A small sample point cloud semantic segmentation method

Abstract

The application belongs to the technical field of point cloud data processing, and particularly relates to a small sample point cloud semantic segmentation method. First, a meta-learning training strategy is used to train a constructed semantic segmentation network model; the semantic segmentation network model comprises a feature extraction network, a multi-prototype generation model and a relationship learning network; the feature extraction network is used for respectively performing semantic feature extraction on a support set and a query set to obtain support set features and query set features, the multi-prototype generation model is used for performing feature selection on the support set features to obtain prototype features, and the relationship learning network is used for learning the similarity relationship between the prototype features and the query features; then, to-be-classified point cloud data is input into the trained semantic segmentation network model to obtain a classification result. The application designs a small sample classification meta-learning method, which can realize classification of a new class by using only a small amount of labeled samples when facing a brand-new scene, and can improve overall classification precision compared with a supervised deep learning method.

Description

Technical Field

[0001] This invention belongs to the field of point cloud data processing technology, specifically relating to a small sample point cloud semantic segmentation method. Background Technology

[0002] Point cloud data, rich in 3D geospatial information, has become a crucial data source in geosciences, playing a vital role in environmental monitoring, building extraction, forest resource surveys, and 3D reconstruction. The rapid development of LiDAR (Light Detection and Ranging) and photogrammetry technologies has made it easier to quickly acquire large-area, large-scale 3D point clouds. Deep learning-based point cloud semantic segmentation aims to quickly label the category of each 3D point and is a key technology for intelligent terrain interpretation and point cloud data applications. However, for large-scale urban scenarios, accurate and rapid point cloud semantic segmentation remains a significant challenge due to the sparse and disordered nature of point cloud data, uneven density distribution, and significant differences in the geometric features of ground features.

[0003] Currently, most deep learning-based point cloud semantic segmentation methods extract geometric or multi-resolution features more suitable for large-scale point clouds by improving network structure. While this can further improve classification accuracy, deep learning models typically require a large number of samples to compensate for their huge parameter space. When dealing with new scenes, existing methods often need to train the network from scratch with sufficient samples. In practice, collecting large-scale datasets is difficult; currently available city-level point cloud datasets are generated at very long intervals. In contrast, with the rapid development of LiDAR sensors and image matching technology, large amounts of unlabeled point cloud data are readily available, while the limited number of labels results in poor point cloud classification performance. Therefore, in practical applications, insufficient point cloud semantic segmentation labels have become the biggest challenge in applying complex deep learning models to unknown scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide a small-sample point cloud semantic segmentation method to solve the problem that supervised deep learning methods fail to achieve good point cloud classification results when there are insufficient semantic segmentation labels.

[0005] To address the aforementioned technical problems, this invention provides a method for semantic segmentation of small-sample point clouds, comprising the following steps:

[0006] 1) The constructed semantic segmentation network model is trained using a meta-learning training strategy; the semantic segmentation network model includes a feature extraction network, a multi-prototype generation model, and a relation learning network; the feature extraction network is used to extract semantic features from the support set and the query set respectively to obtain support set features and query set features, the multi-prototype generation model is used to select features from the support set features to obtain prototype features, and the relation learning network is used to learn the similarity relationship between prototype features and query features.

[0007] 2) Input the point cloud data to be classified into the trained semantic segmentation network model to obtain the classification results.

[0008] Its beneficial effects are as follows: This invention applies meta-learning to point cloud classification and designs a new meta-learning method for small-sample classification of large-scale, large-area point clouds. When facing a completely new scene, only a small number of labeled samples are needed to achieve classification of new categories. It is also applicable to LiDAR point clouds and photogrammetric dense matching point clouds, and has good generalization performance. Compared with supervised deep learning methods, it can improve the overall classification accuracy and improve the classification accuracy of imbalanced categories.

