A point cloud retrieval method based on skeleton enhanced structure perception

Abstract

The application discloses a point cloud retrieval method based on skeleton enhanced structure perception, and belongs to the field of image data processing. The application provides an accurate and efficient solution for three-dimensional model library retrieval, robot target recognition and other tasks. The method first extracts a point cloud skeleton as a structure prior, extracts multi-scale features of the point cloud and the skeleton through a double-branch network, uses a double-branch fusion module, combines multi-scale cross-attention and contrast learning tasks, and explicitly and implicitly enhances the structure perception ability of the point cloud features. Finally, through a multi-scale adaptive local aggregation module, personalized clustering centers are dynamically generated, and discriminative global descriptors are aggregated for retrieval. The application effectively improves the accuracy and robustness of point cloud retrieval.

Description

Technical Field

[0001] This invention belongs to the general field of image data processing or generation, and specifically relates to a point cloud retrieval method based on skeleton-enhanced structure perception. Background Technology

[0002] Point cloud retrieval, a fundamental task in 3D vision, aims to quickly and accurately retrieve models similar to the query model from large-scale point cloud databases. Point cloud data is characterized by its unordered, discrete, and unstructured nature; therefore, constructing discriminative global descriptors from it is the core challenge.

[0003] Point cloud retrieval methods based on deep learning can be mainly divided into two categories: feature learning-based methods and multimodal fusion-based methods.

[0004] Feature-based learning methods, such as PointNet++, construct global descriptors by aggregating local features through hierarchical networks. The structure-awareness of these methods stems from indirect inference of the spatial distribution of point cloud surfaces, a "from the surface to the core" approach. However, this method is prone to losing crucial structural details during feature aggregation, making it difficult to accurately distinguish between objects with similar structures but different surface contours.

[0005] Multimodal fusion-based methods, such as ULIP, attempt to incorporate multimodal information, including images and text, to enhance point cloud feature representations. However, these methods typically rely on complex cross-modal alignment, resulting in high computational complexity, and the introduced information may not fully match the needs of 3D structural understanding.

[0006] Some studies consider explicitly utilizing point cloud skeletons to introduce prior structural information. Point cloud skeletons are compact shape representations that effectively capture the intrinsic topological structure of objects. Existing methods, such as MorphoSkel3D, often use skeletons for downsampling guidance in the preprocessing stage, indirectly guiding the model to focus on important structures by retaining more sampling points near key points in the skeleton. However, such methods fail to achieve collaborative fusion of structural information from the skeleton and point cloud at the deep feature level, limiting the potential of skeleton information in improving the discriminative power of point cloud features.

[0007] In summary, existing technologies have the following shortcomings: (1) traditional feature learning methods are insufficient in perceiving intrinsic structural information; (2) multimodal fusion methods are complex and do not directly utilize structural information; (3) the potential of the skeleton as prior information has not been fully explored and fused at the feature level. Therefore, how to effectively utilize the explicit structural prior of point cloud skeletons to enhance the structural perception capability of point cloud features in a combination of explicit and implicit methods at the deep feature level, and generate more discriminative global descriptors, has become an urgent problem to be solved. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention proposes a point cloud retrieval method with skeleton-enhanced structure perception. This method solves the technical problems of traditional point cloud retrieval methods, such as loss of detail and insufficient discriminative power due to indirect structure inference; existing skeleton information utilization methods mostly remain at the preprocessing level and fail to achieve deep feature fusion; and multimodal fusion methods are highly complex and do not directly utilize structural information.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a point cloud retrieval method based on skeleton-enhanced structure perception, comprising the following steps: Step 1, inputting 3D point cloud data and extracting its skeleton structure, and extracting multi-scale geometric features of the point cloud and multi-scale structural features of the skeleton in parallel through a dual-branch network; Step 2, constructing a dual-branch feature fusion module, using the skeleton structural features extracted in Step 1 as prior guiding information, and achieving explicit fusion of point cloud features and skeleton features through a multi-scale cross-attention mechanism, and introducing a contrastive learning task to implicitly constrain the network to enhance the structural perception capability of point cloud features; Step 3, dynamically aggregating the multi-scale point cloud features enhanced in Step 2 through a multi-scale adaptive local aggregation module, generating a highly discriminative global descriptor, and performing point cloud retrieval based on the global descriptor; Step 4, during model training, constructing a dynamic total loss function composed of a weighted average of classification loss, triplet loss, and contrastive loss, and supervising network training in stages.

