A density-based adaptive hybrid point cloud scene recognition method

Abstract

The present invention belongs to the field of computer vision technology and discloses a density-based adaptive hybrid point cloud scene recognition method. First, a two-stage neighbor search module is designed to dynamically adjust the neighborhood size of point clouds with different distributions based on density. On this basis, a dynamic local feature aggregation network is proposed, combined with a density-driven adjacency matrix to ensure accurate extraction of local details. In addition, an efficient contrast-enhanced linear attention network is further proposed to enable the model to capture more discriminative global context features and better perform scene modeling. Finally, voxel feature extraction is combined to resist local noise interference, and point-by-point features are fused into voxel features to compensate for the loss of details caused by the voxelization process. A large number of experiments on data sets show that the present invention achieves state-of-the-art point cloud scene recognition results with excellent robustness and generalization, verifying the effectiveness of the algorithm.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a density-based adaptive hybrid point cloud scene recognition method. Background Art

[0002] Point cloud scene recognition aims to identify the location represented by the query point cloud by comparing the currently acquired point cloud with its previously collected database point cloud. It plays an important role in tasks such as simultaneous localization and mapping in large-scale environments, loop closure detection, long-term localization, and global positioning, and is a key step in achieving high-precision and robust localization. In recent years, LiDAR sensors have garnered significant attention due to their robustness to illumination variations and stability in complex environments. However, the diversity and complexity of point cloud data in large-scale scenes pose many challenges for generating discriminable and robust global descriptors, including the impact of density variations on feature extraction. Currently, most learning-based point cloud scene recognition methods lack in-depth analysis and modeling of the density variation characteristics of LiDAR point clouds. Dense regions typically contain richer details, while sparse regions are more susceptible to noise interference, resulting in less robust feature extraction and lower retrieval accuracy. Therefore, research on density-based point cloud scene recognition methods that take point cloud density variations into account is not only of great theoretical significance but also particularly valuable for practical applications.

[0003] The IEEE Transactions on Image Processing, 1258–1270, 2022, proposes a lightweight proxy point convolution module that efficiently aggregates local point cloud features and significantly reduces computational resource consumption. However, this method fails to fully account for the density variations of point clouds in large-scale scenes, limiting its ability to capture local structural features in dense areas and resulting in high sensitivity to noise. The IEEE International Conference on Pattern Recognition (ICPR), 3699–3705, 2022, combines voxelization with a pyramid-structured three-dimensional sparse convolution and an efficient channel attention module to effectively capture spatial relationships in point clouds and construct scene descriptors. However, information loss introduced during the voxelization process limits further performance improvements. The IEEE International Conference on Computer Vision (ICCV), 8427–8438, 2023, proposes a hierarchical cross-attention transformer module that fuses point features with voxel features to enhance descriptive capabilities. However, this method fails to effectively aggregate local information in point-by-point features, and the noise in point features further interferes with the expression of voxel features, reducing the overall robustness. Summary of the Invention

[0004] The purpose of this invention is to propose a density-based adaptive hybrid point cloud scene recognition method to improve the accuracy and robustness of LiDAR point cloud geolocation. To address the challenge of point cloud density variation, this invention innovatively incorporates the characteristics of point cloud density variation into location recognition and proposes a density-based adaptive hybrid point cloud scene recognition method to generate distinguishable and robust descriptors, thereby improving the robot's point cloud scene recognition capabilities and providing reliable support for application scenarios such as positioning and navigation.

[0005] The technical solution of the present invention is as follows: a density-based adaptive hybrid point cloud scene recognition method is established to generate a global descriptor by establishing a density-based adaptive hybrid point cloud scene recognition network; the density-based adaptive hybrid point cloud scene recognition network has a dual-branch architecture, including a point-by-point feature extraction network branch and a voxel feature extraction network branch; the features of the two branches are globally average pooled and fused to generate a global descriptor of the scene;

[0006] The point-by-point feature extraction network branch includes a two-stage neighbor search module and two dual-stream feature extraction networks; the dual-stream feature extraction network includes a dynamic local feature aggregation network and a contrast-enhanced linear attention network;