[0009] Furthermore, a meta-learning training strategy comprising three sequentially executed meta-learning stages is employed to train the semantic segmentation network model. The three meta-learning stages are as follows: the first stage is to perform meta-training of the semantic segmentation network model using source data; the second stage is to fine-tune the semantic segmentation network model trained in the first stage using small sample target data; and the third stage is to perform meta-testing using unlabeled target dataset.

[0010] Its beneficial effects are as follows: the first stage enables the network model to acquire learning ability and transferable knowledge, and the second stage enables the network model to effectively adapt to new point cloud scenarios, achieve classification of new categories, and have good generalization performance.

[0011] Furthermore, the feature extraction network is used to extract semantic features in the following manner: extracting geometric features of different multi-scales from the input point cloud data, connecting them with input features of multiple dimensions to obtain the extracted semantic features, and then processing them for output.

[0012] Its beneficial effect is that it can be applied to the classification of land features of different sizes by utilizing geometric features at different scales.

[0013] Furthermore, the feature extraction network is used to process the data through two multilayer perceptrons to produce the output. The input of the first multilayer perceptron is the semantic feature, and the output of the first multilayer perceptron is the input of the second multilayer perceptron. The outputs of the first and second multilayer perceptrons are concatenated to obtain the final output. The first multilayer perceptron is a lightweight multilayer perceptron, and the second multilayer perceptron is a non-linear multilayer perceptron.

[0014] Its beneficial effects are as follows: by using two multilayer perceptrons for processing, the first multilayer perceptron can quickly adapt to different tasks, and the second multilayer perceptron, as a metric learner, can obtain a deep metric space suitable for relation learning networks.

[0015] Furthermore, the multi-prototype generation model is used to select features for each classification by first using random sampling and then clustering.

[0016] Its beneficial effects are as follows: First, using random sampling can reduce the amount of computation of point clouds and improve the processing speed of point clouds, and the features of random sampling reflect the probability distribution of the overall features; then, clustering can be used to select representative features. This method is unsupervised and random, and is suitable for small sample problems.

[0017] Furthermore, the selected features are those located at the cluster centers after clustering.

[0018] Its beneficial effects are: selecting cluster centers means selecting the most representative features, which ensures classification accuracy while improving point cloud processing speed.

[0019] Furthermore, the relation learning network is used to learn the similarity relationship between prototype features and query features by employing a learnable depth metric space: prototype features and query features are concatenated into relation pairs and fed into a two-layer neural network to learn the similarity relationship between them.

[0020] Its beneficial effect is that, compared with the manual selection of similarity calculation, this method has a stronger data adaptability in the learning process.

[0021] Furthermore, the loss function used when training the semantic segmentation network model is:

[0022]

[0023]

[0024] In the formula, Let θ represent the loss, and let θ represent the parameters of the semantic segmentation network model. Let represent an N-way K-shot subtask containing a support set and a query set, where x represents the input point, y represents the true class label of the point, and α represents the learning rate during network training. f represents the gradient of the loss function. [θ] (·) represents the semantic segmentation network model, and L represents the loss of the graph points.

[0025] Furthermore, during the training of the semantic segmentation network model, N categories of tasks are randomly selected each time to train the semantic segmentation network model, and there is no overlap between the tasks.

[0026] Its beneficial effects are: randomly selecting N categories each time can ensure the richness of the task and improve the robustness of the model. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0028] Figure 2 This is a schematic diagram of the task-based meta-learning training process of the present invention;

[0029] Figure 3 This is a structural diagram of the point cloud small sample multi-prototype relation network (MR-PFS network) of the present invention;

[0030] Figure 4 This is a structural diagram of the feature extraction network in the MR-PFS network of the present invention;

[0031] Figure 5 This is a structural diagram of the multi-prototype generation model in the MR-PFS network of the present invention;

[0032] Figure 6 This is a structural diagram of the relation generation network in the MR-PFS network of the present invention;

[0033] Figures 7(a) to 7(d) show the experimental results of segmenting the CAM target dataset using the Proto-3D, attMPTI, Proto-RL, and MPM-RL methods, respectively.