[0010] Preferably, step 1 specifically includes: step 1.1, inputting a 3D point cloud. The corresponding skeleton point cloud was generated using a morphological skeleton extraction method. Step 1.2: Construct an improved PointNet++ network as the feature extraction branch, and introduce a self-attention layer into its Set Abstraction module to enhance the network's long-range context awareness during feature extraction; Step 1.3: Process the 3D point cloud... With the skeleton point cloud Two identical but parameter-independent improved PointNet++ networks are input separately to extract multi-scale point cloud features. Multi-scale skeleton features .

[0011] Preferably, step 2 specifically includes: Step 2.1, constructing a multi-scale cross-attention fusion module, using point cloud features as queries and skeleton features as keys and values ​​to achieve explicit fusion at the feature level; Step 2.2, constructing a contrastive learning task, implicitly guiding the network to learn structural consistency information through feature alignment between "cropped point cloud + skeleton" and "complete point cloud + skeleton"; Step 2.3, using a feature propagation module to enhance features at multiple scales in a top-down manner. Upsampling and fusion are performed to output refined enhanced features. and .

[0012] Preferably, step 2.1 specifically includes the following steps: for each feature scale Based on point cloud features As a query Based on skeletal features As a key AND value Point cloud features enhanced with skeleton information are obtained through cross-attention calculation. The calculation formula is shown in formula (1); (1); where the query matrix Key matrix Value matrix , express The transpose of the matrix, , , For learnable projection matrices, Using the scaling factor, Softmax(·) converts the scaled dot product similarity matrix into a probability distribution. To enhance features.

[0013] Preferably, step 2.2 specifically includes the following steps: Step 2.2.1, processing the original point cloud. Random cropping yields cropped point clouds. Step 2.2.2, will and Input the corresponding features into the network to obtain their respective feature representations. and Step 2.2.3: Construct a contrastive loss based on the InfoNCE loss function. This implicitly drives the network to learn structural consistency; (2); Where, sim( ) represents the cosine similarity, used to calculate the similarity between two feature vectors, with a value range of [−1,1], τ is the temperature hyperparameter, and N is the number of negative samples in the batch. It is a point cloud P cropping crop The feature representation extracted by the network, Fp, is the feature representation of point cloud P extracted by the network, and... Constitute a positive sample. This represents the feature representation of the j-th negative sample. The cosine similarity represents the similarity between positive sample pairs. This represents the cosine similarity between the anchor sample and the j-th negative sample. This represents the score after exponentially scaling the similarity of positive samples. This represents the score after exponentially scaling the similarity of each negative sample. This represents the sum of similarities with all negative samples, and -log(·) indicates that the negative log-likelihood encourages the maximization of similarity between positive samples and negative samples, while suppressing the similarity with negative samples.

[0014] Preferably, step 2.3 specifically includes the following steps: Step 2.3.1, merging high-level features Upsampling to The scale, and with Channel stitching is performed, followed by feature fusion via a multilayer perceptron. ; (3); among which, Indicates an upsampling operation. This indicates a channel splicing operation. Represents a shared multilayer perceptron; step 2.3.2, will Upsampling to The scale, and with Channel stitching is performed, followed by feature fusion via a multilayer perceptron. ; (4).

