A marine biological classification method and system based on graph neural network

Abstract

The present invention relates to the field of neural network technology, and in particular to a method and system for classifying marine organisms based on graph neural networks. The method comprises acquiring point cloud data of marine organisms; constructing a three-dimensional grid model based on the acquired point cloud data, and extracting key morphological features using the three-dimensional grid model; training the graph neural network using the extracted key morphological features; and performing classification prediction based on the trained graph neural network. The present invention can capture the fine three-dimensional structural features of marine organisms by combining high-precision three-dimensional modeling and advanced graph neural network technology, thereby significantly improving the accuracy of classification.

Description

Technical Field

[0001] The present invention relates to the field of neural network technology, and in particular to a method and system for classifying marine organisms based on graph neural network. Background Art

[0002] Marine biological classification is of great significance for marine ecological research, resource management and biodiversity conservation. However, due to the large number of marine biological species and diverse morphologies, and the difficulty in obtaining complete samples of some organisms, traditional classification methods face many challenges. Currently, the widely used classification techniques include morphological classification, DNA barcoding technology and 2D image-based recognition methods. Although morphological classification is intuitive, it relies too much on expert experience, is highly subjective, and is difficult to deal with species with similar morphology; although DNA barcoding technology has high accuracy, it is expensive and difficult to promote and apply in large-scale classification; although 2D image-based recognition methods are efficient, they often ignore the three-dimensional structural information of organisms, resulting in limited classification accuracy. In recent years, the application of three-dimensional modeling technology in the field of biology has become increasingly widespread, providing new possibilities for capturing the three-dimensional structural characteristics of organisms. At the same time, graph neural networks (GNNs) have shown superior performance in classification tasks, especially when processing data with complex relational structures. However, existing marine biological classification methods still have many shortcomings, mainly manifested in: failure to fully utilize the three-dimensional structural characteristics of marine organisms, lack of effective integration of morphological and molecular biological information, and the generalization ability and interpretability of classification models need to be improved. In view of this, the development of a new marine organism classification method that combines three-dimensional modeling and GNN has important theoretical significance and application value, and is expected to break through the limitations of existing technologies and improve the accuracy, efficiency and scalability of marine organism classification. Summary of the invention

[0003] In order to solve the above-mentioned problems, the present invention provides a marine organism classification method and system based on graph neural network.

[0004] In the first aspect, the present invention provides a method for classifying marine organisms based on a graph neural network, which adopts the following technical solution:

[0005] A marine organism classification method based on graph neural network, comprising:

[0006] Obtain point cloud data of marine life;

[0007] A three-dimensional mesh model is constructed based on the acquired point cloud data, and key morphological features are extracted using the three-dimensional mesh model;

[0008] Train the graph neural network using the extracted key morphological features;

[0009] Perform classification prediction based on the trained graph neural network.

[0010] Furthermore, the acquisition of point cloud data of marine organisms includes using an underwater 3D laser scanner to collect point cloud data of marine organisms from multiple angles. To ensure data quality, an iterative closest point ICP algorithm is used to align the multi-angle point clouds. Each collected point is represented as:

[0011]

[0012] Among them, (x i ,y i ,z i ) represents the three-dimensional coordinates of the point, I i Indicates the reflection intensity of a point;

[0013] The minimization error function is expressed as:

[0014]

[0015] Among them, R is the rotation matrix, t is the translation vector, q i are the corresponding points in the target point cloud.

[0016] Furthermore, the three-dimensional mesh model is constructed based on the acquired point cloud data, including applying a Poisson surface reconstruction algorithm to generate an initial mesh model M = (V, F), where V is a vertex set, F is a face set, and the zero isosurface obtained by solving a partial differential equation in Poisson surface reconstruction is the reconstructed surface, and Laplace smoothing is used to optimize the model surface.

[0017] Furthermore, the method of extracting key morphological features using a three-dimensional mesh model includes generating required shape indexes and shape descriptors by calculating curvature information of the model surface, and fusing the shape indexes and shape descriptors with biological data to construct a comprehensive feature vector.