[0034] Figure 7(e) shows the experimental results of the MR-PFS method of the present invention for segmenting the CAM target dataset;

[0035] Figure 7(f) shows the standard segmentation results of the CAM target dataset;

[0036] Figure 7(g) is a schematic diagram of the classification results represented by different grayscale colors in Figures 7(a) to 7(d);

[0037] Figures 8(a) to 8(d) show the experimental results of segmenting the ISPRS target dataset using the Proto-3D, attMPTI, Proto-RL, and MPM-RL methods, respectively.

[0038] Figure 8(e) shows the experimental results of the MR-PFS method of the present invention for segmenting the ISPRS target dataset;

[0039] Figure 8(f) shows the standard segmentation results of the ISPRS target dataset;

[0040] Figure 8(g) is a schematic diagram of the classification results represented by different grayscale colors in Figures 8(a) to 8(f);

[0041] Figures 9(a) to 9(d) show the experimental results of segmenting the DALES target dataset using the Proto-3D, attMPTI, Proto-RL, and MPM-RL methods, respectively.

[0042] Figure 9(e) shows the experimental results of the MR-PFS method of the present invention for segmenting the DALES target dataset;

[0043] Figure 9(f) shows the standard segmentation results of the DALES target dataset;

[0044] Figure 9(g) is a schematic diagram of the classification results represented by different grayscale colors in Figures 9(a) to 9(f). Detailed Implementation

[0045] This invention constructs a semantic segmentation network model, which is a Multi-prototype Relation Network for Point Few-shot Semantic Segmentation (MR-PFS) for few-shot semantic segmentation of 3D point clouds. It employs a meta-learning training strategy comprising three sequentially executed meta-learning stages to train the semantic segmentation network model, i.e., a few-shot learning framework containing a three-step meta-learning process. This is described in detail below.

[0046] Unlike traditional batch-based learning, task-based learning offers a clever solution for networks to learn as much knowledge as possible from a small number of labeled samples. For each training iteration, the MR-PFS network model first randomly accepts a task with N classes, forming a support set of K samples and a query set of L samples. This is also known as the N-way K-shot problem in few-shot learning. Randomly selecting N classes each time ensures task richness, improves model robustness, and eliminates overlap between tasks. However, unlike two-dimensional few-shot tasks, each point cloud set contains three coordinates of M points and additional features as input. During the meta-training process for each task, the support points are explicitly labeled, while the labels for the query set are considered unknown. The MR-PFS network model predicts the labels for the query points under the supervision of the support set. Then, by comparing the predicted labels with the ground truth values, the network parameters are optimized using the loss function. Based on the differences between tasks, the model can gain learning ability by iteratively calculating the loss.

[0047] The workflow of the three-dimensional few-shot learning method designed in this embodiment includes three stages. For example... Figure 2 As shown, specifically, the MR-PFS network model first takes the source dataset as input and acquires learning capabilities and transferable knowledge through a multi-task learning process in meta-training. Then, to enable the network to effectively adapt to new point cloud scenarios, the network model is fine-tuned with a small number of labeled target samples in the few-shot learning stage. Features learned that are relevant to previous tasks are preserved, and the network also focuses on learning differences between specific tasks. Finally, the model is tested on unknown target samples; each task is generated from a small number of labeled samples and an unlabeled test set. These three stages are interconnected with the meta-learner (MR-PFS network model), as follows: Figure 1 As shown. It is important to note that each stage is a meta-learning process, which enables the network model to learn. Moreover, the "small sample" mentioned in this embodiment refers to a small portion of the overall new dataset.

[0048] The specific structure of the point cloud semantic segmentation model used in this embodiment is as follows: Figure 3 As shown, it consists of three parts: a feature extraction network, a multi-prototype generation model, and a relation learning network.