[0015] Preferably, step 3 specifically includes the following steps: Step 3.1, constructing a local adaptive aggregation descriptor submodule, and dynamically generating personalized clustering centers adapted to the current input sample through iterative optimization using a multi-layer Transformer decoder. Step 3.2, based on the final cluster centers For features at multiple scales Calculate the corresponding soft assignment weights for each feature point, and then use these weights to aggregate the residuals between each feature point and the personalized cluster centers to generate descriptors for the corresponding scale. Step 3.3 involves concatenating and fusing the multi-scale descriptors to obtain the initial global descriptor. and through a gating calibration mechanism The channels are adaptively enhanced to output the final global descriptor. .

[0016] Preferably, the core of the local adaptive aggregation descriptor submodule in step 3.1 lies in adapting the cluster centers to the feature distribution of the input samples through an iterative attention mechanism; specifically, it is implemented as follows: Step 3.1.1, let the set of general cluster centers before the update be... It contains k cluster centers, and the input features are ,by As a query (0) Input features as a bond matrix Sum matrix , No. The next update formula is: (5); (6); (7); among which, These are the learnable projection matrices corresponding to the query matrix Q, the key matrix K, and the value matrix V, respectively. The attention weights are obtained; in step 3.1.2, according to the update strategy in step 3.1.1, the personalized cluster centers are obtained after a total of M iterations of updates. Step 3.1.3: Use the momentum update strategy to merge the general cluster centers and the personalized cluster centers to obtain the final cluster centers; (8); among which, The momentum coefficient, As a general cluster center, As a personalized clustering center, It serves as the final cluster center.

[0017] Preferably, in step 3.2, the descriptor Calculated using the following formula: (9); among which, The final set of cluster centers contains k cluster centers. The weight for assigning the i-th feature point to the k-th cluster center. Let i be the local feature descriptor of the i-th point. To perform residual calculations on all points The weighted aggregation; concat(·) means concatenating the vectors corresponding to the k cluster centers.

[0018] The gating calibration mechanism in step 3.3 is calculated using the following formula: (10); among which, It is the Sigmoid activation function. and For learnable parameters, This indicates element-wise multiplication. To concatenate the three scale descriptors to obtain the global descriptor, F output To enhance the global feature descriptor of the output. Preferably, during model training, step 4 uses the total loss function. Classification loss Triple loss And comparative loss The weighted summation consists of: (11); among which, and The weights are dynamically adjusted during training: They are set at the beginning of training. , =0.3 to enhance structure awareness, training settings in the later stages. , =0.8 to enhance feature discrimination.

[0019] The beneficial technical effects brought about by this invention are as follows: 1. Enhanced structural perception ability: By introducing point cloud skeletons as explicit structural priors and designing a dual mechanism of explicit fusion (cross-attention) and implicit guidance (contrastive learning), the network can deeply understand and utilize the internal topological structure of objects, which significantly enhances the structural perception and discrimination ability of point cloud features.

[0020] 2. Feature extraction network improvement: A self-attention layer is introduced into the Set Abstraction module of PointNet++, which enhances the network's ability to model long-range context during local feature extraction.

[0021] 3. Improved discriminativeness of descriptors: A multi-scale local adaptive aggregation module was designed. The Transformer decoder dynamically generates personalized cluster centers that adapt to different samples, replacing the traditional fixed cluster centers. This makes the aggregated global descriptors more sample-adaptive and discriminative.

[0022] 4. Excellent retrieval performance: Experiments on standard datasets such as ModelNet40 and ShapeNet show that the method of this invention reaches or surpasses the current advanced level in retrieval accuracy, verifying its effectiveness. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall process of the skeleton-enhanced structure-aware network (SENet) proposed in this invention.

