A 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion

Abstract

This invention discloses a 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion. First, a multi-view 2D image sequence is acquired as input, and multi-scale feature extraction is performed using a feature pyramid network. Then, a parallel channel-spatial attention mechanism is introduced into the feature extraction network to adaptively enhance key features and suppress noise. High- and low-level features are dynamically fused through adaptively gated multi-scale fusion blocks, and gated residual connections are introduced to ensure the stability of information flow. Multi-scale convolution is used to capture contextual information, and depth maps are predicted step-by-step through cost volume construction and regularization in a cascaded structure. Finally, a 3D point cloud model of the target scene is generated. This invention effectively solves the problem of feature matching difficulties in complex scenes such as weak textures or high reflectivity by using parallel attention and adaptive gated fusion mechanisms, significantly improving the accuracy and robustness of 3D reconstruction, and has good practical value.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and 3D reconstruction technology, specifically relating to a 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion. Background Technology

[0002] Multi-view stereo vision (MVS) technology aims to recover 3D point clouds of a scene from multiple 2D images. In recent years, deep learning has significantly improved the performance of MVS on standard datasets. However, existing methods are mostly based on ideal scene designs, and when faced with weak textures, highly reflective surfaces, and complex occlusions in real-world environments, the reconstruction quality often degrades due to feature matching failures.

[0003] Typically, existing learning-based MVS methods use fixed network structures to extract features, failing to fully consider the differences in importance between feature channels and spatial regions, and lacking the ability to adaptively enhance key information and suppress noise based on image content. Furthermore, existing multi-scale feature fusion strategies are often too simple and fixed, making it difficult to dynamically balance features at different levels, and failing to effectively guarantee the stability of cross-scale information flow, thus limiting the accuracy and robustness of the final reconstruction.

[0004] Therefore, there is an urgent need to develop a multi-view 3D reconstruction method with adaptive feature perception enhancement capabilities and a robust dynamic multi-scale fusion mechanism. Summary of the Invention

[0005] The purpose of this invention is to provide a 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion, which solves the problem that existing methods in complex industrial scenarios such as weak texture and high reflectivity suffer from limited reconstruction accuracy and geometric loss due to insufficient feature extraction and fusion capabilities.

[0006] The technical solution adopted in this invention is a 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion, which is implemented according to the following steps: Step 1: Construct a 3D point cloud reconstruction network, obtain multi-view 2D image sequences as input, and perform multi-scale feature extraction; Step 2: The extracted multi-scale feature maps are fed into the parallel channel-space attention mechanism CASA module, which is divided into a channel attention sub-module and a spatial attention sub-module; Step 3: Input the attention-enhanced features into the adaptive gated multi-scale fusion block AGMF, introduce a gating mechanism, and use multi-scale convolutional kernels to extract rich contextual information; Step 4: Predict a high-precision depth map; Step 5: Filter and fuse the initial point cloud to reconstruct a 3D point cloud model of the target scene.

[0007] The invention is further characterized in that, In step 1, the 3D point cloud reconstruction network sequentially includes a feature extraction module, an attention enhancement module, a feature fusion module, and a depth estimation module. The output of the feature extraction module is connected to the input of the attention enhancement module, the output of the attention enhancement module is connected to the input of the feature fusion module, and the output of the feature fusion module is connected to the input of the depth estimation module.

[0008] Step 2 is implemented in the following steps: Step 201: Feed the feature maps extracted by the Feature Pyramid Network (FPN) into the channel attention submodule and the spatial attention submodule in parallel; Step 202: The channel attention submodule processes the feature map. Simultaneously, global average pooling and global max pooling are performed to obtain average pooling features and max pooling features, respectively. The average pooling features and max pooling features are then fed into a shared multilayer perceptron and processed by an activation function to generate channel attention weights, which are used to adjust the importance of different channels. Step 203: The spatial attention submodule performs average pooling and max pooling on the input feature map along the channel dimension to obtain channel average pooling features and channel max pooling features, respectively; the channel average pooling features and channel max pooling features are concatenated along the channel dimension, and spatial attention weights are generated by convolution operation and activation function on the concatenated features; finally, the channel attention weights and spatial attention weights are applied to the original feature map.

