference gcn between the features of the neighboring points and the broadcast feature, where gcn = groupedpoints - pointstile;

[0098] The local point cloud feature grouped_points, the multiple copy point feature points_tile, and the difference feature gcn are concatenated to obtain the enhanced local point cloud feature grouped_points_agg;

[0099] The enhanced local point cloud features grouped_points_agg are concatenated with the local geometric information grouped_xyz to obtain the final feature representation.

[0100] Segmentation module: Used to input point clouds and their corresponding features into the fully connected layer and segment the semantic information of the point clouds.

[0101] This invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements any of the above-described methods for organ segmentation of maize plant point clouds by fusing dynamic weights and feature enhancement.

[0102] This invention also provides a storage medium storing a computer program, which, when executed by a processor, implements any one of the methods for organ segmentation of maize plant point clouds that integrates dynamic weights and feature enhancement.

Claims

1. A method for organ segmentation of maize plant point clouds that integrates dynamic weights and feature enhancement, characterized in that, Includes the following steps: (1) Input maize plant point cloud data, including the spatial coordinates and additional features of the point cloud; (2) Perform point cloud sampling and aggregation on the input point cloud data to obtain sampled point cloud data, and combine multi-scale clustering, i.e., perform sphere query by different radii to group the point cloud; (3) Feature extraction is performed using the DWE network. The DWE network is composed of the ScoreNet dynamic weight generation module and the feature enhancement module. Dynamic weights are generated through ScoreNet, and the feature-enhanced point cloud is weighted and aggregated according to the dynamic weights. Among them, the ScoreNet dynamic weight generation module is used for weighted processing of local point cloud features. The feature enhancement module enhances the features of point cloud data through the following steps: (S31) Obtain local aggregated point cloud data from the sampled point cloud data; (S32) Extract neighborhood point features from the global point cloud; (S33) Using a broadcast mechanism, the features of the sampling points are broadcast and transmitted to construct a point cloud feature matrix, forming a broadcast feature with the same dimension as the features of neighboring points, and the difference gcn between the features of neighboring points and the broadcast feature is calculated, where ; (S34) The local point feature grouped_points, the multiple copy point feature points_tile, and the difference feature gcn are concatenated to obtain the enhanced local point cloud feature grouped_points_agg; (S35) The enhanced local point cloud features grouped_points_agg are concatenated with the local geometric information grouped_xyz to obtain the final feature representation; (4) Input the point cloud and its corresponding features into the fully connected layer and perform semantic information segmentation of the point cloud.

2. The method for organ segmentation of maize plant point clouds by fusing dynamic weights and feature enhancement as described in claim 1, characterized in that, In step (2), sampling points are obtained using density-based farthest point sampling, including the following steps: (21) Receive point cloud data and set the number of sampling points. The number of neighbors k required for density calculation, an array for storing sampling point indices, and an array for recording the distance of each point to the nearest sampling point; (22) Calculate the Euclidean distance matrix between any two points in the point cloud, find the k nearest neighbors of each point and their corresponding distances, and use the reciprocal of the average nearest distance of each point as the density weight. (23) After obtaining the density weight of each point, a point in the point cloud is randomly selected as the first sampling point. To ensure that all points are likely to be selected when the farthest point is selected for the first time, the distance from all points in the point cloud to the sampling point is initialized to 10. 10 ; (24) Among the currently unselected points, select the point with the largest weighted distance to the existing set of sampling points as the new sampling point. Calculate the distance from all points to the current sampling point and update the nearest distance from each point to the set of selected sampling points. Only retain the smallest distance. The sampling is performed in multiple iterations; the distance between the sampling points is obtained based on the density weights. (25) Output the final selection The index of each sampling point is used to complete the sampling.