[0024] Figure 2 This is a schematic diagram of the structure of the Local Adaptive Aggregation Descriptor (VLAAD), the core submodule of the Multi-Scale Adaptive Local Aggregation Module (MVLAAD) in this invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0026] A point cloud retrieval method based on skeleton-enhanced structure perception includes the following steps: Step 1, inputting 3D point cloud data and extracting its skeleton structure, using a dual-branch network to extract multi-scale geometric features of the point cloud and multi-scale structural features of the skeleton in parallel. Specifically, this includes: Step 1.1, skeleton extraction, inputting 3D point cloud data... The corresponding skeleton point cloud was generated using the morphological skeleton extraction method (MorphoSkel3D). As prior information to enhance structure awareness; Step 1.2, construct an improved PointNet++ network as a feature extraction branch, and introduce a self-attention layer in its Set Abstraction module to enhance the network's long-range context awareness capability during feature extraction; Step 1.3, point cloud feature extraction, the point cloud... With the skeleton point cloud Two identical but parameter-independent improved PointNet++ networks are input separately to extract multi-scale point cloud features. Skeleton feature extraction: extracting skeleton point clouds An improved PointNet++ network with the same branch structure as the point cloud but independent weights is used to extract skeleton features at the corresponding scale. This provides structural guidance signals for subsequent fusion.

[0027] Step 2: Construct a dual-branch feature fusion module. Utilizing the skeleton structure features extracted in Step 1 as prior guiding information, an explicit fusion of point cloud features and skeleton features is achieved through a multi-scale cross-attention mechanism. A contrastive learning task is introduced to implicitly constrain the network, enhancing the structural awareness capability of point cloud features. Specifically, this includes: Step 2.1: Construct a multi-scale cross-attention fusion module, using point cloud features as queries and skeleton features as keys and values, to achieve explicit feature-level fusion. This includes the following steps: For each feature scale... Based on point cloud features As a query Based on skeletal features As a key AND value Point cloud features enhanced with skeleton information are obtained through cross-attention calculation. The calculation formula is shown in formula (1); (1); Among them, the query matrix Key matrix Value matrix , express The transpose of the matrix, , , For learnable projection matrices, Using the scaling factor, Softmax(·) converts the scaled dot product similarity matrix into a probability distribution. To enhance features.

[0028] Step 2.2, construct a contrastive learning task. By aligning features between the "cropped point cloud + skeleton" and the "complete point cloud + skeleton", the network is implicitly guided to learn structural consistency information. Specifically, this includes the following steps: Step 2.2.1, for the original point cloud... Random cropping yields cropped point clouds. ; Step 2.2.2, will and Input the corresponding features into the network to obtain their respective feature representations. and ; Step 2.2.3: Construct a contrastive loss based on the InfoNCE loss function. This implicitly drives the network to learn structural consistency; (2); Where, sim( ) represents the cosine similarity, used to calculate the similarity between two feature vectors, with a value range of [−1,1], τ is the temperature hyperparameter, and N is the number of negative samples in the batch. It is a point cloud P cropping crop The feature representation extracted by the network, Fp, is the feature representation of point cloud P extracted by the network, and... Constitute a positive sample. This represents the feature representation of the j-th negative sample. The cosine similarity represents the similarity between positive sample pairs. This represents the cosine similarity between the anchor sample and the j-th negative sample. This represents the score after exponentially scaling the similarity of positive samples. This represents the score after exponentially scaling the similarity of each negative sample. This represents the sum of similarities with all negative samples, and -log(·) indicates that the negative log-likelihood encourages the maximization of similarity between positive samples and negative samples, while suppressing the similarity with negative samples.

[0029] Step 2.3: Enhance multi-scale features in a top-down manner through the feature propagation module. Upsampling and fusion are performed to output refined enhanced features. and .

[0030] Specifically, it includes the following steps: Step 2.3.1, integrating high-level features. Upsampling to The scale, and with Channel stitching is performed, followed by feature fusion via a multilayer perceptron. : (3); among which, Indicates an upsampling operation. This indicates a channel splicing operation. Represents a shared multilayer perceptron; Step 2.3.2, will Upsampling to The scale, and with Channel stitching is performed, followed by feature fusion via a multilayer perceptron. : (4).

