Small sample point cloud target classification method based on lshw-mamba network

Abstract

The small sample point cloud target classification method based on LSHW-Mamba network relates to the technical field of 3D point cloud processing and deep learning, and solves the problems that the existing technology focuses on feature mining in the Euclidean space domain, and under the condition of insufficient target samples, the feature representation ability of the target geometric structure is insufficient, and it is difficult to capture the key information for distinguishing similar objects. The present application explicitly analyzes the spectral response of the local geometric structure of the point cloud through the spherical harmonic function, combines the local spherical harmonic wavelet transform to realize the geometric feature extraction of the point cloud, simultaneously introduces the geometric gating mechanism to optimize the feature propagation efficiency of the Mamba model, models the target features in parallel with the Transformer attention branch, considers the relevance of the local and global features of the target, and improves the discrimination ability of the network model for the target under the small sample scene. The present application effectively solves the core problems of insufficient feature discrimination and decreased classification accuracy under the small sample condition.

Description

Technical Field

[0001] This invention relates to the fields of 3D point cloud processing and deep learning technology, specifically to a small sample point cloud target classification method based on Local Spherical Harmonic Wavelet (LSHW)-Mamba network. Background Technology

[0002] As a direct digital representation of the physical world, 3D point clouds play an increasingly important role in fields such as autonomous driving, robot navigation, and intelligent detection due to their precise ability to depict spatial geometric information. Compared with 2D images, point cloud data naturally exhibits inherent characteristics such as disorder, sparsity, and irregular distribution. In few-shot learning scenarios, the limited number of labeled samples combined with the complex characteristics of point clouds presents even more severe technical challenges for point cloud object classification. How to achieve efficient and accurate point cloud category classification using limited labeled samples has become a key problem that urgently needs to be solved in the field of 3D vision.

[0003] LiDAR is a core sensor in the field of environmental perception, capable of accurately capturing 3D structural information of a scene using pulsed light waves. It plays an indispensable role in point cloud target classification tasks in key areas such as autonomous driving, localization, and map building. However, in practical work, there are often small sample scenarios where target annotation data is scarce, leading to a significant decrease in the classification accuracy of traditional deep learning point cloud classification methods, which severely limits their deployment in complex real-world scenarios.

[0004] In small-sample (Few-Shot) scenarios, due to the sparsity of data and the unstructured nature of point clouds, network models struggle to extract robust features from a limited number of multi-class samples, easily leading to confused classifications of geometrically similar objects. Existing methods such as PointNet++, Transformer, and other mainstream approaches primarily focus on feature mining in the Euclidean space domain. When there are insufficient target samples, their ability to represent the geometric structure features of the target is inadequate, making it difficult to capture the key information for distinguishing similar objects.

[0005] To address the aforementioned issues, this invention proposes a point cloud target classification network based on local spherical harmonic wavelet transform and efficient spatially gated Mamba. By capturing target feature information in the frequency domain, it enhances the network model's ability to perceive and encode high-frequency details of target point clouds, improving feature discriminability and robustness. This significantly outperforms existing methods such as Transformer and self-supervised pre-training, alleviating the problem of insufficient target feature representation under small sample conditions. Summary of the Invention

[0006] To address the problem that existing technologies, which focus on feature mining in the Euclidean space domain, lack sufficient target samples, resulting in inadequate representation of target geometric features and difficulty in capturing key information to distinguish similar objects, this invention provides a small-sample point cloud target classification method based on the LSHW-Mamba network to alleviate the problem of insufficient target feature representation under small sample conditions.

[0007] A few-sample point cloud object classification method based on LSHW-Mamba network is implemented by the following steps:

[0008] Step 1: Group and standardize the point cloud, and represent the set of relative coordinates of local point clouds in local point cloud clusters in tensor form;

[0009] Step 2: Design a space-frequency domain dual feature encoding module to extract target geometric features from Euclidean space and spherical harmonic spectral domain respectively, and obtain feature sequences as input to the LSHW-Mamba encoder;

[0010] Step 3: Using stacked LSHW-Mamba encoders, global correlation modeling is achieved by leveraging efficient spatial gating and state space modeling to obtain the final feature sequence;

[0011] Step 4: Use the classification head to predict the category of the final feature sequence to achieve target classification of small sample point clouds.