[0018] Furthermore, the method of training the graph neural network using the extracted key morphological features includes using the graph neural network to synthesize the graph structure and train the classification model, wherein a similarity matrix is ​​constructed, and based on the similarity matrix, a k-nearest neighbor algorithm is used to determine the edge set of the graph, and a graph convolutional neural network GCN is used to complete the training. The operation of GCN is expressed as:

[0019]

[0020] in, is the degree matrix, is the adjacency matrix, is the node feature of the lth layer, is a learnable parameter, is the ReLU activation function.

[0021] Furthermore, the method of training the graph neural network using the extracted key morphological features also includes introducing a multi-head attention mechanism to optimize the graph convolutional neural network in order to improve the expressive power of the model, which is specifically expressed as follows:

[0022]

[0023]

[0024] Where k represents the number of attention heads, a k and W k is a learnable parameter;

[0025] The loss function is:

[0026]

[0027] in, is the cross entropy loss, is the L2 regularization term, which is used to prevent the model from overfitting. It is a contrast loss, which is used to enhance the discriminability of features; and is the balance coefficient, which is determined by cross-validation.

[0028] Furthermore, the classification prediction based on the trained graph neural network includes performing classification prediction through a trained GNN model:

[0029]

[0030] Evaluate the confidence of the prediction:

[0031]

[0032]

[0033] When C < threshold1 (threshold1=0.75) or Entropy > threshold2 (hreshold2=0.6), the expert review mechanism is triggered.

[0034] The second aspect is a marine biological classification system based on graph neural network, including:

[0035] The data acquisition module is configured to acquire point cloud data of marine organisms;

[0036] A three-dimensional construction module is configured to construct a three-dimensional mesh model based on the acquired point cloud data, and extract key morphological features using the three-dimensional mesh model;

[0037] The model training module is configured to train the graph neural network using the extracted key morphological features;

[0038] The classification prediction module is configured to perform classification prediction based on the trained graph neural network.

[0039] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device for a method of classifying marine organisms based on a graph neural network.

[0040] In a fourth aspect, the present invention provides a terminal device including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the method for classifying marine organisms based on graph neural networks.

[0041] In summary, the present invention has the following beneficial technical effects:

[0042] (1) Improve classification accuracy: By combining high-precision 3D modeling and advanced graph neural network technology, the present invention can capture the fine 3D structural features of marine organisms, significantly improving classification accuracy. Compared with traditional 2D image classification methods, the present invention can better distinguish species with similar morphology but different 3D structures, reducing misclassification.

[0043] (2) Enhanced generalization ability: The use of the GNN model and multi-head attention mechanism enables the present invention to have stronger generalization ability when facing unknown or rare species. This feature is of great significance for marine biodiversity research and new species discovery.

[0044] (3) Improve classification efficiency: This invention realizes an automated process from data collection to classification prediction, greatly improving the efficiency of marine biological classification. Especially when dealing with large-scale samples, it can significantly save time and labor costs compared to traditional methods that rely on expert experience.

[0045] (4) Integration of multi-source information: By integrating three-dimensional morphological features, DNA sequence information and ecological environment parameters, the present invention provides a comprehensive classification basis. This multi-dimensional information integration helps to improve the reliability and scientificity of classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a schematic diagram of a marine organism classification method based on a graph neural network according to Example 1 of the present invention. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0048] Example 1

[0049] Reference Figure 1 , a method for classifying marine organisms based on a graph neural network in this embodiment includes:

[0050] Obtain point cloud data of marine life;

[0051] A three-dimensional mesh model is constructed based on the acquired point cloud data, and key morphological features are extracted using the three-dimensional mesh model;

[0052] Train the graph neural network using the extracted key morphological features;

[0053] Perform classification prediction based on the trained graph neural network.

[0054] Specifically, the following steps are included:

[0055] S1 Data Collection,

[0056] A high-precision underwater 3D laser scanner is used to collect point cloud data of marine life from multiple angles. Each collected point is represented as:

[0057]

[0058] Among them, (x i ,y i ,z i ) represents the three-dimensional coordinates of the point, I i Represents the reflection intensity of the point. To ensure data quality, the iterative closest point (ICP) algorithm is used to align the multi-angle point cloud to minimize the error function:

[0059]

[0060] Among them, R is the rotation matrix, t is the translation vector, q i are the corresponding points in the target point cloud.