[0031] Step 3: Dynamically aggregate the multi-scale point cloud features enhanced in Step 2 using the multi-scale adaptive local aggregation module to generate a highly discriminative global descriptor, and perform point cloud retrieval based on the global descriptor; specifically including the following steps: Step 3.1: Construct a Local Adaptive Aggregation Descriptor Submodule (VLAAD), and dynamically generate personalized clustering centers adapted to the current input sample through iterative optimization using a multi-layer Transformer decoder. .

[0032] The core of the Local Adaptive Aggregation Descriptor Submodule lies in adapting the cluster centers to the feature distribution of the input samples through an iterative attention mechanism.

[0033] The specific implementation is as follows: Step 3.1.1, let the set of general cluster centers before the update be... It contains k cluster centers, and the input features are ,by As a query (0) Input features as a bond matrix Sum matrix . No. The next update formula is: (5); (6); (7); among which, These are the learnable projection matrices corresponding to the query matrix Q, the key matrix K, and the value matrix V, respectively. The resulting attention weights; Step 3.1.2: Following the update strategy in Step 3.1.1, the personalized cluster centers are obtained after a total of M iterations. ; Step 3.1.3: Using a momentum update strategy, the general cluster centers and personalized cluster centers are fused to obtain the final cluster centers. (8); among which, The momentum coefficient, As a general cluster center, As a personalized clustering center, It serves as the final cluster center.

[0034] Step 3.2, based on the final cluster centers For features at multiple scales Calculate the corresponding soft assignment weights for each feature point, and then use these weights to aggregate the residuals between each feature point and the personalized cluster centers to generate descriptors for the corresponding scale. ; descriptor Calculated using the following formula: (9); in, The final set of cluster centers contains k cluster centers. The weight for assigning the i-th feature point to the k-th cluster center. Let i be the local feature descriptor of the i-th point. To perform residual calculations on all points The weighted aggregation; concat(·) means concatenating the vectors corresponding to the k cluster centers.

[0035] Step 3.3 involves concatenating and fusing the multi-scale descriptors to obtain the initial global descriptor. and through a gating calibration mechanism The channels are adaptively enhanced to output the final global descriptor. .

[0036] The gating calibration mechanism is calculated using the following formula: (10); among which, It is the Sigmoid activation function. and For learnable parameters, This indicates element-wise multiplication. To concatenate the three scale descriptors to obtain the global descriptor, F output To enhance the global feature descriptor of the output.

[0037] Step 4: During model training, construct a dynamic total loss function consisting of a weighted average of classification loss, triplet loss, and contrastive loss, and supervise network training in stages.

[0038] Total loss function used Classification loss Triple loss And comparative loss The weighted summation consists of: (11); among which, and The weights are dynamically adjusted during training: They are set at the beginning of training. , =0.3 to enhance structure awareness, training settings in the later stages. , =0.8 to enhance feature discrimination.

[0039] This invention proposes a point cloud retrieval method based on skeleton-enhanced structure awareness, the overall network architecture of which is as follows: Figure 1 As shown, it mainly includes three core stages: multi-scale feature extraction, dual-branch feature fusion and guidance, and multi-scale adaptive local aggregation and retrieval.

[0040] Phase 1: Multi-scale feature extraction (corresponding to step 1).

[0041] The input is a 3D point cloud containing 1024 points. First, the MorphoSkel3D algorithm was used to extract its skeleton point cloud. Number of skeleton points Typically much smaller Subsequently, and The data are fed into a dual-branch improved PointNet++ network. Each branch contains three downsampling stages (Set Abstraction layers). The improvement of this invention lies in integrating a self-attention module within each SA layer, enabling the network to consider long-range inter-point relationships within the local region while aggregating local neighborhood information, thereby extracting features with stronger context awareness. Finally, the point cloud branch outputs the features. Skeleton branch output features .

[0042] Phase 2: Dual-branch feature fusion and structure-aware guidance (corresponding to step 2).