[0007] The specific steps are as follows: Input point cloud X p Input them into the point-by-point feature extraction network branch and the voxel feature extraction network branch respectively; in the point-by-point feature extraction network branch, input the point cloud X p The binary space adjacency matrix A is obtained through the two-stage neighbor search module and the first two-stream feature extraction network respectively. The binary space adjacency matrix A is adaptively extracted from the dynamic local feature aggregation network to extract the point-by-point local features F. l ; Input point cloud X p In contrast to the enhanced linear attention network, a point-by-point global feature F is generated. g , and the point-by-point local features F l Add them together to get point-by-point features; the point-by-point features are used as the input of the second dual-stream feature extraction network, and the point-by-point features are output in the same feature extraction method. The point-by-point features output by the two dual-stream feature extraction networks are added together to get the final point-by-point features F p ;

[0008] In the voxel feature extraction network branch, the input point cloud X p Voxelize and feed into the voxel feature extraction network, and fuse the final point-by-point feature F after the first voxel feature extraction based on efficient attention p , then upsample the fusion features to generate voxel features F v, which is consistent with the final point-wise feature F p Global average pooling is performed separately and the global descriptors of the scene point cloud are obtained by splicing. The density-based adaptive hybrid point cloud scene recognition network is supervised by the truncated smoothed average precision approximation loss function to calculate the similarity between the global descriptors and realize point cloud scene recognition.

[0009] Furthermore, the specific process of the two-stage neighbor point search module is as follows:

[0010] For the original input point cloud X p , calculate the distance d of each point to the origin O, and use the gamma distribution to represent the overall trend of the point cloud with distance change:

[0011]

[0012] Where s and θ are the shape parameter and scale parameter respectively, Γ(·) is the gamma function, and f(d) is the point cloud distance prior function;

[0013] Perform the first stage of neighbor point selection and set the number of neighbor points k s1 Restricted to Within the range, more neighboring points are selected for point clouds near non-blind areas, and fewer neighboring points are selected for distant point clouds:

[0014]

[0015] In the second stage, adaptive neighbor point adjustment is performed based on the local point cloud density. The search radius r of the local area at different distances is defined as:

[0016] r=glog(αd+β) (3)

[0017] Among them, γ, α, and β are hyperparameters; the search radius is used to determine the number of point clouds in the local area, and this is used as the local density of the point cloud. The Euclidean distance between point cloud pairs is defined as d b ; For the i-th point, the number of neighboring points obtained in the first stage is The local density is Dynamically adjust the number of neighboring points so that dense point clouds have more neighboring points, while sparse point clouds have neighboring points that are limited to the sphere defined by r:

[0018]

[0019] According to the distance d between the point cloud pairs b , and combined with the number of neighboring points obtained by the two-stage neighbor search module, a point cloud binary space adjacency matrix A is constructed, where each element indicates whether a given point is adjacent to another point.

[0020] Furthermore, the specific process of the dynamic local feature aggregation network is as follows:

[0021] For the two dual-stream feature extraction networks, the initial point cloud features I of their respective dynamic local feature aggregation networks are the original input point cloud X p The point cloud features after one-dimensional convolution and the point cloud features after one-dimensional convolution of the output of the first two-stream feature extraction network; the initial point cloud features I are preprocessed by one-dimensional convolution to obtain features F, which are then multiplied by the binary space adjacency matrix A and averaged by the number of neighboring points k:

[0022]

[0023] Among them A i 、U i Represents the adjacent vector and proxy point features of the i-th point respectively; U=[U1,…,U N ] represents all proxy point features; then, the association between the proxy point and the target point is calculated to obtain the new aggregate feature F':

[0024] F'=g Θ (UF) (6)

[0025] where g Θ It is a nonlinear function with learnable parameters; finally, F' is connected to I through the residual structure to obtain the output I' of the dynamic local feature aggregation network.