[0049] The feature extraction network uses discrete point cloud data as input to extract multi-scale features. Its specific structure is as follows: Figure 4As shown. Due to network and memory limitations, existing point-based deep learning networks cannot directly input large-scale point clouds when dealing with airborne LiDAR point clouds, which are mainly composed of terrain data. Furthermore, the complex and diverse point cloud regions cause differences in land cover types and imbalances in data volume among various point cloud types, leading to oscillations in the optimization process. Simultaneously, the large coordinate scale of airborne data makes it unsuitable for direct extraction of local features. Therefore, at the network input front end, this invention performs block processing, coordinate scale adjustment, and data augmentation on the point cloud to make the model more adaptable to large-scale point clouds. This part belongs to the front-end processing operations of the network input. During feature extraction, considering the experience of traditional methods in feature extraction and the significant geometric differences in land cover attributes, this invention focuses on the geometric features of aerial point clouds. It utilizes hierarchical convolutions at different scales to encode the geometric features of different neighborhoods of points. While edge convolution can obtain local features by constructing a directed dynamic graph, this invention improves upon this by dividing the neighborhood into different sizes to achieve geometric feature extraction at different scales, thus adapting to land cover of different sizes. These geometric features are concatenated with encoded input features from multiple dimensions to extract semantic features. A lightweight multilayer perceptron (MLP) is then used to quickly adapt to different tasks. Finally, a non-linear multilayer perceptron (MLP) is used as a metric learner to obtain a deep metric space suitable for the subsequent relational network. Specifically, the connection is as follows: the output of the first MLP serves as the input of the second MLP, and the final output is obtained by concatenating the outputs of the first and second MLPs. Furthermore, combining... Figure 1 It can be seen that the support set data is processed by a feature extraction network to obtain support set features, and the query set data is processed by a feature extraction network to obtain query set features. Among these, Figure 4 N×3 and N×C in Indicates the data size; N×3 represents N points, each with three-dimensional coordinates, N×C in The other dimensional features of each point are input.

[0050] The multi-prototype generation model aims to represent the data distribution supporting the sample set using multiple prototypes, while significantly reducing the computational cost of point clouds. Specifically, considering the large scale of the point cloud, points are selected using random sampling as the initial prototype for each class. This approach better describes the distribution, and the specific structure is as follows: Figure 5As shown. Since point cloud partitions are typically located within a 100-square-meter area and contain a limited number of ground objects, clustering is further used to select the most representative features as the final prototype. From a computational perspective, random sampling significantly improves the speed of point cloud processing compared to farthest-point sampling, and randomly sampled features reflect the probability distribution of the overall features. Although clustering is unsupervised and random in feature selection, it is well-suited for small-sample problems. Figure 1 The support set features are used to generate prototype features through a multi-prototype generation model.

[0051] The design of the relational learning network aims to enable the model to learn from data with different distributions, thereby learning the categories of query set samples. Its specific structure is as follows: Figure 6 As shown. Due to the differences in the distribution of feature vectors extracted from features and generated from multiple prototypes, directly calculating the similarity distance between embedded features can obscure their semantic relationships. Therefore, this embodiment does not use a distance-based similarity calculation method, but instead uses a learnable depth metric space to compare the relationships between features. Prototype features and query set features are first concatenated into relation pairs, and then fed into a two-layer neural network to learn their similarity. The learning process is more data-adaptive compared to manually selecting a similarity calculation method. This embodiment sets the principle that relation pairs with the same category combination have scores closer to 1, while relation pairs with different category combinations have scores closer to 0. In practice, the prototype and query set features are combined with a concatenation operator, and then the relation network generates a relation score for each query point. Two fully convolutional layers with loss operations are implemented in the relation learning module, fully connected layers, ReLU function, and batch normalization. The combined pairwise features are finally transformed into a distribution with a sigmoid activation function, and a mean squared error loss function is selected to regress the relation scores of the labels, enabling the network to learn.