[0043] like Figure 1 As shown in the middle, and Cross-attention fusion is performed at three scales (Equation (1)) to obtain enhanced features. Meanwhile, during training, a cropped point cloud will be constructed. and compare it with the complete skeleton. The input is fed into the network for forward propagation, and the resulting features are calculated. Features of the complete input pair Calculate the comparative loss together (Formula (2)). This loss serves as an implicit supervision signal, does not participate in forward inference, and is only used to optimize network parameters during the training phase. Finally, The refined features are obtained by fusing them through the feature propagation module (formulas (3)(4)). and .

[0044] Phase 3: Adaptive aggregation and descriptor generation (corresponding to step 3).

[0045] The core of this stage is the Multi-Scale Adaptive Local Aggregation Module (MVLAAD), whose internal VLAAD sub-module structure is as follows: Figure 2 As shown.

[0046] Dynamic clustering: Setting initial general cluster centers (For example, K=64 centers, dimension D=256). With refined high-level features As input .like Figure 2 As shown, As a query sequence As a key-value sequence, it undergoes iterative optimization through a two-layer Transformer decoder (M=2). The calculations for each layer are shown in equations (5) to (7). After two layers of decoding, the result is... , and then with Momentum fusion is performed according to formula (8) (assuming α=0.5), resulting in... .

[0047] Multiscale aggregation: using To each , and Soft-assigned weights are calculated for the three feature maps. For each feature point... Calculate its relationship with the cluster center The similarity is then normalized to weights using softmax. Then aggregate the residuals according to formula (9). Generate descriptors for the corresponding scale. .

[0048] Gated fusion: The concatenation is then compressed to 512 dimensions through a linear layer, resulting in... Then, the final 512-dimensional global descriptor is obtained by calibrating through the gating mechanism of formula (10). .

[0049] Training and retrieval (corresponding to step 4).

[0050] The model uses the Adam optimizer with an initial learning rate of 0.001, multiplied by 0.7 every 30 epochs. The total loss function is shown in formula (11). The dynamic weight adjustment strategy is set as follows: for the first 100 epochs, , ; From epoch 100 to 200, linear adjustment to , Then it remained unchanged, with a total of 250 epochs of training.

[0051] During retrieval, the point cloud to be queried is input into the trained network to obtain its... Descriptor. Calculate the Euclidean distance between the descriptor and all pre-calculated descriptors in the database, sort them in ascending order of distance, and return the top K most similar models as the search results.

[0052] To verify the effectiveness of this invention, comparative experiments were conducted on the ModelNet40, ShapeNetPart, and ShapeNet datasets.

[0053] ModelNet40 contains 12,311 3D CAD models across 40 categories, with 9,843 samples in the training set and 2,468 samples in the test set. It is the standard benchmark in the field of point cloud retrieval.

[0054] ShapeNet contains approximately 50,000 models across 55 categories, offering greater retrieval challenges due to its wider range of categories and more diverse shapes.

[0055] ShapeNetPart is a subset of ShapeNet, containing 8887 models across 8 categories, with 6013 models in the training set and 2874 models in the test set.

[0056] The evaluation metrics used were mean average precision (mAP) and top 10 precision (P@10).

[0057] The experimental setup was standardized: 1024 points were input into a point cloud, data augmentation was performed using random rotation and translation, the skeleton was extracted using MorphoSkel3D, the optimizer was Adam, the initial learning rate was 0.001, and the training lasted for 250 epochs.

[0058] The model of this invention (denoted as SENet) is compared with more than ten advanced methods, including PointNet, PointNet++, DGCNN, PCT, CF3D, and *MorphoSkel3D (using only skeleton-guided downsampling). The main comparison results are shown in Table 1.

[0059] Table 1 shows the performance comparison of different methods on the point cloud retrieval dataset; .

[0060] Experimental results show that: 1. The overall performance is excellent; the SENet model proposed in this invention achieves the best or near-best retrieval performance on both ShapeNetPart and ShapeNet datasets, and its performance on ModelNet40 is comparable to that of state-of-the-art methods, proving its effectiveness and generalization ability.

