Large-scale point cloud geometric compression method based on dual-branch neural network

Abstract

This invention discloses a large-scale point cloud geometric compression method based on a dual-branch neural network, comprising: representing the point cloud as a sequence of nodes, acquiring data through dense and sparse feature windows, and inputting the acquisition results of the two windows into a Transformer-based dual-branch neural network including dense and sparse context branches. This effectively extracts sparse large-scale context and local detail context. The dense and sparse feature sampling results are complementary, avoiding the compression of all point clouds and thus reducing space occupation. Furthermore, the sparse and dense context features are fused to determine the node occupancy code and probability distribution. Then, the occupancy code and probability distribution are encoded using an entropy model, ultimately achieving efficient compression of the large-scale point cloud. This invention can reduce the space occupied by large-scale point cloud compression and has high efficiency, and can be widely applied in the field of point cloud compression.

Description

Technical Field

[0001] This invention relates to the field of point cloud compression technology, and in particular to a large-scale point cloud geometric compression method based on a dual-branch neural network. Background Technology

[0002] Point clouds, as a flexible form of 3D data representation, typically consist of the coordinates of points in three-dimensional space and related attributes (such as color and reflectivity). Point clouds have been widely used in many fields such as autonomous driving, virtual reality, and robotics. However, point clouds occupy a huge amount of storage space, making storage and transmission difficult, and urgently requiring efficient point cloud compression methods. Some deep learning-based methods employ sparse 3D convolutions to better model the dependencies between encoded data and encoded voxels. In existing CNN-based models, spatial context information is extracted better than handcrafted models. However, due to their limited receptive domain, they collect geometric context within a relatively narrow range, which leads to poor performance on large point clouds.

[0003] Therefore, existing large-scale point cloud compression technologies still have shortcomings such as low efficiency and large space requirements. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a large-scale point cloud geometric compression method based on a dual-branch neural network that reduces space occupation and is highly efficient.

[0005] One aspect of this invention provides a large-scale point cloud geometric compression method based on a dual-branch neural network, comprising:

[0006] The input point cloud is represented in the form of an octree, and the octree is represented as a node sequence through the ancestor node aggregation module;

[0007] The node sequence is sampled using a dense feature window to obtain a dense feature sampling result; the node sequence is sampled using a sparse feature window to obtain a sparse feature sampling result; the window length of the dense feature window is smaller than the window length of the sparse feature window;

[0008] The dense feature sampling results are input into the dense context branch of the Transformer-based dual-branch neural network to obtain dense context features; the sparse feature sampling results are input into the sparse context branch of the Transformer-based dual-branch neural network to obtain sparse context features.

[0009] The dense context features and sparse context features are fused by the feature fusion module to obtain fused features;

[0010] The dual-branch neural network determines the occupancy code and probability distribution of each node in the dense feature sampling result and the sparse feature sampling result based on the fusion features.

[0011] The occupancy code and its probability distribution are input into the entropy model to obtain the bitstream output by the entropy model after encoding. Preferably, the sampling of the node sequence using a dense feature window to obtain the dense feature sampling result includes:

[0012] Dense nodes in the node sequence are acquired by sliding a dense feature window.

[0013] The occupancy code, level, and quadrant number of each dense node are connected to form a dense main feature vector;

[0014] The embedding of each dense node is obtained by performing a learnable embedding matrix on the dense principal feature vector;

[0015] The embeddings of the ancestors of each dense node are aggregated to form a dense fusion embedding matrix, which serves as the dense feature sampling result.

[0016] Preferably, the step of inputting the dense feature sampling result into the dense context branch of a Transformer-based two-branch neural network to obtain dense context features includes:

[0017] The dense feature sampling results are input into a dense encoder network in a Transformer-based dual-branch neural network to obtain dense context features.

[0018] Preferably, the step of sampling the node sequence using a sparse feature window to obtain sparse feature sampling results includes:

[0019] Sparse nodes in the node sequence are acquired by sliding a sparse feature window.

[0020] The hierarchy of each sparse node is connected to the quadrant number to form a sparse principal feature vector;

[0021] The embedding of each sparse node is obtained by performing a learnable embedding matrix on the sparse principal feature vector;

[0022] The embeddings of the ancestors of each sparse node are aggregated to form a sparse fusion embedding matrix, which serves as the sparse feature sampling result.