[0052] The method will be further explained in detail below by applying it to a specific example. Assume the source dataset for meta-training is a densely matched point cloud of images from the Birmingham region of the Sensat Urban dataset, with a spatial coverage of 1.2 × 10⁻⁶. 6 m 2 The dataset includes 13 types of terrain features, totaling 569,147,075 points; target point cloud data includes commonly used ISPRS lidar point clouds, containing 9 types of terrain features, totaling 1,165,598 points; and DALES lidar point clouds, with a spatial coverage of 2×10. 6 m 2 It contains 8 land cover classes, 12,219,779 points, and densely matched point clouds of imagery from the Cambridge region of the Sensat Urban dataset, with a spatial coverage of 3.2 × 10⁻⁶.6 m 2 The dataset contains 12 land cover types and 2,278,514,725 points. For the preparation of small sample point cloud data, a sliding window of 50m × 50m with 10m intervals was first used to divide the four aerial point cloud datasets into blocks. For each block, 4096 points were randomly sampled.

[0053] 1) Constructing such Figure 1 The semantic segmentation network model shown is the MR-PFS network model.

[0054] 2) Perform meta-training on the source data ( Figure 1 (M-TA).

[0055] On the Birmingham point cloud as the source data, the MR-PFS network is first meta-trained using labeled samples in a supervised learning manner. For each training iteration, an N-way K-shot task is randomly selected, where N classes are randomly chosen each time to ensure task richness and improve model robustness, while ensuring no overlap between tasks. During the meta-training process for each task, the support points are explicitly labeled, while the labels for the query set are considered unknown.

[0056] Support set in each task and query set Simultaneously input into the MR-PFS network (such as...) Figure 3 As shown), firstly, deep features of two subsets are obtained through a point-structure-based feature extraction network (e.g., ...). Figure 4 (As shown); secondly, for each class of the support set, 100 points are randomly sampled as the initial prototypes for each class, and clustering is further performed, with the features of the cluster centers selected as the final multi-prototypes (e.g., ...). Figure 5 (as shown); and then the prototype features and query features The connections yield relation pairs, and for each pair of features, the relation network generates a relation score r. i,k for:

[0057]

[0058] In the formula, G represents the vector concatenation operation. φ (·) represents a relational network function containing two learning modules. Each module includes a fully connected layer, an activation function, and batch normalization. The combined pairwise features are transformed into a [0,1] distribution through the relational network.

[0059] Finally, the relation network is trained in a supervised manner based on whether the categories of the query set and the prototype are consistent. The mean squared error loss function is chosen to regress the relation scores of the labels, and all parameters θ of the MR-PFS network are optimized. Based on the differences between tasks, the model can predict new tasks and calculate the loss through multiple iterations. Acquiring learning ability:

[0060]

[0061]

[0062] In the formula, Let θ represent the loss, and let θ represent the parameters of the semantic segmentation network model. Let represent an N-way K-shot subtask containing a support set and a query set, where x represents the input point, y represents the true class label of the point, and α represents the learning rate during network training. f represents the gradient of the loss function. [θ] (·) represents the MR-PFS network function, and L represents the loss of the graph points.

[0063] 3) Fine-tuning on a small target dataset ( Figure 1 (FT in the text).

[0064] The model has already acquired learning capabilities after being trained on the source dataset for multiple tasks. When there are a small number of labeled samples in the target domain, the model is fine-tuned to make this meta-learning capability more adaptable to different target domains. Specifically, multiple different sub-tasks are generated for the small sample dataset of target point clouds in the same meta-training manner, and the training process is also supervised by multiple tasks. The only difference is that the initial training parameters of the model are selected from the parameters obtained by meta-training, so the fine-tuning process is more efficient than the meta-training process.

[0065] 4) Meta-testing on unlabeled target datasets ( Figure 1 After passing the M-TE test, the trained semantic segmentation network model can be obtained.