[0026] Furthermore, the specific process of the contrast-enhanced linear attention network is as follows:

[0027] For two two-stream feature extraction networks, their respective input features F of the contrast-enhanced linear attention network are in are the original input point clouds X p Point cloud features after one-dimensional convolution, point cloud features after one-dimensional convolution of the output of the first two-stream feature extraction network; input feature F in , first generate the initial query matrix Q0, key matrix K0 and value matrix V0 through the projection matrix; then, through a matrix K with learnable parameters p Perform position encoding on K0, use ReLU function to ensure the non-negativity of features, and map function f q (·) Adjust the direction of point features while performing feature normalization to enhance the contrast between point clouds:

[0028]

[0029] Then, the focal feature F is extracted by focal linear attention FA f :

[0030] F f =FA(Q,K,V)=Q(K T V) (8)

[0031] In order to further enhance the diversity of features, channel attention CA is adopted to perform one-dimensional convolution on the value matrix V to generate channel features F c ; Then, the channel feature F c With the focus feature F f Combined, we get the contrast enhancement feature F ce :

[0032] F ce =CEA(Q,K,V)=FA(Q,K,V)+CA(V) (9)

[0033] Contrast enhancement feature F ce Perform channel transformation and compare it with the input feature F in Add together to generate the output F of the contrast-enhanced linear attention network out :

[0034] F out =r Θ (CEA(Q,K,V))+F in (10)

[0035] where r Θ is a nonlinear function with learnable parameters;

[0036] Finally, the output features of the contrast-enhanced linear attention network and the dynamic point-by-point feature aggregation network are added to generate point-by-point fusion features; two dual-stream feature extraction networks are performed to capture multi-scale features, and the output results are spliced ​​to form the point-by-point feature F pf .

[0037] Furthermore, the specific process of the voxel feature extraction network branch is as follows:

[0038] Input point cloud X p Quantized into sparse voxels, the voxel space coordinates X are obtained v , and extract voxel input features F through 3D sparse convolution v Then, a double-layer pyramid structure is used to extract voxel features. Each layer of the pyramid is composed of a three-dimensional sparse convolution and an efficient channel attention module, which extracts local features and captures the relationship between different channels respectively. A feature fusion module is introduced after the first efficient channel attention module to combine the dense point-by-point features F around the voxel into a single feature. pf Interpolate the distance weights to create sparse features F p2v , and with the voxel feature F vf Perform fusion and output fusion feature F fuse; After the second efficient channel attention module, transposed convolution is used to adjust the feature space resolution, realize multi-scale feature fusion, and generate voxel feature maps with different receptive fields.

[0039] Beneficial effects of the present invention: The present invention innovatively incorporates point cloud density change features into point cloud scene recognition, and proposes a density-based adaptive hybrid point cloud scene recognition method to generate distinguishable and robust descriptors, thereby achieving reliable positioning. First, the present invention designs a two-stage neighbor search module to dynamically adjust the neighborhood size of point clouds with different distributions based on density. On this basis, the present invention proposes a dynamic local feature aggregation network, combined with a density-driven adjacency matrix, to ensure accurate extraction of local details. In addition, the present invention further proposes an efficient contrast-enhanced linear attention network, which enables the model to capture more discriminative global context features and better perform scene modeling. Finally, the present invention combines voxel feature extraction to resist local noise interference, and fuses point-by-point features into voxel features to compensate for the loss of details caused by the voxelization process. A large number of experiments on data sets show that the present invention achieves state-of-the-art point cloud scene recognition results with excellent robustness and generalization, verifying the effectiveness of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Flowchart of the density-based adaptive hybrid point cloud scene recognition method. DETAILED DESCRIPTION

[0041] Figure 1 This is the main flow chart of the technical solution of the present invention. Figure 1 As shown, the density-based adaptive hybrid point cloud scene recognition algorithm proposed in the present invention includes the following steps:

[0042] Step 1, training data: The present invention constructs a relevant training set from the NCLT dataset and provides input point clouds for the network;

[0043] Step 2, model training: The network framework proposed in the present invention is a dual-branch architecture, consisting of a point-by-point feature extraction network branch and a voxel feature extraction network branch. The point-by-point feature extraction network branch includes two dual-stream feature extraction networks for aggregating multi-scale features, where the dynamic local feature aggregation network integrates the features of neighboring points obtained by the two-stage neighbor search module, while the contrast-enhanced linear attention network focuses on global context features. The voxel feature extraction network branch adopts a pyramid structure, where each scale uses sparse convolution and efficient channel attention modules for feature extraction. The point-by-point features are fused to enhance the voxel features through a feature fusion module. Finally, the features from the two branches are globally average pooled and fused to generate a global descriptor of the scene. The proposed network is supervised by a truncated smoothed average precision approximation loss function to calculate the similarity between descriptors to achieve point cloud scene recognition. Next, we will introduce these in detail.