[0012] The beneficial effects of this invention are:

[0013] 1. The small-sample point cloud target classification method described in this invention proposes a Space-Frequency Dual Feature Encoding (SFFE) mechanism. This overcomes the limitation of existing methods such as Transformer and self-supervised pre-training, which focus solely on Euclidean space features during feature extraction. By introducing Local Spherical Harmonic Wavelet (LSHW), the point cloud is projected into the spherical coordinate spectral domain, achieving complementary feature modeling in both the spatial and frequency domains. This mechanism can deeply mine the geometric spectral response features of the target point cloud, accurately capture subtle geometric differences between similar objects, significantly improve the robustness of the model in small-sample scenarios, effectively address the core problem of insufficient discriminative features, and outperform Transformer and self-supervised pre-training methods based on single spatial domain modeling.

[0014] 2. The small-sample point cloud target classification method described in this invention designs an efficient spatial geometric gating (ESGG) unit, specifically addressing the redundancy of the Transformer attention mechanism and the lack of 3D spatial structure perception in linear sequence models (including traditional Mamba variants). This unit dynamically modulates the Mamba state flow using the absolute coordinates of local group centers in the point cloud, reconstructing the model's ability to perceive the spatial structure of the 3D point cloud. Simultaneously, it integrates Transformer to complete local-global target association modeling, balancing the efficiency and accuracy of feature modeling and effectively reducing attention computation redundancy. Compared to existing self-supervised pre-training methods, it achieves better spatial feature alignment without relying on large amounts of unlabeled data, adapting to scenarios with scarce small-sample data and further enhancing the model's classification performance.

[0015] 3. To verify the effectiveness of the proposed model, multiple comparative experiments were conducted on the ModelNet40 standard dataset. The experimental results show that LSHW-Mamba achieves excellent performance in various few-sample point cloud classification tasks, and can provide technical support and theoretical reference for 3D object classification tasks in fields such as autonomous driving and robot perception. Attached Figure Description

[0016] Figure 1 This is a framework diagram of the small sample point cloud target classification method based on LSHW-Mamba network described in this invention;

[0017] Figure 2 A visual example of the ModelNet40 dataset. Detailed Implementation

[0018] Specific Implementation Method 1: Combination Figure 1 This implementation describes a small-sample point cloud target classification method based on the LSHW-Mamba network. This method utilizes a small-sample point cloud classification network based on spherical harmonic wavelets and efficient geometrically gated Mamba to construct a framework for local geometric spectrum perception and global feature modeling. By capturing target feature information in the frequency domain, this method enhances the network model's ability to perceive and encode high-frequency details of the target point cloud, improves feature discriminability and robustness, and alleviates the problem of insufficient target feature representation under small-sample conditions.

[0019] like Figure 1As shown, the point cloud target classification method in this embodiment first groups and standardizes the point cloud through Farthest Point Sampling (FPS) and K-Nearest Neighbor (KNN); then, a Space-Frequency Dual Feature Encoding (SFDFE) module is designed to extract target geometric features from Euclidean space and spherical harmonic spectrum domain, respectively; then, a stacked LSHW-Mamba encoder is used to achieve global association modeling by utilizing efficient spatial gating and state space model; finally, a classification head is used for category prediction. The specific implementation process is as follows:

[0020] Step 1. Group and standardize the point cloud based on the FPS algorithm and KNN; the specific process is as follows:

[0021] Step 1-1. Given the input point cloud First, the FPS algorithm is used to analyze the input point cloud. Selected from Each center point , As a variable hyperparameter, it is set to 256 from the perspective of network model operating efficiency. Batch size;

[0022] Step 1-2. For each center point ,exist Search for its recent Each neighboring point forms a local point set. ;

[0023] Steps 1-3. To enhance the translation invariance of features, local points are transformed into relative coordinates with respect to the center point. : The set of relative coordinates of all local points is a tensor. This tensor will serve as the input for subsequent space-frequency domain dual feature encoding.