3. The method for organ segmentation of maize plant point clouds by fusing dynamic weights and feature enhancement as described in claim 1, characterized in that, In step (3), the ScoreNet dynamic weight generation module that constitutes the DWE network generates dynamic weights based on the input point cloud data features and additional features. The specific steps are as follows: (31) The ScoreNet dynamic weight generation module generates dynamic weights for each point based on the local features of the point cloud input, representing the importance of the local region in the segmentation process. The values ​​are between 0 and 1. (32) The generated dynamic weights are used to perform weighted aggregation on the sampled and grouped point cloud data, amplifying the contribution of important features to the results, and conversely reducing the contribution of less important features.

4. A point cloud organ segmentation system for maize plants that integrates dynamic weights and feature enhancement, characterized in that, include Point cloud data module: used to input corn plant point cloud data, including the spatial coordinates and additional features of the point cloud; Sampling module: Used to sample and aggregate the input point cloud data to obtain sampled point cloud data, and combine it with multi-scale clustering, i.e., grouping the point cloud by querying spheres with different radii; The DWE network module is used for feature extraction using the DWE network. The DWE network is a fusion of the ScoreNet dynamic weight generation module and the feature enhancement module. It generates dynamic weights through ScoreNet and performs weighted aggregation of the feature-enhanced point cloud based on these dynamic weights. The ScoreNet dynamic weight generation module is used for weighted processing of local point cloud features. The feature enhancement module performs feature enhancement on the point cloud data as follows: It obtains locally aggregated point cloud data from the sampled point cloud data; it extracts neighborhood point features from the global point cloud; using a broadcast mechanism, it broadcasts the sampled point features to construct a point cloud feature matrix, forming a broadcast feature with the same dimension as the neighborhood point features, and calculates the difference gcn between the neighborhood point features and the broadcast feature. The local point feature grouped_points, the multiple copy point feature points_tile, and the difference feature gcn are concatenated to obtain the enhanced local point cloud feature grouped_points_agg; the enhanced local point cloud feature grouped_points_agg is then concatenated with the local geometric information grouped_xyz to obtain the final feature representation. Segmentation module: Used to input point clouds and their corresponding features into the fully connected layer and segment the semantic information of the point clouds.

5. The maize plant point cloud organ segmentation system according to claim 4, characterized in that, In the sampling module, sampling points are obtained using density-based farthest point sampling, as follows: Receive point cloud data and set the number of sampling points. The number of neighbors k required for density calculation, an array for storing sampling point indices, and an array for recording the distance of each point to the nearest sampling point; Calculate the Euclidean distance matrix between any two points in the point cloud, find the k nearest neighbors of each point and their corresponding distances, and use the reciprocal of the average nearest distance of each point as the density weight. After obtaining the density weight of each point, a point in the point cloud is randomly selected as the first sampling point. To ensure that all points have a chance of being selected when choosing the farthest point for the first time, the distance from all points in the point cloud to the sampling point is initialized to 10. 10 ; Among the currently unselected points, select the point with the largest weighted distance to the existing set of sampled points as the new sampled point. Calculate the distance from all points to the current sampled point and update the nearest distance from each point to the set of selected sampled points. Only retain the minimum distance. Sub-iteration sampling; The distance between sampling points is obtained based on density weights; Output the final selection The index of each sampling point is used to complete the sampling.

6. The maize plant point cloud organ segmentation system integrating dynamic weights and feature enhancement according to claim 4, characterized in that, In the DWE network module, the ScoreNet dynamic weight generation module, which constitutes the DWE network, generates dynamic weights based on the input point cloud data features and additional features, as follows: The ScoreNet dynamic weight generation module generates dynamic weights for each point based on the local features of the point cloud input. These weights represent the importance of the local region in the segmentation process, and the values ​​range from 0 to 1. The generated dynamic weights are used to perform weighted aggregation on the sampled and grouped point cloud data, amplifying the contribution of important features to the results and conversely reducing the contribution of less important features.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the computer program is loaded into the processor, it implements a method for organ segmentation of maize plant point cloud by fusing dynamic weights and feature enhancements as described in any one of claims 1-3.

8. A storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for organ segmentation of maize plant point cloud based on any one of claims 1-3, which integrates dynamic weights and feature enhancement.

Images