[0044] Step 2.1: For the original input point cloud X p , calculate the distance d of each point to the origin O, and use the gamma distribution to represent the overall trend of the point cloud with distance change:

[0045]

[0046] Where s and θ are the shape parameter and scale parameter respectively, Γ(·) is the gamma function, and f(d) is the point cloud distance prior function;

[0047] Then, the first stage of neighbor point selection is performed, and the number of neighbor points is limited to Within the range, more neighboring points are selected for point clouds near non-blind areas, and fewer neighboring points are selected for distant point clouds:

[0048]

[0049] In the second stage, adaptive neighbor point adjustment is performed based on the local point cloud density. The search radius r of the local area at different distances is defined as:

[0050] r=glog(αd+β) (3)

[0051] Among them, γ, α, and β are hyperparameters; the search radius is used to determine the number of point clouds in the local area, and this is used as the local density of the point cloud. The Euclidean distance between point cloud pairs is defined as d b , for the i-th point, the number of neighboring points obtained in the first stage is The local density is Dynamically adjust the number of neighboring points so that dense point clouds have more neighboring points, while sparse point clouds have neighboring points that are limited to the sphere defined by r:

[0052]

[0053] According to the distance d between the point cloud pairs b , and combined with the number of neighboring points obtained by the two-stage neighbor search module, a point cloud binary space adjacency matrix A is constructed, where each element indicates whether a given point is adjacent to another point.

[0054] Step 2.2: For the two dual-stream feature extraction networks, the initial point cloud features I of their respective dynamic local feature aggregation networks are the original input point cloud X p The point cloud features after one-dimensional convolution and the point cloud features after one-dimensional convolution of the output of the first two-stream feature extraction network; the initial point cloud features I are preprocessed by one-dimensional convolution to obtain features F, which are then multiplied by the binary space adjacency matrix A and averaged by the number of neighboring points k:

[0055]

[0056] Among them A i 、U i Represent the adjacent vector and proxy point features of the i-th point respectively. U=[U1,…,U N ] represents all proxy point features; then, the association between the proxy point and the target point is calculated to obtain the new aggregate feature F':

[0057] F'=g Θ (UF) (6)

[0058] where g Θ It is a nonlinear function with learnable parameters; finally, F' is connected to I through the residual structure to obtain the output I' of the dynamic local feature aggregation network.

[0059] Step 2.3: For the two two-stream feature extraction networks, each compares the input features F of the enhanced linear attention network in are the original input point clouds X p Point cloud features after one-dimensional convolution, point cloud features after one-dimensional convolution of the output of the first two-stream feature extraction network; input feature F in , first generate the initial query matrix Q0, key matrix K0 and value matrix V0 through the projection matrix; then, through a matrix K with learnable parameters p Perform position encoding on K0, use ReLU function to ensure the non-negativity of features, and map function f q (·) Adjust the direction of point features while performing feature normalization to enhance the contrast between point clouds:

[0060]

[0061] in

[0062] Then, the focal feature F is extracted by focal linear attention (FA) f :

[0063] F f =FA(Q,K,V)=Q(K T V) (8)

[0064] In order to further enhance the diversity of features, channel attention (CA) is adopted to perform one-dimensional convolution on the value matrix V to generate channel features F c ; Then, the channel feature F c With the focus feature F f Combined, we get the contrast enhancement feature F ce :

[0065] F ce =CEA(Q,K,V)=FA(Q,K,V)+CA(V) (9)

[0066] Contrast enhancement feature F ce Perform channel transformation and compare it with the input feature F in Add together to generate the output F of the contrast-enhanced linear attention network out :

[0067] F out =r Θ (CEA(Q,K,V))+F in (10)

[0068] where r Θ is a nonlinear function with learnable parameters;

[0069] Finally, the output features of the contrast-enhanced linear attention network and the dynamic point-by-point feature aggregation network are added to generate point-by-point fusion features; two dual-stream feature extraction networks are performed to capture multi-scale features, and the output results are spliced ​​to form the point-by-point feature F pf .