[0009] 4. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 3, characterized in that the operation of the channel attention submodule in step 202 is as shown in formula (1): (1) In formula (1), the feature map extracted in step 201 is used. Each was subjected to global average pooling. and global max pooling After the operation, the data is sent to the shared multilayer sensor. After Activation function Processing to obtain channel attention weights .

[0010] The operation of the spatial attention submodule in step 203 is shown in the following formula (2): (2) In formula (2), the feature map extracted in step 201 is used. Average pooling is performed along the channel dimension. and max pooling The results are concatenated and then subjected to convolution operations. and Activation function Spatial attention weights are obtained after the operation. .

[0011] Step 3 is implemented in the following steps: Step 301: Use convolutional networks to learn the relationship between high-level abstract features and low-level detailed features, generate adaptive fusion weights, and combine high- and low-level features in a weighted manner accordingly. Step 302: Introduce a gating mechanism to generate gating coefficients, adaptively control the proportion of low-level features flowing into the next level in the form of residuals, and maintain stable information flow; Step 303: Process and stitch the fused features using convolutional kernels of different sizes to enhance the multi-scale expressive power of the features and restore spatial resolution.

[0012] The generation and weighted combination operations of the adaptive fusion weights in step 301 are shown in formulas (3) and (4): (3) (4) In formula (3), the high-level features are included. With low-level features splicing Through multilayer perceptron and Sigmoid activation function Obtain adaptive fusion weights In formula (4), this weight is used. right and Perform weighted combination to obtain the fused features .

[0013] The gating mechanism and residual information control operation described in step 302 are shown in the following formulas (5) and (6): (5) (6) Formula (5) utilizes 1x1 convolution. and Sigmoid activation function Gating weights are generated based on high- and low-level features. In formula (6), the final output feature is... The fusion features obtained from step 301 With gated weights Low-level features of regulation It is formed by adding the residuals together.

[0014] In step 4, the cascaded network adopts a coarse-to-fine strategy. In each level, the cost volume is constructed based on the planar scanning algorithm using the features after feature enhancement and fusion. Then, the cost volume is regularized by a three-dimensional convolutional neural network and the current level's depth map is generated by regression. The coarse depth map generated by the previous level is used to guide the next level to perform a more refined depth estimation.

[0015] Step 5 is implemented in the following steps: Using the camera intrinsic and extrinsic parameters corresponding to the input image, the multi-view 2D depth map predicted in step 4 is back-projected onto the 3D space to generate the initial 3D point cloud for each view. Then multi-view Figure 1 Consistency verification calculates the reprojection error of the point cloud between different views, and combines photometric consistency and geometric consistency to remove noise caused by occlusion or mismatch. Finally, the filtered point clouds from each viewpoint are fused to obtain the final 3D point cloud model of the target scene.

[0016] The beneficial effects of this invention are that the 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion, mainly by introducing a parallel channel-spatial attention mechanism (CASA), can more effectively perceive key information in complex scenes during the feature extraction stage. Considering the negative impact of weak textures and highly reflective areas on feature matching accuracy, this attention mechanism is integrated into the backbone network to adaptively enhance salient features and suppress noise, thereby improving the model's feature representation ability and final reconstruction accuracy. From the perspective that high-precision depth estimation requires the effective combination of deep semantic information and shallow geometric details, an adaptive gated multi-scale fusion block (AGMF) is designed. Through dynamic weight adjustment and gated residual connections, it captures the dependencies between cross-scale features, ensuring the stability and richness of information flow. Furthermore, integrating the above components into a coarse-to-fine cascaded network architecture effectively alleviates the reconstruction failure problem caused by feature inconsistency in highly challenging industrial scenarios, significantly improving the robustness and generalization ability of the algorithm in practical applications, and demonstrating significant practical value. Attached Figure Description