[0061] 2. Effectiveness of the skeleton-based deep feature fusion strategy; By comparing *MorphoSkel3D (using only the skeleton for downsampling) and SENet (this invention, performing deep feature fusion), it can be found that on the more complex and challenging ShapeNetPart dataset, the advantages of this invention are significant (mAP: 98.9% vs 97.24%, P@10: 99.4% vs 98.79%). This strongly demonstrates that the strategy proposed in this invention, which uses the skeleton as a prior for explicit and implicit deep feature fusion, can more effectively enhance the network's perception and understanding of complex point cloud structures than using only the skeleton for preprocessing.

[0062] In summary, this invention significantly improves the accuracy of point cloud retrieval through an innovative skeleton-enhanced structure perception mechanism and an adaptive feature aggregation strategy.

[0063] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A point cloud retrieval method based on skeleton-enhanced structure perception, characterized in that, Includes the following steps: Step 1: Input 3D point cloud data and extract its skeleton structure. Extract multi-scale geometric features of the point cloud and multi-scale structural features of the skeleton in parallel through a dual-branch network. Step 2: Construct a dual-branch feature fusion module. Use the skeleton structure features extracted in Step 1 as prior guiding information. A multi-scale cross-attention mechanism is used to achieve explicit fusion of point cloud features and skeleton features. A contrastive learning task is introduced to implicitly constrain the network to enhance the structural perception ability of point cloud features. Step 3: Through the multi-scale adaptive local aggregation module, the multi-scale point cloud features enhanced in Step 2 are dynamically aggregated to generate a highly discriminative global descriptor, and point cloud retrieval is performed based on the global descriptor; Step 4: During model training, construct a dynamic total loss function consisting of a weighted average of classification loss, triplet loss, and contrastive loss, and supervise network training in stages.

2. The point cloud retrieval method according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1, Input 3D point cloud The corresponding skeleton point cloud was generated using a morphological skeleton extraction method. ; Step 1.2: Construct an improved PointNet++ network as the feature extraction branch, and introduce a self-attention layer in its Set Abstraction module to enhance the network's long-range context awareness during the feature extraction process. Step 1.3, convert the 3D point cloud With skeleton point cloud By inputting two modified PointNet++ networks with identical structures but independent parameters, multi-scale point cloud features are extracted. Multi-scale skeleton features .

3. The point cloud retrieval method according to claim 2, characterized in that, Step 2 specifically includes: Step 2.1: Construct a multi-scale cross-attention fusion module, using point cloud features as queries and skeleton features as keys and values, to achieve explicit fusion at the feature level; Step 2.2: Construct a contrastive learning task. By aligning the features between "cropped point cloud + skeleton" and "complete point cloud + skeleton", the network is implicitly guided to learn structural consistency information. Step 2.3: Enhance multi-scale features in a top-down manner through the feature propagation module. Upsampling and fusion are performed to output refined enhanced features. and .

4. The point cloud retrieval method according to claim 3, characterized in that, Step 2.1 specifically includes the following steps: For each feature scale Using point cloud features As a query Based on skeletal features As a key AND value Point cloud features enhanced with skeleton information are obtained through cross-attention calculation. The calculation formula is shown in formula (1); （1）； Among them, the query matrix Key matrix Value matrix , express The transpose of the matrix, , , For learnable projection matrices, Using the scaling factor, Softmax(·) converts the scaled dot product similarity matrix into a probability distribution. To enhance features.