[0023] Preferably, the step of inputting the sparse feature sampling result into the sparse context branch of a Transformer-based two-branch neural network to obtain sparse context features includes:

[0024] The sparse feature sampling results are input into the encoder of the sparse transformer in the Transformer-based dual-branch neural network to obtain sparse context features.

[0025] Preferably, it further includes:

[0026] For each node in the current frame, use the KNN algorithm to search for the y nearest neighbor nodes in the previous frame;

[0027] Connect the embeddings of each node and its corresponding neighboring nodes to obtain the inter-frame node embeddings.

[0028] Preferably, both the dense context branch and the sparse context branch include two Transformers and one MLP layer; the feature mixing module includes one MLP layer and three Transformers.

[0029] Another aspect of this invention provides a large-scale point cloud geometry compression device based on a dual-branch neural network, comprising:

[0030] The node sequence acquisition unit is used to represent the input point cloud in the form of an octree, and to represent the octree as a node sequence through the ancestor node aggregation module;

[0031] A feature sampling unit is used to sample the node sequence using a dense feature window to obtain a dense feature sampling result; and to sample the node sequence using a sparse feature window to obtain a sparse feature sampling result; wherein the window length of the dense feature window is smaller than the window length of the sparse feature window;

[0032] The context feature extraction unit is used to input the dense feature sampling result into the dense context branch of the Transformer-based two-branch neural network to obtain dense context features; and to input the sparse feature sampling result into the sparse context branch of the Transformer-based two-branch neural network to obtain sparse context features.

[0033] The context feature fusion unit is used to fuse the dense context features and the sparse context features through the feature mixing module to obtain fused features;

[0034] The probability distribution determination unit is used to determine the occupancy code and the probability distribution of the occupancy code of each node in the dense feature sampling result and the sparse feature sampling result based on the fusion features through the dual-branch neural network.

[0035] A compression coding unit is used to input the occupancy code and its probability distribution into the entropy model to obtain the bitstream output by the entropy model after encoding.

[0036] Another aspect of the present invention provides an electronic device, including a processor and a memory;

[0037] The memory is used to store programs;

[0038] The processor executes the program to implement the above-described method.

[0039] Another aspect of this invention provides a computer-readable storage medium storing a program that is executed by a processor to implement the above-described method.

[0040] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the method described above.

[0041] This invention represents point clouds as a sequence of nodes, which are then acquired through dense and sparse feature windows. The acquisition results from both windows are input into a Transformer-based dual-branch neural network, which includes dense and sparse context branches. This effectively extracts sparse large-scale context and local detail context. The dense and sparse feature sampling results are complementary, avoiding the compression of all point clouds and thus reducing space usage. Furthermore, the sparse and dense context features are fused to determine the node occupancy code and probability distribution. Finally, the occupancy code and probability distribution are encoded using an entropy model, ultimately achieving efficient compression of large-scale point clouds. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1A flowchart illustrating a large-scale point cloud geometric compression method based on a dual-branch neural network, provided for an embodiment of the present invention;

[0044] Figure 2 An example structure diagram of a two-branch neural network based on Transformer provided in an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of node aggregation provided in an embodiment of the present invention;

[0046] Figure 4 An example flowchart of a large-scale point cloud geometric compression method based on a dual-branch neural network provided in an embodiment of the present invention;

[0047] Figure 5 This is a structural block diagram of a large-scale point cloud geometry compression device based on a dual-branch neural network, provided for an embodiment of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] Reference Figure 1 This invention provides a method for large-scale point cloud geometric compression based on a dual-branch neural network, specifically including the following steps:

[0050] Step S100: Represent the input point cloud in the form of an octree, and represent the octree as a node sequence through the ancestor node aggregation module.

[0051] Specifically, due to the compact and progressive representation of point clouds by octrees, this invention can employ octrees in point cloud compression. Each node in an octree represents a cubic bounding box in three-dimensional space. The root node represents the largest bounding box of the point cloud. From the root to the leaf, the octree recursively divides the bounding box into eight sub-octets. The occupancy state of a node (8-bit binary code) represents the occupancy state of the eight sub-bounding boxes. The octree can be constructed as a sequence of occupancy codes, as follows:

[0052] X={x i |x i ∈N + ,x i ∈[1,255],i=0,1,…,n}

[0053] Where N represents the number of nodes in the octree.