[0061] The aligned point cloud data is processed through a bilateral filtering algorithm to process the data noise. At the same time, biological data D is collected, including DNA sequence S and ecological environment parameters E = {temperature T, salinity S a , depth D, pH value} to ensure the accuracy and timeliness of the data.

[0062] S2 3D modeling,

[0063] Based on the collected high-quality point cloud data, this step aims to construct an accurate three-dimensional mesh model and extract key morphological features.

[0064] First, the Poisson surface reconstruction algorithm is applied to generate the initial mesh model M = (V, F), where V is the vertex set and F is the face set. Poisson reconstruction solves the partial differential equation: V, where X is the indicator function, and V is the normal vector field of the oriented point cloud. The zero isosurface of X obtained by solving the equation is the reconstructed surface. The model surface is optimized using Laplace smoothing, which can be specifically expressed as:

[0065]

[0066] in, is the smoothing coefficient, in the present invention The value of is 0.3, is the edge weight, for The first-order neighborhood vertex set of .

[0067] Then, the key morphological features are extracted, the principal curvatures k1 and k2 are calculated, and then the Gaussian curvature K and the mean curvature H are obtained:

[0068]

[0069] About extracting key morphological features:

[0070] The main curvatures k1 and k2 are calculated using the formula, which can be expressed as:

[0071]

[0072] in, and is the vertex position after optimization by the Laplace smoothing optimization model, Represents an operation that computes eigenvalues.

[0073] Then the Gaussian curvature K and mean curvature H calculated using k1 and k2 are used as key morphological features.

[0074] Based on the curvature information, generate the required shape index S I And the shape descriptor SHOT, which can be described as:

[0075]

[0076] The SHOT descriptor is a statistical representation of local surface geometry, and a 364-dimensional vector is selected in the present invention.

[0077] The choice of vector dimension is arbitrary, and the choice of 364-dimensional vector can be removed. The purpose of calculating SI and shape descriptors is to fuse all the calculated results to construct a comprehensive vector, so that we can use different information to get a better result during training.

[0078] Finally, the morphological features are fused with the biological data to construct a comprehensive feature vector:

[0079] The process of feature fusion is shown in the following formula F: Is a feature fusion function, using this function, we can Variables are fused into one.

[0080]

[0081] in, It is a feature fusion function, which is implemented through a neural network and maps features of different dimensions and scales into a unified feature space.

[0082] S3 GNN model training,

[0083] This step uses the powerful expression ability of graph neural networks to construct a graph structure G = (V, E) based on the feature vector F obtained by three-dimensional modeling and train the classification model. First, construct the similarity matrix:

[0084]

[0085] in, is the Gaussian kernel width, set to the median of the Euclidean distance of the eigenvector F. Based on the similarity matrix , use the k nearest neighbor algorithm to determine the edge set E of the graph, k is set to log(N) + 1, and N is the total number of samples.

[0086] The training is completed using a graph convolutional neural network (GCN). The core operations of GCN are:

[0087]

[0088] in, is the degree matrix, is the adjacency matrix, is the node feature of the lth layer, is a learnable parameter, is the ReLU activation function.

[0089] In order to better improve the expressiveness of the model, the present invention introduces a multi-head attention mechanism to optimize the graph convolutional neural network, which can be specifically expressed as:

[0090]

[0091]

[0092] Where k represents the number of attention heads, a k and W k are learnable parameters.

[0093] The loss function can be designed as:

[0094]

[0095] in, is the cross entropy loss, is the L2 regularization term, which is used to prevent the model from overfitting. It is a contrast loss that can enhance the discriminability of features. and is the balance coefficient, which is determined by cross-validation.

[0096] S4 Classification prediction,

[0097] Classify and predict new marine biological samples. First, repeat steps S1-S2 to obtain the feature vector F of the new sample. x . x Embedded into the trained graph structure:

[0098]

[0099] Among them, sim is the cosine similarity function, and TopK returns the K nodes with the highest similarity.

[0100] Classification prediction is performed through the trained GNN model G:

[0101]

[0102] Evaluate the confidence of the prediction:

[0103]

[0104]

[0105] When C < threshold1 (threshold1=0.75) or Entropy > threshold2 (hreshold2=0.6), the expert review mechanism is triggered.