[0024] Step 2. Design the Space-Frequency Domain Dual Feature Encoding (SFDFE) module to extract and fuse target geometric features from Euclidean space and spherical harmonic spectral domain respectively, enriching the target feature information; the specific process is as follows:

[0025] Step 2-1. Extract target spatial features by spatial domain branch;

[0026] This branch utilizes a self-attention mechanism to capture the Euclidean spatial dependencies between local point pairs. First, a multilayer perceptron (MLP) is used to process the relative coordinates of the point cloud. Convert to initial features This lays the foundation for subsequent feature processing:

[0027]

[0028] To enable the network model to explicitly perceive geometric structures, a Transformer layer was applied within each local point cloud group, and a factor based on relative position difference was injected when calculating the attention score. The geometric bias term, and Local points and Relative to the center point The relative coordinates are used to simulate nonlinearity through the softmax function:

[0029]

[0030] in, For position encoding functions, For query vector, Given the key vectors, first calculate the similarity between the two. The vector dimension was introduced to avoid excessively high similarity values ​​due to excessively high vector dimensions. After attention aggregation and residual connection, the updated features are obtained. Finally, max pooling is performed on the feature information of each point. This yields the target spatial features representing the local region. : .

[0031] Step 2-2. Extract target frequency domain features by using local spherical harmonic branching in the frequency domain;

[0032] By using spherical harmonic analysis to project local geometry into the spectral domain, low-frequency (overall target contour information) and high-frequency (target edge) information are separated.

[0033] First, the relative coordinates of the point cloud Convert to spherical coordinates Then calculate the position of each point. Response of spherical harmonic basis functions under first-order bandwidth Meanwhile, M learnable wavelet filters (including low-pass and band-pass filters) are defined. Each of them Finally, the response intensity of each point to these filters is obtained through projection calculations, thereby capturing the frequency domain geometric features:

[0034]

[0035]

[0036] In the formula, It is the response of the m-th learnable wavelet filter in the spherical harmonic basis function. Learnable weight parameters at the location, Let be the value of the m-th learnable wavelet filter at the k-th bandwidth. It is a multilayer perceptron used to process frequency band information, where K represents K neighboring points. It is the frequency domain geometric feature of the i-th center point, with dimension . .

[0037] Steps 2-3. Feature fusion;

[0038] The target spatial features obtained from the spatial domain branch and the target frequency domain features obtained from the frequency domain branch are concatenated and mapped to the dimension through a linear layer. This forms the initial state of the input sequence. Simultaneously, learnable category labels are introduced. This helps network models better focus on category-related features:

[0039]

[0040] Finally, the input sequence is formed. .

[0041] In this embodiment, the order Q of the spherical harmonic bandwidth determines the frequency domain resolution. Q=0 corresponds to a constant term (global DC component), and Q=1~4 corresponds to spatial fluctuation patterns at different frequencies. The higher the order, the better it can depict the subtle geometric undulations of the target surface. In this embodiment, setting the bandwidth to 4 can cover the low-frequency global contour (Q=0~1) and mid-to-high frequency detail features (Q=2~4) of the local area of ​​the point cloud. Among them, Q=2~3 corresponds to the curvature changes, edge transitions and other distinguishable structures of the target surface, and Q=4 can capture fine-grained geometric information such as tiny protrusions and grooves, which perfectly matches the core requirement of distinguishing similar objects in small sample scenarios.