[0017] Figure 1 This is the overall architecture diagram of the 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 This invention presents a 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion. First, a feature pyramid network is used as the backbone to extract multi-scale features from multi-view images to obtain information at different resolution levels. To further improve model performance, considering the differences in importance of different feature channels and spatial regions in complex scenes, a parallel channel-spatial attention mechanism is introduced to adaptively enhance effective features and suppress noise interference. Subsequently, an adaptively gated multi-scale fusion block is employed, which not only achieves dynamic weighted fusion of high- and low-level features but also ensures the stability of cross-scale information flow through gated residual connections and captures richer contextual information using multi-scale convolution. Finally, a cascaded network structure is combined for coarse-to-fine depth estimation, effectively addressing the technical problems of limited reconstruction accuracy and geometric incompleteness caused by insufficient feature extraction and fusion capabilities in traditional methods under complex industrial scenes with weak textures and high reflectivity.

[0020] Example 2 This invention is based on a 3D point cloud reconstruction method using attention-enhanced multi-scale feature fusion, combined with... Figure 1 The specific steps are as follows: Step 1: First, construct a 3D point cloud reconstruction network, obtain multi-view 2D image sequences as input, and use the Feature Pyramid Network (FPN) to extract multi-scale features from the input images to obtain feature maps at different resolution levels. Step 2: The extracted multi-scale feature maps are fed into the parallel channel-space attention mechanism CASA module. The network is divided into two parallel branches, which capture the global dependencies between channels through the channel attention submodule and focus on key spatial regions through the spatial attention submodule, thereby adaptively enhancing effective features and suppressing noise in both channel and spatial dimensions. Step 3: Input the attention-enhanced features into the adaptive gated multi-scale fusion block AGMF. By learning adaptive weights, high-level abstract features and low-level detailed features are dynamically fused. A gating mechanism is introduced to control the flow of residual information to ensure stability. Multi-scale convolutional kernels are used to extract rich contextual information and gradually restore the spatial resolution of the feature map. Step 4: In each level of the cascaded network structure, the cost volume is constructed using the processed features and regularized by a three-dimensional convolutional neural network. A coarse-to-fine strategy is used to progressively regress and predict a high-precision depth map. Step 5: Using camera parameters, backproject the multi-view 2D depth map obtained in Step 4 onto 3D space to generate an initial point cloud, and then use multi-view... Figure 1 Consistency verification filters and fuses the initial point cloud to reconstruct a 3D point cloud model of the target scene.

[0021] Example 3 This invention is based on a 3D point cloud reconstruction method using attention-enhanced multi-scale feature fusion, combined with... Figure 1 The specific steps are as follows: Step 1: First, construct a 3D point cloud reconstruction network, obtain multi-view 2D image sequences as input, and use the Feature Pyramid Network (FPN) to extract multi-scale features from the input images to obtain feature maps at different resolution levels. In step 1, the 3D point cloud reconstruction network sequentially includes a feature extraction module, an attention enhancement module, a feature fusion module, and a depth estimation module. The output of the feature extraction module is connected to the input of the attention enhancement module, the output of the attention enhancement module is connected to the input of the feature fusion module, and the output of the feature fusion module is connected to the input of the depth estimation module.

[0022] Step 2: The extracted multi-scale feature maps are fed into the parallel channel-space attention mechanism CASA module. The network is divided into two parallel branches, which capture the global dependencies between channels through the channel attention submodule and focus on key spatial regions through the spatial attention submodule, thereby adaptively enhancing effective features and suppressing noise in both channel and spatial dimensions. Step 2 is implemented in the following steps: Step 201: Feed the feature maps extracted by the Feature Pyramid Network (FPN) into the channel attention submodule and the spatial attention submodule in parallel; Step 202: The channel attention submodule processes the feature map. Simultaneously, global average pooling and global max pooling are performed to obtain average pooling features and max pooling features, respectively. The average pooling features and max pooling features are then fed into a shared multilayer perceptron and processed by an activation function to generate channel attention weights, which are used to adjust the importance of different channels. The operation of the channel attention submodule in step 202 is shown in formula (1): (1) In formula (1), the feature map extracted in step 201 is used. Each was subjected to global average pooling. and global max pooling After the operation, the data is sent to the shared multilayer sensor. After Activation function Processing to obtain channel attention weights .