5. The point cloud retrieval method according to claim 4, characterized in that, Step 2.2 specifically includes the following steps: Step 2.2.1, process the original point cloud Random cropping yields cropped point clouds. ; Step 2.2.2, will and Input each feature into the network to obtain its corresponding feature representation. and ; Step 2.2.3: Construct a contrastive loss based on the InfoNCE loss function. This implicitly drives the network to learn structural consistency; （2）； Where, sim( ) represents the cosine similarity, used to calculate the similarity between two feature vectors, with a value range of [−1,1], τ is the temperature hyperparameter, and N is the number of negative samples in the batch. It is a point cloud P cropping crop The feature representation extracted by the network, Fp, is the feature representation of point cloud P extracted by the network, and... Constitute a positive sample. This represents the feature representation of the j-th negative sample. The cosine similarity represents the similarity between positive sample pairs. This represents the cosine similarity between the anchor sample and the j-th negative sample. This represents the score after exponentially scaling the similarity of positive samples. This represents the score after exponentially scaling the similarity of each negative sample. This represents the sum of similarities with all negative samples, and -log(·) indicates that the negative log-likelihood encourages the maximization of similarity between positive samples and negative samples, while suppressing the similarity with negative samples.

6. The point cloud retrieval method according to claim 3, characterized in that, Step 2.3 specifically includes the following steps: Step 2.3.1, integrate high-level features Upsampling to The scale, and with Channel stitching is performed, followed by feature fusion via a multilayer perceptron. ; （3）； in, Indicates an upsampling operation. This indicates a channel splicing operation. Represents a shared multilayer perceptron; Step 2.3.2, will Upsampling to The scale, and with Channel stitching is performed, followed by feature fusion via a multilayer perceptron. ; （4）。 7. The point cloud retrieval method according to claim 5, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Construct a local adaptive aggregation descriptor submodule, and dynamically generate personalized cluster centers adapted to the current input sample through iterative optimization using a multi-layer Transformer decoder. ; Step 3.2, based on the final cluster centers For features at multiple scales Calculate the corresponding soft assignment weights for each feature point, and then use these weights to aggregate the residuals between each feature point and the personalized cluster centers to generate descriptors for the corresponding scale. ; Step 3.3 involves concatenating and fusing the multi-scale descriptors to obtain the initial global descriptor. and through a gating calibration mechanism The channels are adaptively enhanced to output the final global descriptor. .

8. The point cloud retrieval method according to claim 7, characterized in that, The core of the local adaptive aggregation descriptor submodule in step 3.1 lies in adapting the cluster centers to the feature distribution of the input samples through an iterative attention mechanism; specifically, it is implemented as follows: Step 3.1.1, let the set of general cluster centers before the update be... It contains k cluster centers, and the input features are ,by As a query (0) Input features as a bond matrix Sum matrix , No. The next update formula is: （5）； （6）； （7）； in, These are the learnable projection matrices corresponding to the query matrix Q, the key matrix K, and the value matrix V, respectively. The resulting attention weights; Step 3.1.2: Following the update strategy in Step 3.1.1, the personalized cluster centers are obtained after a total of M iterations. ; Step 3.1.3: Use the momentum update strategy to merge the general cluster centers and the personalized cluster centers to obtain the final cluster centers; （8）； in, The momentum coefficient, As a general cluster center, As a personalized clustering center, It serves as the final cluster center.

9. The point cloud retrieval method according to claim 8, characterized in that, In step 3.2, the descriptor Calculated using the following formula: （9）； in, The final set of cluster centers contains k cluster centers. The weight for assigning the i-th feature point to the k-th cluster center. Let i be the local feature descriptor of the i-th point. To perform residual calculations on all points The weighted aggregation; concat(·) concatenates the vectors corresponding to the k cluster centers; The gating calibration mechanism in step 3.3 is calculated using the following formula; （10）； in, It is the Sigmoid activation function. and For learnable parameters, This indicates element-wise multiplication. To concatenate the three scale descriptors to obtain the global descriptor, F output To enhance the global feature descriptor of the output.

10. The point cloud retrieval method according to claim 1, characterized in that, During model training, the total loss function used in step 4 is... Classification loss Triple loss And comparative loss The weighted summation consists of: （11）； in, and The weights are dynamically adjusted during training: They are set at the beginning of training. , =0.3 to enhance structure awareness, training settings in the later stages. , =0.8 to enhance feature discrimination.

Images