[0042] The wavelet filter constructs a multi-band decomposition structure using one low-pass filter and two band-pass filters. The low-pass filter captures low-frequency components in the frequency domain, while the band-pass filters focus on different mid-to-high frequency ranges. The low-pass filter (corresponding to Q=0~1) extracts the overall shape contour of the local region, the first band-pass filter (corresponding to Q=2~3) captures the main structural edges and curvature changes of the target, and the second band-pass filter (corresponding to Q=3~4) captures subtle geometric features. The three filters form a hierarchical geometric information representation of the global contour, main structure, and fine-grained details.

[0043] The frequency domain feature size maps the response features of the three frequency bands into a 128-dimensional vector, achieving efficient compression and semantic aggregation of frequency domain information. The 128-dimensional feature vector is not simply stacked, but rather nonlinearly fuses the frequency domain responses of the target's global contour, main structure, and fine-grained details through MLP. Each dimension corresponds to a combined geometric spectrum pattern, enhancing the discriminative power of the features.

[0044] Step 3. Target features are efficiently progressively layered using stacked LSHW-Mamba encoders. Among these... As the input sequence of the first layer of the stacked LSHW-Mamba encoder, let the output of the (l-1)th layer be... , No. Layer computation involves efficient spatial geometry gating and global relational modeling using the Mamba state-space model; the specific process is as follows:

[0045] Step 3-1. Efficient spatial geometry gating;

[0046] Let the input features be First, set the center point of each local point set. Through linear layers (Linear), LayerNorm layers, and SiLU function mapping, the information is transformed into geometric context information that reflects spatial relationships. :

[0047]

[0048] Calculate the global max-pooling vector of the current input features, and summarize the global feature information to obtain... , It is a transform operator for learnable wavelet filters. It takes the maximum value of the response intensity of all wavelet filters, which is actually... ,in , It is the response of the spherical harmonic basis function.

[0049] Combined with the geometric context information obtained in the first step With current input features Generate dynamic weights that can adaptively adjust. This achieves the combination of geometric location and semantic features:

[0050]

[0051] Utilizing the generated dynamic weights For input features Modulation is performed to highlight the features of key spatial regions, resulting in modulated features: ,in A learnable scaling factor. Modulated features. With the category tag After updating, the sequence is reassembled to obtain the modulated sequence. .

[0052] Step 3-2. Mamba sequence modeling;

[0053] The modulated sequence The input bidirectional state space Bi-SSM module performs two scans on the sequence, one forward and one backward, to capture the dependencies between features. The results of the two scans are then summed, and finally, residual connections are used to preserve the original feature information and avoid gradient vanishing.

[0054]

[0055] Step 4. After feature enhancement by multiple LSHW-Mamba encoding modules ( Layer encoding yields a final feature sequence containing local-global correlations and complementary spatial-frequency information about the target. To fully utilize this information and obtain classification results, the specific process is as follows:

[0056] Step 4-1. Directly extract the features corresponding to the category labels from the final feature sequence. ;

[0057] Step 4-2. Perform max pooling on the final features of all local groups to highlight the most salient features of the target. ;

[0058] Step 4-3. Perform average pooling on the final features of all local groups to preserve the overall feature trend of the target. .

[0059] Step 4-4. Finally, these three types of features are concatenated and fused, mapped to the target class space through a linear layer, and the probability of each class is calculated using the Softmax function to complete the classification prediction. The formula is as follows:

[0060]

[0061] in It is a multilayer perceptron used for class prediction, and the probability is calculated by Softmax to obtain the final class probability distribution Y.

[0062] Specific Implementation Method Two: Combination Figure 2 This embodiment describes a verification example of a small-sample point cloud target classification method based on an LSHW-Mamba network as described in Specific Embodiment 1:

[0063] This implementation is based on the PyTorch framework and performed on an NVIDIA 4070TI Super GPU. During data preprocessing, the input point cloud is uniformly downsampled to 1024 points. To construct the local geometry, the number of center point samples is set to 256 (G=256), and the number of K-nearest neighbors for each local neighborhood is 16. The LSHW-Mamba encoder consists of 12 stacked layers (L=12), with the feature embedding dimension set to... The number of heads in the multi-head attention mechanism is set to 6. The wavelet feature dimension in the frequency domain branch is set to... To prevent overfitting, the dropout rate for the stochastic depth was set to 0.2. The AdamW optimizer was used.