[0023] Step 203: The spatial attention submodule performs average pooling and max pooling on the input feature map along the channel dimension to obtain channel average pooling features and channel max pooling features, respectively; the channel average pooling features and channel max pooling features are concatenated along the channel dimension, and spatial attention weights are generated by convolution operation and activation function on the concatenated features; finally, the channel attention weights and spatial attention weights are applied to the original feature map.

[0024] The operation of the spatial attention submodule in step 203 is shown in the following formula (2): (2) In formula (2), the feature map extracted in step 201 is used. Average pooling is performed along the channel dimension. and max pooling The results are concatenated and then subjected to convolution operations. and Activation function Spatial attention weights are obtained after the operation. .

[0025] Step 3: Input the attention-enhanced features into the adaptive gated multi-scale fusion block AGMF. By learning adaptive weights, high-level abstract features and low-level detailed features are dynamically fused. A gating mechanism is introduced to control the flow of residual information to ensure stability. Multi-scale convolutional kernels are used to extract rich contextual information and gradually restore the spatial resolution of the feature map. Step 3 is implemented in the following steps: Step 301: Use convolutional networks to learn the relationship between high-level abstract features and low-level detailed features, generate adaptive fusion weights, and combine high- and low-level features in a weighted manner accordingly. The generation and weighted combination operations of the adaptive fusion weights in step 301 are shown in formulas (3) and (4): (3) (4) In formula (3), the high-level features are included. With low-level features splicing Through multilayer perceptron and Sigmoid activation function Obtain adaptive fusion weights In formula (4), this weight is used. right and Perform weighted combination to obtain the fused features .

[0026] Step 302: Introduce a gating mechanism to generate gating coefficients, adaptively control the proportion of low-level features flowing into the next level in the form of residuals, and maintain stable information flow; The gating mechanism and residual information control operation described in step 302 are shown in the following formulas (5) and (6): (5) (6) Formula (5) utilizes 1x1 convolution. and Sigmoid activation function Gating weights are generated based on high- and low-level features. In formula (6), the final output feature is... The fusion features obtained from step 301 With gated weights Low-level features of regulation It is formed by adding the residuals together.

[0027] Step 303: Process and stitch the fused features using convolutional kernels of different sizes to enhance the multi-scale expressive power of the features and restore spatial resolution.

[0028] Step 4: In each level of the cascaded network structure, the cost volume is constructed using the processed features and regularized by a three-dimensional convolutional neural network. A coarse-to-fine strategy is used to progressively regress and predict a high-precision depth map. In step 4, the cascaded network adopts a coarse-to-fine strategy. In each level, the cost volume is constructed based on the planar scanning algorithm using the features after feature enhancement and fusion. Then, the cost volume is regularized by a three-dimensional convolutional neural network and the current level's depth map is generated by regression. The coarse depth map generated by the previous level is used to guide the next level to perform a more refined depth estimation.

[0029] Step 5: Using camera parameters, backproject the multi-view 2D depth map obtained in Step 4 onto 3D space to generate an initial point cloud, and then use multi-view... Figure 1 Consistency verification filters and fuses the initial point cloud to reconstruct a 3D point cloud model of the target scene.

[0030] Step 5 is implemented in the following steps: Using the camera intrinsic and extrinsic parameters corresponding to the input image, the multi-view 2D depth map predicted in step 4 is back-projected onto the 3D space to generate the initial 3D point cloud for each view. Then multi-view Figure 1 Consistency verification calculates the reprojection error of the point cloud between different views, and combines photometric consistency and geometric consistency to remove noise caused by occlusion or mismatch. Finally, the filtered point clouds from each viewpoint are fused to obtain the final 3D point cloud model of the target scene.

[0031] Example 4 This invention is based on a 3D point cloud reconstruction method using attention-enhanced multi-scale feature fusion, combined with... Figure 1 The specific steps are as follows: Step 1: First, construct a 3D point cloud reconstruction network, obtain multi-view 2D image sequences as input, and use the Feature Pyramid Network (FPN) to extract multi-scale features from the input images to obtain feature maps at different resolution levels. In step 1, the 3D point cloud reconstruction network sequentially includes a feature extraction module, an attention enhancement module, a feature fusion module, and a depth estimation module. The output of the feature extraction module is connected to the input of the attention enhancement module, the output of the attention enhancement module is connected to the input of the feature fusion module, and the output of the feature fusion module is connected to the input of the depth estimation module.