[0066] At this point, facing the new cross-domain scenario, meta-testing is performed on the unlabeled samples of the target dataset. The data is also divided according to the multi-task mode of meta-learning. Unlike training, samples are no longer randomly selected. Instead, each task is iterated sequentially to ensure that all unlabeled points can obtain their class labels through the network. Finally, the class labels of each point are obtained by voting on the test point set.

[0067] 5) Input the point cloud data to be classified into the trained semantic segmentation network model to obtain the corresponding classification results for each point cloud.

[0068] The effectiveness of the present invention will be verified and explained through the following simulation experiments.

[0069] 1) Simulation conditions. The hardware consists of an Intel Core i9-7900 CPU, an Nvidia GeForce RTX 3090 Ti GPU, and 128GB of RAM. The software uses the PyTorch library to implement the method of this invention.

[0070] 2) Parameter settings: In the MR-PFS network, a point embedding structure network is used to extract 320-dimensional features. 100 initial prototypes are selected through random sampling, and m=10 features from each cluster center are generated as multiple prototype features. For the pre-training process of feature learning, the Adam optimizer and a learning rate of 0.001 are used to train the data 200 times. For the meta-learning process, random sampling and meta-training are performed 3000 times.

[0071] 3) Simulation Results: The model was trained on the Birmingham source dataset and tested on the other three datasets. For each target dataset, 10% of the points were randomly selected as labeled samples for model fine-tuning, and the remaining 90% were used for testing. The mean Intersection over Union (mIoU) was used as the evaluation metric.

[0072] The multi-prototype relation learning network MR-PFS of this invention is compared with four meta-learning networks based on two-dimensional prototype network ProtoNet, two-dimensional relation network RelationNet, and the recently proposed attMPTI three-dimensional attention-based multi-prototype reasoning improvement network: Proto-3D, Proto-RL, attMPTI, and MPM-RL. Figures 7(a)-7(d) Figure 7(e) shows the experimental results of segmenting the CAM target dataset using the Proto-3D, attMPTI, Proto-RL, and MPM-RL methods, respectively; Figure 7(f) shows the experimental results of segmenting the CAM target dataset using the MR-PFS method of this invention; Figure 7(f) shows the standard segmentation results of the CAM target dataset. Figures 8(a)-8(f) For ISPRS target dataset and Figures 7(a)-7(f) Corresponding experimental results; Figures 9(a)-9(f) For the DALES target dataset and Figures 7(a)-7(f) The corresponding experimental results are shown in Table 1. Table 1 compares the final classification results of various methods. Figure 7(g) shows the classification results represented by different grayscale colors in Figures 7(a)-7(d); Figure 8(g) shows the classification results represented by different grayscale colors in Figures 8(a)-8(f); and Figure 9(g) shows the classification results represented by different grayscale colors in Figures 9(a)-9(f).

[0073] Table 1

[0074]

[0075] Experimental results show that: (1) Compared with existing meta-learning classification methods, the present invention has superior performance in classifying large-scale point clouds, especially point clouds in different scene target domains; (2) The present invention can achieve better small-sample point cloud classification results than other advanced meta-learning classification methods; (3) Compared with the above methods, the present invention effectively alleviates the mixed misclassification phenomenon in other methods in complex scenes, more finely and accurately classifies detailed targets, and well preserves the boundary information of ground objects.

[0076] In summary, this invention can quickly adapt to different target scenarios and achieve classification of new categories by fine-tuning the model on target datasets with only a small number of labeled samples. It has good generalization performance and can improve the overall classification accuracy compared to supervised deep learning methods. It can also improve the classification accuracy of imbalanced categories, which is of great significance in practical applications.

[0077] Specifically, the present invention has the following advantages:

[0078] (1) A novel small sample classification meta-learning method for large-scale, large-area point clouds was designed. When facing a completely new scene, only a small number of labeled samples are needed to achieve the classification of new categories. It is applicable to both lidar point clouds and photogrammetric dense matching point clouds, and has good generalization performance. Compared with supervised deep learning methods, it can improve the overall classification accuracy and improve the classification accuracy of imbalanced categories.