[0064] To comprehensively verify the performance and advantages of the proposed method in point cloud classification, the LSHW-Mamba network model was trained and tested using the ModelNet40 dataset. Several mainstream methods based on different backbone networks were also selected for comparative experiments, as shown in Table 1. The visualization results of the ModelNet40 dataset are shown below. Figure 2 As shown.

[0065] Table 1

[0066]

[0067] In supervised learning scenarios without large-scale pre-training, LSHW-Mamba significantly outperforms mainstream methods such as PointNet, DGCNN, and Transformer in four tasks: 5-way 10-shot, 5-way 20-shot, 10-way 10-shot, and 10-way 20-shot. Specifically, it achieves an accuracy of 92.5% in the 5-way 10-shot task and 92.6% in the 10-way 20-shot task. With the introduction of self-supervised pre-training, its performance is further improved, with accuracy increases of 1.7% and 2.9% in the 5-way 20-shot and 10-way 20-shot tasks, respectively. It also outperforms Point-BERT across all test metrics and outperforms Point-MAE in more challenging high-sample-volume tasks. This is mainly due to its spatial-frequency dual feature encoding mechanism and efficient spatial gating sequence modeling capabilities, enabling it to extract more discriminative geometric features and stably capture the global spatial structure of the target in small-sample scenarios.

[0068] In addition, based on the experimental configuration of the ModelNet40 dataset (1024 input point clouds, 256 groups, 12 layers of LSHW-Mamba encoder), this invention is quantitatively compared with mainstream methods such as PointNet, Transformer, and Point-BERT. In terms of parameter scale, LSHW-Mamba has a total of approximately 18.7M parameters, which is 20.4% and 25.5% lower than Transformer (23.5M) and Point-BERT (25.1M), respectively. This is attributed to Mamba's linear sequence modeling and lightweight design with efficient spatial gating. In terms of computational complexity, using FLOPs as an indicator, the FLOPs of a single forward propagation in this invention are approximately 3.2G, which is 44.8% lower than Transformer (5.8G). This avoids the secondary complexity of the self-attention mechanism, and the frequency domain transformation only adds about 0.3G FLOPs, which is offset by feature complementarity. In terms of inference efficiency, on an NVIDIA 4070TI SuperGPU, the inference time of a single sample in this invention is approximately 12.6ms, which is 17.6% and 42.2% higher than PointNet (15.3ms) and Transformer (21.8ms), respectively. This meets the requirements of real-time applications and fully verifies the advantages of this invention in parameter efficiency and inference speed, laying the foundation for its deployment in practical scenarios such as autonomous driving and robot perception.

[0069] To verify the effectiveness of each core module in LSHW-Mamba, ablation experiments were conducted in 5-way and 10-way settings on ModelNet40. Using the base Mamba network with the SFDFE mechanism and ESGG module removed as a baseline, the results are shown in Table 2.

[0070] Table 2

[0071]

[0072] Introducing the SFDFE mechanism alone significantly improved the classification performance of the network model. In the 5-way 10-shot setting, accuracy increased from 89.2% to 93.0%, and in the 10-way 20-shot setting, it increased from 91.0% to 92.6%. This demonstrates that relying solely on features extracted from spatial coordinates is insufficient in small-sample scenarios. The SFDFE module, through local spherical harmonic wavelet transform, supplements the geometric details in the frequency domain, resulting in richer and more discriminative feature representations.