[0070] Step 2.4: Input point cloud X p Quantized into sparse voxels, the voxel space coordinates X are obtained v , and extract voxel input features F v Then, a double-layer pyramid structure is used to extract voxel features. Each layer of the pyramid is composed of a three-dimensional sparse convolution and an efficient channel attention module, which respectively extracts local features and captures the relationship between different channels. In order to solve the information loss caused by sparse quantization, a feature fusion module is introduced after the first efficient channel attention module to combine the dense point-by-point features F around the voxel into a single feature. pf Interpolate the distance weights to create sparse features F p2v, and with the voxel feature F vf Perform fusion and output fusion feature F fuse ; After the second efficient channel attention module, transposed convolution is used to adjust the feature space resolution, realize multi-scale feature fusion, and generate voxel feature maps with different receptive fields; finally, global average pooling is applied to the point-by-point features and voxel features respectively, and the results are concatenated to generate a global scene descriptor; the network is supervised by the truncated smoothed average precision approximation loss function, and the similarity between descriptors is calculated to realize point cloud scene recognition.

[0071] Step 3, Point Cloud Retrieval: During the testing phase, we first input the library point cloud into a density-based adaptive hybrid point cloud recognition network to generate global descriptors for each point cloud and store these descriptors in the database. Next, we input the query point cloud into the proposed network to obtain its global descriptor. This descriptor is then compared with the descriptors of the database point clouds to find the closest geographic object to the query point cloud among point clouds collected at different times, thus completing the point cloud retrieval task.

[0072] To verify the effectiveness of our algorithm for large-scale point cloud location recognition, we tested it on the NCLT dataset, using the 2012-01-22 sequence as the database. The experimental results show that our proposed algorithm achieved an average Recall@1 of 81.2% across the six sequences used. Table 1 shows the Recall@1 results for our method using different sequences as queries.

[0073]

[0074]

Claims

1. A density-based adaptive hybrid point cloud scene recognition method, characterized in that: Establishing a density-based adaptive hybrid point cloud scene recognition network to generate a global descriptor; the density-based adaptive hybrid point cloud scene recognition network is a dual-branch architecture, including a point-by-point feature extraction network branch and a voxel feature extraction network branch; The features of the two branches are globally average pooled and fused to generate a global descriptor of the scene; The point-by-point feature extraction network branch includes a two-stage neighbor search module and two dual-stream feature extraction networks; the dual-stream feature extraction network includes a dynamic local feature aggregation network and a contrast-enhanced linear attention network; The specific steps are as follows: Input point cloud X p Input them into the point-by-point feature extraction network branch and the voxel feature extraction network branch respectively; in the point-by-point feature extraction network branch, input the point cloud X p Input into the two-stage neighbor search module and the first two-stream feature extraction network respectively; The binary space adjacency matrix A is obtained through the two-stage neighbor point search module, and the binary space adjacency matrix A is adaptively extracted from the dynamic local feature aggregation network point by point local features F l ; Input point cloud X p In contrast to the enhanced linear attention network, a point-by-point global feature F is generated. g , and the point-by-point local features F l Add them together to get point-by-point features; the point-by-point features are used as the input of the second dual-stream feature extraction network, and the point-by-point features are output in the same feature extraction method. The point-by-point features output by the two dual-stream feature extraction networks are added together to get the final point-by-point features F p ; In the voxel feature extraction network branch, the input point cloud X p Voxelize and feed into the voxel feature extraction network, and fuse the final point-by-point feature F after the first voxel feature extraction based on efficient attention p , then upsample the fusion features to generate voxel features F v , which is consistent with the final point-by-point feature F p Global average pooling is performed separately and the global descriptors of the scene point cloud are obtained by splicing. The density-based adaptive hybrid point cloud scene recognition network is supervised by the truncated smoothed average precision approximation loss function to calculate the similarity between the global descriptors and realize point cloud scene recognition.

2. The density-based adaptive hybrid point cloud scene recognition method according to claim 1, characterized in that: The specific process of the two-stage neighbor point search module is as follows: For the original input point cloud X p , calculate the distance d of each point to the origin O, and use the gamma distribution to represent the overall trend of the point cloud with distance change: Where s and θ are the shape parameter and scale parameter respectively, Γ(·) is the gamma function, and f(d) is the point cloud distance prior function; Perform the first stage of neighbor point selection and set the number of neighbor points k s1 Restricted to Within the range, more neighboring points are selected for point clouds near non-blind areas, and fewer neighboring points are selected for distant point clouds: In the second stage, adaptive neighbor point adjustment is performed based on the local point cloud density. The search radius r of the local area at different distances is defined as: r=glog(αd+β) (3) Among them, γ, α, and β are hyperparameters; the search radius is used to determine the number of point clouds in the local area, and this is used as the local density of the point cloud. The Euclidean distance between point cloud pairs is defined as d b ; For the i-th point, the number of neighboring points obtained in the first stage is The local density is Dynamically adjust the number of neighboring points so that dense point clouds have more neighboring points, while sparse point clouds have neighboring points that are limited to the sphere defined by r: According to the distance d between the point cloud pairs b , and combined with the number of neighboring points obtained by the two-stage neighbor search module, a point cloud binary space adjacency matrix A is constructed, where each element indicates whether a given point is adjacent to another point.