[0032] Step 2: The extracted multi-scale feature maps are fed into the parallel channel-space attention mechanism CASA module. The network is divided into two parallel branches, which capture the global dependencies between channels through the channel attention submodule and focus on key spatial regions through the spatial attention submodule, thereby adaptively enhancing effective features and suppressing noise in both channel and spatial dimensions. Step 2 is implemented in the following steps: Step 201: Feed the feature maps extracted by the Feature Pyramid Network (FPN) into the channel attention submodule and the spatial attention submodule in parallel; Step 202: The channel attention submodule processes the feature map. Simultaneously, global average pooling and global max pooling are performed to obtain average pooling features and max pooling features, respectively. The average pooling features and max pooling features are then fed into a shared multilayer perceptron and processed by an activation function to generate channel attention weights, which are used to adjust the importance of different channels. Step 203: The spatial attention submodule performs average pooling and max pooling on the input feature map along the channel dimension to obtain channel average pooling features and channel max pooling features, respectively; the channel average pooling features and channel max pooling features are concatenated along the channel dimension, and spatial attention weights are generated by convolution operation and activation function on the concatenated features; finally, the channel attention weights and spatial attention weights are applied to the original feature map.

[0033] Step 3: Input the attention-enhanced features into the adaptive gated multi-scale fusion block AGMF. By learning adaptive weights, high-level abstract features and low-level detailed features are dynamically fused. A gating mechanism is introduced to control the flow of residual information to ensure stability. Multi-scale convolutional kernels are used to extract rich contextual information and gradually restore the spatial resolution of the feature map. Step 4: In each level of the cascaded network structure, the cost volume is constructed using the processed features and regularized by a three-dimensional convolutional neural network. A coarse-to-fine strategy is used to progressively regress and predict a high-precision depth map. Step 5: Using camera parameters, backproject the multi-view 2D depth map obtained in Step 4 onto 3D space to generate an initial point cloud, and then use multi-view... Figure 1 Consistency verification filters and fuses the initial point cloud to reconstruct a 3D point cloud model of the target scene.

[0034] Example 5 This invention is based on a 3D point cloud reconstruction method using attention-enhanced multi-scale feature fusion, combined with... Figure 1 The specific steps are as follows: Step 1: First, construct a 3D point cloud reconstruction network, obtain multi-view 2D image sequences as input, and use the Feature Pyramid Network (FPN) to extract multi-scale features from the input images to obtain feature maps at different resolution levels. In step 1, the 3D point cloud reconstruction network sequentially includes a feature extraction module, an attention enhancement module, a feature fusion module, and a depth estimation module. The output of the feature extraction module is connected to the input of the attention enhancement module, the output of the attention enhancement module is connected to the input of the feature fusion module, and the output of the feature fusion module is connected to the input of the depth estimation module.

[0035] Step 2: The extracted multi-scale feature maps are fed into the parallel channel-space attention mechanism CASA module. The network is divided into two parallel branches, which capture the global dependencies between channels through the channel attention submodule and focus on key spatial regions through the spatial attention submodule, thereby adaptively enhancing effective features and suppressing noise in both channel and spatial dimensions. Step 2 is implemented in the following steps: Step 201: Feed the feature maps extracted by the Feature Pyramid Network (FPN) into the channel attention submodule and the spatial attention submodule in parallel; Step 202: The channel attention submodule processes the feature map. Simultaneously, global average pooling and global max pooling are performed to obtain average pooling features and max pooling features, respectively. The average pooling features and max pooling features are then fed into a shared multilayer perceptron and processed by an activation function to generate channel attention weights, which are used to adjust the importance of different channels. The operation of the channel attention submodule in step 202 is shown in formula (1): (1) In formula (1), the feature map extracted in step 201 is used. Each was subjected to global average pooling. and global max pooling After the operation, the data is sent to the shared multilayer sensor. After Activation function Processing to obtain channel attention weights .