[0073] Introducing the ESGG module separately also improved model performance, achieving an accuracy of 91.6% in a 5-way 10-shot setting, 2.4 percentage points higher than the baseline model; and increasing from 92.5% to 93.6% in a 5-way 20-shot setting. This is because the ESGG module dynamically adjusts the target's feature flow using the absolute coordinates of local group centers, allowing the linear sequence model to regain the ability to perceive three-dimensional spatial structure, helping the model to more accurately locate the target's key geometric regions, thereby making accurate class classifications.

[0074] The model achieves optimal performance when both the SFDFE and ESGG modules are added simultaneously: accuracy rises to 96.1% in the 5-way 10-shot setting, an improvement of 6.9 percentage points over the baseline model; and in the 10-way 20-shot setting, accuracy increases from 91.0% to 95.6%, an improvement of 4.6 percentage points. This indicates that the two modules are complementary: the SFDFE mechanism is responsible for extracting high-quality target features in the local spatial and frequency domains, while the ESGG module guides the network at the global level to perform sequence modeling that better fits the logic of the 3D structure. Together, they enable the model to achieve further performance breakthroughs in few-shot classification tasks.

[0075] In addition, the SFDFE module demonstrates stable performance in low-shot, multi-way scenarios, achieving accuracies of 93.0% and 90.4% in 5-way 10-shot and 10-way 10-shot tasks, respectively. This is mainly due to the strong discriminative ability of frequency domain features for fine-grained geometric differences, making it suitable for demanding scenarios with many categories and scarce samples. The ESGG module shows even more significant performance improvements in multi-shot scenarios, with accuracies of 1.1 percentage points higher than the baseline in both 5-way 20-shot and 10-way 20-shot tasks. This is because its spatial gating mechanism can better utilize the structural information of sufficient samples and strengthen global correlation modeling. When the two modules are combined, the model maintains stable performance in all shot and way combinations, achieving accuracies of 96.1%, 96.5%, 92.8%, and 95.6% in 5-way 10-shot, 5-way 20-shot, 10-way 10-shot, and 10-way 20-shot tasks, respectively, covering a wide range of low-sample scenarios.

[0076] The above description is a specific illustration of the present invention, and not a limitation thereof. Those skilled in the art can make various equivalent technical solutions without departing from the scope of the present invention; therefore, all equivalent technical solutions should be included within the protection scope of the present invention.

[0077] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0078] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A few-sample point cloud object classification method based on LSHW-Mamba network, characterized by: This method is implemented by the following steps: Step 1: Group and standardize the point cloud, and represent the set of relative coordinates of local point clouds in local point cloud clusters in tensor form; Step 2: Design a space-frequency domain dual feature encoding module to extract target geometric features from Euclidean space and spherical harmonic spectral domain respectively, and obtain feature sequences as input to the LSHW-Mamba encoder; Step 3: Using stacked LSHW-Mamba encoders, global correlation modeling is achieved by leveraging efficient spatial gating and state space modeling to obtain the final feature sequence; Step 4: Use the classification head to predict the category of the final feature sequence to achieve target classification of small sample point clouds.

2. The small-sample point cloud target classification method based on LSHW-Mamba network according to claim 1, characterized in that: The specific process of step one is as follows: Step 11: Use the FPS algorithm to analyze the input point cloud. Selected from Each center point , B is a variable hyperparameter, and B is the batch size. Steps 1 and 2: For each center point ,exist Search for its recent Each neighboring point forms a local point set; Step 13: Concentrate the local point cloud Transformed relative to the center point relative coordinates The set of relative coordinates of all local point clouds is a tensor. .

3. The small-sample point cloud target classification method based on LSHW-Mamba network according to claim 1, characterized in that: The specific process of step two is as follows: Step 2: Extract target spatial features from Euclidean space using the spatial domain branch of the space-frequency domain dual feature encoding module; Step 22: Extract the target frequency domain features from the spherical harmonic spectrum domain using the frequency domain local spherical harmonic small ball branch of the space-frequency domain dual feature coding module; Step 23: Fuse the spatial features extracted in Step 21 and the frequency domain features extracted in Step 22 to obtain the final input sequence.