[0054] Step S110: Sample the node sequence using a dense feature window to obtain a dense feature sampling result; sample the node sequence using a sparse feature window to obtain a sparse feature sampling result; the window length of the dense feature window is less than the window length of the sparse feature window.

[0055] Specifically, the process of obtaining dense feature sampling results may include the following:

[0056] S1. Use a dense feature window to slide and collect dense nodes in the node sequence.

[0057] S2. Connect the occupancy code, level and quadrant number of each dense node to form a dense main feature vector.

[0058] S3. Obtain the embedding of each dense node by performing a learnable embedding matrix on the dense principal feature vector.

[0059] S4. Aggregate the embeddings of the ancestors of each dense node to form a dense fusion embedding matrix, which serves as the dense feature sampling result.

[0060] Specifically, in dense context branches, this invention constructs a sliding context window of length N1 and encodes all octree nodes X window by window. Given an octree, the code window it occupies... This invention assigns the occupancy code x to each node. k Level l k and quadrant number o k Connect them to form the main feature vector e″ k =[x k ,l k ,o k The embedding of each octree node is obtained by performing a learnable embedding matrix w on the principal eigenvector, i.e., e′. k =e″ k w. To further extract the context of ancestor nodes, we aggregate the embeddings of the A-1 ancestors of each node to construct a fused embedding matrix, i.e. Then, X i The embedding matrix is ​​represented as

[0061] Specifically, the process of obtaining sparse feature sampling results may include the following:

[0062] S1. Use a sparse feature window to slide and collect sparse nodes in the node sequence.

[0063] S2. Connect the level and quadrant number of each sparse node to form a sparse principal feature vector.

[0064] S3. Obtain the embedding of each sparse node by performing a learnable embedding matrix on the sparse principal feature vector.

[0065] S4. Aggregate the embeddings of the ancestors of each sparse node to form a sparse fusion embedding matrix, which serves as the sparse feature sampling result.

[0066] Specifically, after sparse feature window sampling, the octree node input sparse context extraction branch is used for feature extraction. In the sparse context branch, this invention designs a sliding context window of length N2 (N2>N1). Given the occupied code sequence in the dense context branch... This invention uses X i The sparse window is placed symmetrically to the center. Specifically, there is the following octree sequence with sparse context branches:

[0067]

[0068] Then, the hierarchy and quadrant number of each node are connected to form the principal feature vector e″. k =[l k ,o k Following a process similar to dense context extraction, this invention can employ a learnable embedding matrix w. S And aggregate the information of the a-1 ancestors for each node to take advantage of large-scale context. Then, The embedding matrix can be formulated as follows:

[0069]

[0070] in, It is the embedding matrix of a single node.

[0071] Step S120: Input the dense feature sampling result into the dense context branch of the Transformer-based dual-branch neural network to obtain dense context features; input the sparse feature sampling result into the sparse context branch of the Transformer-based dual-branch neural network to obtain sparse context features.

[0072] Reference Figure 2 This invention provides an example structure diagram of a two-branch neural network based on Transformer.

[0073] Specifically, a Transformer-based dense encoder network can be used to utilize the context of nodes within a dense feature window, as shown in the following formula:

[0074]

[0075] Where DenseEncoder represents a Transformer-based encoder network, and F... i Dense X represents i The aggregated feature matrix.

[0076] Specifically, an encoder based on a sparse transformer is used to generate X. i Sparse Feature matrix:

[0077]

[0078] Where SparseEncoder represents an encoder network based on a sparse transformer, f i S This represents the feature vector of a single octree node.

[0079] Step S130: The dense context features and sparse context features are fused through the feature fusion module to obtain fused features.

[0080] Specifically, the window features in dense and sparse context branches can be aggregated as follows:

[0081]

[0082] Here, FusionEncoder represents a transformer-based fusion encoder network, while It is a concatenation of features in dense and sparse context branches. Then, X i The estimated probability distribution is:

[0083]

[0084] θ represents the neural network parameters.