[0106] Every time 5%-10% of the original training set size is accumulated, the graph structure is rebuilt to reflect the change in data distribution. Classification accuracy, F1 score, confusion matrix and other indicators are calculated for every 1,000 new samples, and model hyperparameters or structures are adjusted based on the evaluation results.

[0107] Example 2

[0108] This embodiment provides a marine life classification system based on a graph neural network, including:

[0109] The data acquisition module is configured to acquire point cloud data of marine organisms;

[0110] A three-dimensional construction module is configured to construct a three-dimensional mesh model based on the acquired point cloud data, and extract key morphological features using the three-dimensional mesh model;

[0111] The model training module is configured to train the graph neural network using the extracted key morphological features;

[0112] The classification prediction module is configured to perform classification prediction based on the trained graph neural network.

[0113] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for example, a method for classifying marine organisms based on a graph neural network.

[0114] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the marine life classification method based on graph neural network.

[0115] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A marine organism classification method based on graph neural network, characterized in that: include: Obtain point cloud data of marine life; A three-dimensional mesh model is constructed based on the acquired point cloud data, and key morphological features are extracted using the three-dimensional mesh model; Train the graph neural network using the extracted key morphological features; Classification prediction based on trained graph neural network; The acquisition of point cloud data of marine organisms includes using an underwater 3D laser scanner to collect point cloud data of marine organisms from multiple angles. To ensure data quality, an iterative closest point ICP algorithm is used to align the multi-angle point clouds. Each collected point is represented as: in,( x i ,y i ,z i ) represents the three-dimensional coordinates of the point, I i Indicates the reflection intensity of a point; The minimization error function is expressed as: Among them, R is the rotation matrix, t is the translation vector, q i is the corresponding point in the target point cloud; The three-dimensional mesh model is constructed based on the acquired point cloud data, including applying a Poisson surface reconstruction algorithm to generate an initial mesh model. M = (V, F) , V is the vertex set, F For a set of patches, the zero isosurface obtained by solving the partial differential equation in Poisson surface reconstruction is the reconstructed surface, and the model surface is optimized using Laplace smoothing; The method of extracting key morphological features using a three-dimensional mesh model includes generating a required shape index and shape descriptor by calculating curvature information of the model surface, and constructing a comprehensive feature vector by fusing the shape index and shape descriptor with biological data, wherein the biological data includes DNA sequences and ecological environment parameters; The method of training the graph neural network using the extracted key morphological features includes constructing a graph structure based on the key morphological features and training a classification model, wherein a similarity matrix is ​​constructed, and based on the similarity matrix, a k-nearest neighbor algorithm is used to determine the edge set of the graph, and a graph convolutional neural network GCN is used to complete the training. The operation of GCN is expressed as follows: in, is the degree matrix, is the adjacency matrix, For the l The node characteristics of the layer, is a learnable parameter, is the ReLU activation function; The method of training the graph neural network by using the extracted key morphological features also includes introducing a multi-head attention mechanism to optimize the graph convolutional neural network in order to improve the expressive power of the model, which is specifically expressed as follows: in k represents the number of attention heads, a k and W k is a learnable parameter; The loss function is: in, is the cross entropy loss, for L2 Regularization term, used to prevent the model from overfitting. It is a contrast loss, which is used to enhance the discriminability of features; and is the balance coefficient, determined by cross-validation; The classification prediction based on the trained graph neural network includes classification prediction through the trained GNN model G: Evaluate the confidence of the prediction: when C < threshold 1 , threshold 1 =0.75 or Entropy > threshold 2 , threshold 2 =0.6, the expert review mechanism is triggered. is the feature vector of the new sample.

2. A marine biological classification system based on graph neural network, executing the method according to claim 1, characterized in that: include: The data acquisition module is configured to acquire point cloud data of marine organisms; A three-dimensional construction module is configured to construct a three-dimensional mesh model based on the acquired point cloud data, and extract key morphological features using the three-dimensional mesh model; The model training module is configured to train the graph neural network using the extracted key morphological features; The classification prediction module is configured to perform classification prediction based on the trained graph neural network.

3. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

4. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method as claimed in claim 1 .

Images