4. The small-sample point cloud target classification method based on LSHW-Mamba network according to claim 3: characterized in that: In step two, the specific process of extracting target space features from Euclidean space is as follows: The relative coordinates of the point cloud are converted into initial features by MLP. A Transformer layer is used in each local point cloud group, and a geometric bias term based on the relative position difference is injected when calculating the attention score. After attention aggregation and residual connection, the updated features are obtained. Finally, max pooling is performed on the updated features of each point cloud to obtain the target spatial features of the local region. .

5. The small-sample point cloud target classification method based on LSHW-Mamba network according to claim 4: characterized in that: In step two, the specific process of extracting the target frequency domain features from the spherical harmonic spectral domain is as follows: Convert the relative coordinates of the point cloud to spherical coordinates; calculate the coordinates of each point cloud element. Response of spherical harmonic basis functions under first-order bandwidth Meanwhile, define M learnable wavelet filters. Finally, the response intensity of each point to the wavelet filter is obtained through projection calculation; Extract target frequency domain features based on the response intensity. It can be expressed as follows: ； ； In the formula, For the m-th learnable wavelet filter in the spherical harmonic basis function response Learnable weight parameters at the location, For a multilayer perceptron used to process frequency band information, K represents the K neighboring points. Let be the value of the m-th learnable wavelet filter at the k-th bandwidth. For dimensions.

6. The small-sample point cloud target classification method based on LSHW-Mamba network according to claim 5: In steps two and three, the target spatial features obtained in step two-one and the target frequency domain features obtained in step two-two are concatenated and mapped to the dimension through a linear layer. This forms the initial state of the input sequence; simultaneously, learnable category labels are introduced. The final input sequence is obtained, expressed as follows: ； In the formula, and These are the target spatial features and the target frequency domain features, respectively. Final input sequence for: .

7. The small-sample point cloud target classification method based on LSHW-Mamba network according to claim 6: the specific process of step three is as follows: Step 31: High-efficiency spatial geometry gating; Let the input features be First, set the center point of each local point set. Through linear layers, LayerNorm layers, and SiLU function mapping, it is transformed into geometric context information that reflects spatial positional relationships. It can be expressed as follows: ； Calculate the global max-pooling vector of the current input features, and summarize the global feature information to obtain... , For the transform operator of learnable wavelet filters, To maximize the response intensity for all wavelet filters; Based on the geometric context information With current input features Generate dynamic weights that can adaptively adjust. This achieves the combination of geometric location and semantic features: ； Using generated dynamic weights Local group features Modulate, and then with the category label By splicing, the modulated features are obtained: ,in, The learnable scaling factor; the modulated features With category tags After updating, the sequences are reassembled to generate the modulated sequence. ; Step 3.2: Mamba sequence modeling; The sequence modulated in step three-one The bidirectional state-space model is input and scanned twice, in both the forward and reverse directions, to capture the dependencies between the features before and after the scan. The results of the two scans are then summed, and finally, the original feature information is preserved through residual connection to obtain the final feature sequence. 。 8. The few-sample point cloud target classification method based on LSHW-Mamba network according to claim 7, characterized in that: The specific process of step four is as follows: Step 41: Extract the features corresponding to the category labels from the final feature sequence. ; Step 4.2: Perform max pooling on the final features of all local groups to obtain the most salient feature information of the target. ; Step 43: Perform average pooling on the final features of all local groups to preserve the overall trend features of the target. ; Step 44, and finally steps 41, 42, and 43, feature concatenation and fusion are performed. The result is mapped to the target class space through a linear layer, and the probability of each class is calculated using the Softmax function to complete the classification prediction. The formula is as follows: ； In the formula, For a multilayer perceptron used for class prediction, the final class probability distribution is obtained by calculating the probabilities using Softmax.

Images