[0036] Step 203: The spatial attention submodule performs average pooling and max pooling on the input feature map along the channel dimension to obtain channel average pooling features and channel max pooling features, respectively; the channel average pooling features and channel max pooling features are concatenated along the channel dimension, and spatial attention weights are generated by convolution operation and activation function on the concatenated features; finally, the channel attention weights and spatial attention weights are applied to the original feature map.

[0037] The operation of the spatial attention submodule in step 203 is shown in the following formula (2): (2) In formula (2), the feature map extracted in step 201 is used. Average pooling is performed along the channel dimension. and max pooling The results are concatenated and then subjected to convolution operations. and Activation function Spatial attention weights are obtained after the operation. .

[0038] Step 3: Input the attention-enhanced features into the adaptive gated multi-scale fusion block AGMF. By learning adaptive weights, high-level abstract features and low-level detailed features are dynamically fused. A gating mechanism is introduced to control the flow of residual information to ensure stability. Multi-scale convolutional kernels are used to extract rich contextual information and gradually restore the spatial resolution of the feature map. Step 3 is implemented in the following steps: Step 301: Use convolutional networks to learn the relationship between high-level abstract features and low-level detailed features, generate adaptive fusion weights, and combine high- and low-level features in a weighted manner accordingly. Step 302: Introduce a gating mechanism to generate gating coefficients, adaptively control the proportion of low-level features flowing into the next level in the form of residuals, and maintain stable information flow; Step 303: Process and stitch the fused features using convolutional kernels of different sizes to enhance the multi-scale expressive power of the features and restore spatial resolution.

[0039] Step 4: In each level of the cascaded network structure, the cost volume is constructed using the processed features and regularized by a three-dimensional convolutional neural network. A coarse-to-fine strategy is used to progressively regress and predict a high-precision depth map. Step 5: Using camera parameters, backproject the multi-view 2D depth map obtained in Step 4 onto 3D space to generate an initial point cloud, and then use multi-view... Figure 1 Consistency verification filters and fuses the initial point cloud to reconstruct a 3D point cloud model of the target scene.

[0040] Example 6 This embodiment describes a 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion. On one hand, it primarily introduces a parallel channel-spatial attention mechanism (CASA) during the feature extraction stage, enabling more effective perception of key feature information in complex scenes (such as areas with weak texture or high reflectivity). Considering the impact of noise interference and differences in feature importance on matching accuracy, this attention mechanism is integrated into the backbone network to adaptively enhance salient features and suppress invalid information, thereby improving the model's feature representation ability and final reconstruction accuracy. From the perspective that high-precision depth estimation requires effectively combining deep semantic information with shallow geometric details, an adaptive gated multi-scale fusion block (AGMF) is designed. Through dynamic weight adjustment and gated residual connections, it captures the dependencies between cross-scale features, ensuring the stability and richness of information flow.

[0041] On the other hand, integrating the above components into a coarse-to-fine cascaded network architecture not only effectively alleviates the reconstruction failure problem caused by feature inconsistency in highly challenging industrial scenarios, but also significantly improves the algorithm's generalization ability on multiple standard benchmark datasets, which has great practical value.

Claims

1. A 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion, characterized in that, The specific steps are as follows: Step 1: Construct a 3D point cloud reconstruction network, obtain multi-view 2D image sequences as input, and perform multi-scale feature extraction; Step 2: The extracted multi-scale feature maps are fed into the parallel channel-space attention mechanism CASA module, which is divided into a channel attention sub-module and a spatial attention sub-module; Step 3: Input the attention-enhanced features into the adaptive gated multi-scale fusion block AGMF, introduce a gating mechanism, and use multi-scale convolutional kernels to extract rich contextual information; Step 4: Predict a high-precision depth map; Step 5: Filter and fuse the initial point cloud to reconstruct a 3D point cloud model of the target scene.

2. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 1, characterized in that, In step 1, the 3D point cloud reconstruction network sequentially includes a feature extraction module, an attention enhancement module, a feature fusion module, and a depth estimation module; the output of the feature extraction module is connected to the input of the attention enhancement module, the output of the attention enhancement module is connected to the input of the feature fusion module, and the output of the feature fusion module is connected to the input of the depth estimation module.

3. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 2, characterized in that, Step 2 is implemented in the following steps: Step 201: Feed the feature maps extracted by the Feature Pyramid Network (FPN) into the channel attention submodule and the spatial attention submodule in parallel; Step 202: The channel attention submodule processes the feature map. Simultaneously, global average pooling and global max pooling are performed to obtain average pooling features and max pooling features, respectively. The average pooling features and max pooling features are then fed into a shared multilayer perceptron and processed by an activation function to generate channel attention weights, which are used to adjust the importance of different channels. Step 203: The spatial attention submodule performs average pooling and max pooling on the input feature map along the channel dimension to obtain channel average pooling features and channel max pooling features, respectively; the channel average pooling features and channel max pooling features are concatenated along the channel dimension, and spatial attention weights are generated by convolution operation and activation function on the concatenated features; finally, the channel attention weights and spatial attention weights are applied to the original feature map.

4. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 3, characterized in that, The operation of the channel attention submodule in step 202 is shown in formula (1): （1） In formula (1), the feature map extracted in step 201 is used. Each was subjected to global average pooling. and global max pooling After the operation, the data is sent to the shared multilayer sensor. After Activation function Processing to obtain channel attention weights .

5. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 4, characterized in that, The operation of the spatial attention submodule in step 203 is shown in the following formula (2): （2） In formula (2), the feature map extracted in step 201 is used. Average pooling is performed along the channel dimension. and max pooling The results are concatenated and then subjected to convolution operations. and Activation function Spatial attention weights are obtained after the operation. .

6. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 5, characterized in that, Step 3 is implemented in the following steps: Step 301: Use convolutional networks to learn the relationship between high-level abstract features and low-level detailed features, generate adaptive fusion weights, and combine high- and low-level features in a weighted manner accordingly. Step 302: Introduce a gating mechanism to generate gating coefficients, adaptively control the proportion of low-level features flowing into the next level in the form of residuals, and maintain stable information flow; Step 303: Process and stitch the fused features using convolutional kernels of different sizes to enhance the multi-scale expressive power of the features and restore spatial resolution.

7. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 6, characterized in that, The generation and weighted combination operations of the adaptive fusion weights in step 301 are shown in formulas (3) and (4): （3） （4） In formula (3), the high-level features are included. With low-level features splicing Through multilayer perceptron and Sigmoid activation function Obtain adaptive fusion weights In formula (4), this weight is used. right and Perform weighted combination to obtain the fused features .

8. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 7, characterized in that, The gating mechanism and residual information control operation in step 302 are shown in the following formulas (5) and (6): （5） （6） Formula (5) utilizes 1x1 convolution. and Sigmoid activation function Gating weights are generated based on high- and low-level features. In formula (6), the final output feature is... The fusion features obtained from step 301 With gated weights Low-level features of regulation It is formed by adding the residuals together.

9. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 8, characterized in that, In step 4, the cascaded network adopts a coarse-to-fine strategy. In each level, the cost volume is constructed based on the planar scanning algorithm using the features after feature enhancement and fusion. Then, the cost volume is regularized by a three-dimensional convolutional neural network and the current level's depth map is generated by regression. The coarse depth map generated by the previous level is used to guide the next level to perform a more refined depth estimation.

10. The 3D point cloud reconstruction method based on attention-enhanced multi-scale feature fusion according to claim 9, characterized in that, Step 5 is implemented in the following steps: Using the camera intrinsic and extrinsic parameters corresponding to the input image, the multi-view 2D depth map predicted in step 4 is back-projected onto the 3D space to generate the initial 3D point cloud for each view. Then, multi-view consistency verification is performed to calculate the reprojection error of the point cloud between different views, and noise caused by occlusion or mismatch is eliminated by combining photometric consistency and geometric consistency. Finally, the filtered point clouds from each viewpoint are fused to obtain the final 3D point cloud model of the target scene.

Images