[0085] Step S140: The dual-branch neural network determines the occupancy code and probability distribution of each node in the dense feature sampling result and the sparse feature sampling result based on the fusion features.

[0086] Specifically, this invention can use the cross-entropy between the predicted occupied code probability distribution and the actual occupied code as the loss function, as follows:

[0087]

[0088] Step S150: Input the occupancy code and its probability distribution into the entropy model to obtain the bitstream output by the entropy model after encoding.

[0089] Specifically, lossless entropy encoding of octrees requires a probability distribution of P(X). Since the actual P(X) is unknown, an entropy model is proposed to obtain a good approximation of P(X). The compression method needs to find the best approximation of the actual distribution of the occupancy code to minimize the Shannon cross-entropy. Where Q(X) represents the approximate distribution.

[0090] In one alternative implementation, the backbone of the neural network of the present invention consists of Transformer and MLP layers. Each individual branch includes two Transformers and one MLP layer. For the fusion of two branches, one MLP layer and three Transformers are used.

[0091] To utilize the contextual information of dynamic point clouds, this invention can also design an inter-frame context aggregation module to extract cross-frame context based on encoded geometric information. (Refer to...) Figure 3 This invention provides a schematic diagram of node aggregation. Figure 3 In this context, E represents the embedding. For each node in the current frame, the KNN algorithm is used to search for the y nearest neighbor nodes in the previous frame. The embeddings of these nodes are then concatenated to obtain the inter-frame node embedding.

[0092] For example, the inter-frame embedding of the k-th frame is:

[0093]

[0094] in, Indicates inter-frame features. This represents the feature vector of the y-th neighboring node in the previous frame, using the aggregated node information on the two branches. The estimated probability distribution of intra-frame modes is formulated as follows:

[0095]

[0096] Where θ represents the neural network parameters.

[0097] To describe the invention in more detail, specific examples will be used to illustrate its practical application.

[0098] Reference Figure 4 This invention provides an example flowchart of a large-scale point cloud geometric compression method based on a dual-branch neural network.

[0099] Specifically, the input point cloud is first represented using an octree. Then, through an ancestor node aggregation module, the octree is represented using a node sequence. Next, a sliding window is used to sample the node sequence, and the nodes within the window are input into a Transformer-based two-branch neural network. The neural network is used to estimate the occupancy code of each node, obtaining the probability distribution of the occupancy code. The probability distribution of the occupancy code and the occupancy code are then input into an entropy model, which encodes the data and outputs the final bitstream.

[0100] Reference Figure 5 This invention provides a large-scale point cloud geometry compression device based on a dual-branch neural network, comprising:

[0101] The node sequence acquisition unit is used to represent the input point cloud in the form of an octree, and to represent the octree as a node sequence through the ancestor node aggregation module;

[0102] A feature sampling unit is used to sample the node sequence using a dense feature window to obtain a dense feature sampling result; and to sample the node sequence using a sparse feature window to obtain a sparse feature sampling result; wherein the window length of the dense feature window is smaller than the window length of the sparse feature window;

[0103] The context feature extraction unit is used to input the dense feature sampling result into the dense context branch of the Transformer-based two-branch neural network to obtain dense context features; and to input the sparse feature sampling result into the sparse context branch of the Transformer-based two-branch neural network to obtain sparse context features.

[0104] The context feature fusion unit is used to fuse the dense context features and the sparse context features through the feature mixing module to obtain fused features;

[0105] The probability distribution determination unit is used to determine the occupancy code and the probability distribution of the occupancy code of each node in the dense feature sampling result and the sparse feature sampling result based on the fusion features through the dual-branch neural network.

[0106] A compression coding unit is used to input the occupancy code and its probability distribution into the entropy model to obtain the bitstream output by the entropy model after encoding.

[0107] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform... Figure 1 The method shown.