[0079] (2) A point cloud few-shot learning framework comprising three meta-learning stages was constructed. Through a task-based learning strategy, the designed network can learn how to transfer knowledge from the source domain to different target domains through three steps: meta-training, fine-tuning, and meta-testing. It can quickly adapt to unknown categories with only a small number of labeled samples from new datasets, without needing to retrain from scratch.

[0080] (3) A semantic segmentation network for small sample points cloud based on three-dimensional point structure MR-PFS is designed, including the construction of a multi-scale point cloud feature extraction network, which focuses on the geometric features of large-scale point clouds; the construction of a multi-prototype generation model, which can quickly model the complex feature distribution of labeled point cloud samples; and the construction of a learnable relational network to label unlabeled points in the depth metric space, which is more consistent with the depth feature distribution.

Claims

1. A method for semantic segmentation of point clouds with few samples, characterized in that, Includes the following steps: 1) The constructed semantic segmentation network model is trained using a meta-learning training strategy; the semantic segmentation network model includes a feature extraction network, a multi-prototype generation model, and a relation learning network; The feature extraction network is used to extract semantic features from the support set and query set respectively to obtain support set features and query set features. The semantic feature extraction method is as follows: hierarchical convolutions of different scales are used to extract different neighborhood geometric features of the input point cloud data, and these features are concatenated with input features of multiple dimensions to obtain the extracted semantic features. These features are then processed by two multilayer perceptrons for output. The input of the first multilayer perceptron is the semantic features, and the output of the first multilayer perceptron is the input of the second multilayer perceptron. The outputs of the first and second multilayer perceptrons are concatenated to obtain the final output. The first multilayer perceptron is a lightweight multilayer perceptron, and the second multilayer perceptron is a non-linear multilayer perceptron. The multi-prototype generation model is used to select prototype features from the support set features, and for each classification, feature selection is performed by first random sampling and then clustering. Relation learning networks are used to learn the similarity relationships between prototype features and query features; 2) Input the point cloud data to be classified into the trained semantic segmentation network model to obtain the classification results.

2. The small-sample point cloud semantic segmentation method according to claim 1, characterized in that, A meta-learning training strategy consisting of three sequentially executed meta-learning stages is adopted to train the semantic segmentation network model. The three meta-learning stages are as follows: the first stage is to perform meta-training on the semantic segmentation network model using source data. The second stage involves fine-tuning the semantic segmentation network model trained in the first stage using small sample target data; the third stage involves meta-testing using unlabeled target datasets.

3. The small-sample point cloud semantic segmentation method according to claim 1, characterized in that, In step 2), the point cloud data to be classified is the data after block processing, coordinate scale processing, and data augmentation processing.

4. The small-sample point cloud semantic segmentation method according to claim 1, characterized in that, The selected features are those that are located at the cluster centers after clustering.

5. The small-sample point cloud semantic segmentation method according to claim 1, characterized in that, Relation learning networks are used to learn similarity relationships between prototype features and query features by employing a learnable deep metric space: prototype features and query features are concatenated into relation pairs and fed into a two-layer neural network to learn the similarity relationship between them.

6. The small-sample point cloud semantic segmentation method according to any one of claims 1 to 5, characterized in that, The loss function used when training the semantic segmentation network model is: In the formula, Indicates loss, The parameters represent the semantic segmentation network model. This represents a subtask of an N-way K-shot that includes a support set and a query set. x Indicates the input point. y Indicates the true category label of the point. This represents the learning rate during network training. The gradient of the loss function is represented. This represents a semantic segmentation network model. This indicates the loss of a point on the graph.

7. The small-sample point cloud semantic segmentation method according to any one of claims 1 to 5, characterized in that, During the training of the semantic segmentation network model, N categories of tasks are randomly selected each time to train the semantic segmentation network model, and there is no overlap between the tasks.

Images