3. The density-based adaptive hybrid point cloud scene recognition method according to claim 1, characterized in that: The specific process of the dynamic local feature aggregation network is as follows: For the two dual-stream feature extraction networks, the initial point cloud features I of their respective dynamic local feature aggregation networks are the original input point cloud X p The point cloud features after one-dimensional convolution and the point cloud features after one-dimensional convolution of the output of the first two-stream feature extraction network; the initial point cloud features I are preprocessed by one-dimensional convolution to obtain features F, which are then multiplied by the binary space adjacency matrix A and averaged by the number of neighboring points k: Among them A i 、U i Represent the adjacent vector and proxy point features of the i-th point respectively; U=[U1,…,U N ] represents all proxy point features; then, the association between the proxy point and the target point is calculated to obtain the new aggregate feature F': F'=g Θ (U-F) (6) where g Θ It is a nonlinear function with learnable parameters; finally, F' is connected to I through the residual structure to obtain the output I' of the dynamic local feature aggregation network.

4. The density-based adaptive hybrid point cloud scene recognition method according to claim 1, characterized in that: The specific process of the contrast-enhanced linear attention network is as follows: For two two-stream feature extraction networks, their respective input features F of the contrast-enhanced linear attention network are in are the original input point clouds X p Point cloud features after one-dimensional convolution, point cloud features after one-dimensional convolution of the output of the first two-stream feature extraction network; input feature F in , first generate the initial query matrix Q0, key matrix K0 and value matrix V0 through the projection matrix; then, through a matrix K with learnable parameters p Perform position encoding on K0, use ReLU function to ensure the non-negativity of features, and map function f q (·) Adjust the direction of point features while performing feature normalization to enhance the contrast between point clouds: Then, the focal feature F is extracted by focal linear attention FA f : F f =FA(Q,K,V)=Q(K T V) (8) In order to further enhance the diversity of features, channel attention CA is adopted to perform one-dimensional convolution on the value matrix V to generate channel features F c ; Then, the channel feature F c With the focus feature F f Combined, we get the contrast enhancement feature F ce : F ce =CEA(Q,K,V)=FA(Q,K,V)+CA(V) (9) Contrast enhancement feature F ce Perform channel transformation and compare it with the input feature F in Add together to generate the output F of the contrast-enhanced linear attention network out : F out =r Θ (CEA(Q,K,V))+F in (10) where r Θ is a nonlinear function with learnable parameters; Finally, the output features of the contrast-enhanced linear attention network and the dynamic point-by-point feature aggregation network are added to generate point-by-point fusion features; two dual-stream feature extraction networks are performed to capture multi-scale features, and the output results are spliced ​​to form the point-by-point feature F pf .

5. The large-scale point cloud scene recognition method based on density-adaptive hybrid network according to claim 1 is characterized in that: The specific process of the voxel feature extraction network branch is as follows: Input point cloud X p Quantized into sparse voxels, the voxel space coordinates X are obtained v , and extract voxel input features F through 3D sparse convolution v Then, a double-layer pyramid structure is used to extract voxel features. Each layer of the pyramid is composed of a three-dimensional sparse convolution and an efficient channel attention module, which extracts local features and captures the relationship between different channels respectively. A feature fusion module is introduced after the first efficient channel attention module to combine the dense point-by-point features F around the voxel into a single feature. pf Interpolate the distance weights to create sparse features F p2v , and with the voxel feature F vf Perform fusion and output fusion feature F fuse ; After the second efficient channel attention module, transposed convolution is used to adjust the feature space resolution, realize multi-scale feature fusion, and generate voxel feature maps with different receptive fields.

Images