[0108] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0109] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0110] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0112] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0113] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0114] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0115] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0116] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A large-scale point cloud geometric compression method based on a dual-branch neural network, characterized in that, include: The input point cloud is represented in the form of an octree, and the octree is represented as a node sequence through the ancestor node aggregation module; The node sequence is sampled using a dense feature window to obtain dense feature sampling results; The node sequence is sampled using a sparse feature window to obtain sparse feature sampling results; The window length of the dense feature window is smaller than the window length of the sparse feature window; The dense feature sampling results are input into the dense context branch of a Transformer-based two-branch neural network to obtain dense context features; The sparse feature sampling results are input into the sparse context branch of a Transformer-based dual-branch neural network to obtain sparse context features. The dense context features and sparse context features are fused by the feature fusion module to obtain fused features; The dual-branch neural network determines the occupancy code and probability distribution of each node in the dense feature sampling result and the sparse feature sampling result based on the fusion features. The occupancy code and its probability distribution are input into the entropy model to obtain the bitstream output by the entropy model after encoding. The sampling of the node sequence using a dense feature window to obtain dense feature sampling results includes: Dense nodes in the node sequence are acquired by sliding a dense feature window. The occupancy code, level, and quadrant number of each dense node are connected to form a dense main feature vector; The embedding of each dense node is obtained by performing a learnable embedding matrix on the dense principal feature vector; Aggregate each of the dense nodes The ancestor's embedding is used to form a dense fusion embedding matrix, which serves as the dense feature sampling result; The step of sampling the node sequence using a sparse feature window to obtain sparse feature sampling results includes: Sparse nodes in the node sequence are acquired by sliding a sparse feature window. The hierarchy of each sparse node is connected to the quadrant number to form a sparse principal feature vector; The embedding of each sparse node is obtained by performing a learnable embedding matrix on the sparse principal feature vector; Aggregate each of the sparse nodes The ancestor's embedding is used to form a sparse fusion embedding matrix, which serves as the sparse feature sampling result.

2. The large-scale point cloud geometric compression method based on a dual-branch neural network according to claim 1, characterized in that, The step of inputting the dense feature sampling result into the dense context branch of a Transformer-based two-branch neural network to obtain dense context features includes: The dense feature sampling results are input into a dense encoder network in a Transformer-based dual-branch neural network to obtain dense context features.

3. The large-scale point cloud geometric compression method based on a dual-branch neural network according to claim 1, characterized in that, The step of inputting the sparse feature sampling result into the sparse context branch of a Transformer-based two-branch neural network to obtain sparse context features includes: The sparse feature sampling results are input into the encoder of the sparse transformer in the Transformer-based dual-branch neural network to obtain sparse context features.

4. The large-scale point cloud geometric compression method based on a dual-branch neural network according to claim 1, characterized in that, Also includes: For each node in the current frame, use the KNN algorithm to search for the y nearest neighbor nodes in the previous frame; Connect the embeddings of each node and its corresponding neighboring nodes to obtain the inter-frame node embeddings.

5. The large-scale point cloud geometric compression method based on a dual-branch neural network according to any one of claims 1-4, characterized in that, Both the dense context branch and the sparse context branch include two Transformers and one MLP layer; the feature mixing module includes one MLP layer and three Transformers.

6. A large-scale point cloud geometric compression device based on a dual-branch neural network, characterized in that, The apparatus is used to implement the large-scale point cloud geometric compression method based on a dual-branch neural network as described in claim 1, and the apparatus includes: The node sequence acquisition unit is used to represent the input point cloud in the form of an octree, and to represent the octree as a node sequence through the ancestor node aggregation module; A feature sampling unit is used to sample the node sequence using a dense feature window to obtain a dense feature sampling result; and to sample the node sequence using a sparse feature window to obtain a sparse feature sampling result; wherein the window length of the dense feature window is smaller than the window length of the sparse feature window; The context feature extraction unit is used to input the dense feature sampling result into the dense context branch of the Transformer-based two-branch neural network to obtain dense context features; and to input the sparse feature sampling result into the sparse context branch of the Transformer-based two-branch neural network to obtain sparse context features. The context feature fusion unit is used to fuse the dense context features and the sparse context features through the feature mixing module to obtain fused features; The probability distribution determination unit is used to determine the occupancy code and the probability distribution of the occupancy code of each node in the dense feature sampling result and the sparse feature sampling result based on the fusion features through the dual-branch neural network. A compression coding unit is used to input the occupancy code and its probability distribution into the entropy model to obtain the bitstream output by the entropy model after encoding.

7. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 5.

Images