Point cloud attribute coding, compression method and system based on lifting structure

Abstract

The application provides a point cloud attribute coding and compression method and system based on lifting structure, comprising: obtaining to-be-encoded point cloud data; obtaining a first lifting representation of the attribute signal based on the geometric coordinates by using a forward first lifting structure; obtaining a second lifting representation by passing the first lifting representation through a forward second lifting structure; obtaining a hidden layer representation according to the second lifting structure; quantizing the hidden layer representation by using a uniform scalar quantizer, and encoding the quantized hidden layer representation to obtain a point cloud attribute signal code stream. The application effectively utilizes the topological structure correlation of the point cloud signal, eliminates the redundant information of the point cloud attribute in the spatial and channel dimensions, and further realizes high-performance compression of the point cloud attribute; the reversible characteristics of the lifting structure are inherited, a compact multi-scale representation is generated, and strong support is provided for high-fidelity reconstruction.

Description

Technical Field

[0001] This invention relates to a solution in the field of point cloud data compression technology, specifically, to a point cloud attribute encoding / decoding method, compression method, and system based on a lifting structure. Background Technology

[0002] 3D point clouds, as a discrete data representation of continuous surfaces, possess six degrees of freedom in spatial representation, making them naturally suitable for depicting 3D objects and scenes in the real world. With the rapid development of large-scale 3D sensing devices and acquisition systems, their quantity and scale are expanding dramatically, and they are widely used in emerging frontier fields such as immersive communication, virtual / augmented / mixed reality, cultural heritage protection, and autonomous driving. They not only greatly enrich and facilitate modern life but also demonstrate enormous commercial application potential in smart cities, industrial inspection, and other scenarios. A 3D point cloud is a discrete set of points constructed by discretizing the continuous surfaces of objects in 3D space. These points depend on the topological structure of the continuous surfaces, exhibiting a highly irregular distribution in 3D space. Each point not only contains geometric information (i.e., 3D coordinates) but also carries additional attribute information such as color, reflectivity, and normal vectors. Furthermore, the data scale is often enormous; a single point cloud can consist of millions of points. In application scenarios with stringent requirements for real-time data transmission, the storage and transmission needs of 3D point cloud data have far exceeded the computing resources and local storage capacity of terminal devices due to the dual constraints of network bandwidth and storage capacity. This makes the efficient compression of 3D point cloud geometry and attributes a critical problem that urgently needs to be solved.

[0003] A literature review of existing technologies revealed that early point cloud attribute compression schemes used a phased coding approach to represent and encode point cloud attribute signals, primarily encompassing methods such as Discrete Fourier Transform, Wavelet Transform, Lifting Structure, and Sparse Representation. These phased coding framework-based point cloud attribute compression schemes separate the transformation module from the quantization and entropy coding modules, making overall collaborative optimization difficult and limiting compression performance. Deep learning-based point cloud attribute compression methods construct neural networks to jointly optimize the transformation, quantization, and entropy models in an end-to-end manner, thereby more accurately fitting the complex distribution of point clouds. X. Sheng et al., in their 2022 paper "Deep-PCAC: An end-to-end deep lossy compression framework for point cloud attributes" published in *IEEE Transactions on Multimedia*, designed a second-order point convolution and performed end-to-end compression of point cloud attributes based on a traditional variational autoencoder structure. However, this method is highly complex and has limited compression performance. To reduce computational complexity, J. Wang et al., in their 2022 paper "Sparse tensor-based point cloud attribute compression" (Sparse-PCAC) published at the *International Conference on Multimedia Information Processing and Retrieval*, further utilized sparse convolution to perform end-to-end compression of point cloud attributes. More recently, Z. Guo et al., in their 2024 paper "TSC-PCAC: voxel transformer and sparse convolution-based point cloud attribute compression for 3D broadcasting" (TSC-PCAC) published in *IEEE Transactions on Broadcasting*, introduced sparse convolution into the Transformer to mine local and global correlations between points in the point cloud. However, these methods neglect adaptively mining the topological correlations of the point cloud to obtain multi-scale representations of point cloud attributes in a reversible manner. Therefore, when processing highly irregularly distributed point cloud data, these methods struggle to efficiently represent point clouds in a compact form while maintaining high-fidelity reconstruction results.

[0004] Furthermore, in 2017, the MPEG international standardization organization officially released a Call for Proposals (CfP) for 3D point cloud compression technology, initiating the development of relevant international standards. After years of technical iterations and multiple rounds of working group reviews, the standardization project has entered the final draft stage. Initially, three test models (TMC1-TMC3) were released, targeting three types of point cloud compression: static point clouds (Category 1), dynamically time-varying point clouds (Category 2), and dynamically acquired point clouds (Category 3). TMC1 and TMC3 were technically integrated to form TMC13, which is the current geometry-based point cloud compression (G-PCC). G-PCC aims to utilize the geometric structure of point clouds to design novel transform coding strategies: in geometric data processing, it employs an octree geometry codec and a triangular mesh geometry codec; in attribute data compression, it utilizes Region Adaptive Wavelet Transform (RAHT) and Lifting Transform (Lift) to fully leverage spatial correlation. TMC2, also known as Video-based Point Cloud Compression (V-PCC), primarily targets dynamic, time-varying point cloud compression. Its core technology involves projecting 3D point cloud data into a 2D video sequence, then using mature video coding standards (such as HEVC and VVC) for efficient compression. However, G-PCC also employs a staged coding model and cannot adaptively utilize the topological information of the point cloud for end-to-end compression, thus limiting its coding performance. V-PCC, on the other hand, requires projection transformation, which introduces irreversible distortion. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a point cloud attribute encoding / decoding method, compression method, and system based on a lifting structure.

[0006] According to a first aspect of the present invention, a point cloud attribute encoding method based on lifting structure is provided, comprising:

[0007] Obtain the point cloud data to be encoded, including the geometric coordinates and attribute signals of each point;

[0008] Using a forward first lifting structure, a first lifted representation of the attribute signal is obtained based on the geometric coordinates. The first lifting structure achieves spatial adaptive partitioning by sensing the geometric information of the point cloud. A cross-attention filtering mechanism is designed to predict and update attributes using the correlation of geometric attributes.

[0009] The first boosted representation is passed through the forward boosting structure to obtain the second boosted representation. The second boosting structure consists of one or more pairs of channel dimensionality upscaling and channel dimensionality downscaling modules based on the forward boosting structure, as well as several optional channel dimension invariant modules based on the forward boosting structure.

[0010] The second boosted representation is passed through one or more sparse convolution-based downsampling layers to obtain the hidden layer representation;

[0011] The hidden layer representation is quantized using a uniform scalar quantizer, and the quantized hidden layer representation is then used for encoding to obtain the point cloud attribute signal bitstream.

[0012] Preferably, obtaining a first lifted representation of the attribute signal based on the geometric coordinates using a forward first lifting structure includes:

[0013] The geometric information of the point cloud is perceived to achieve adaptive spatial partitioning. The points in the point cloud are hierarchically and uniformly partitioned according to the geometric information C of the point cloud data P to be encoded.

[0014] For the attribute signals of each level after the division, a cross-attention filtering mechanism is designed using geometric attribute correlation to perform a forward first boosting operation and obtain a first boosted representation.

[0015] Preferably, the geometric information of the perceived point cloud enables adaptive spatial partitioning. Based on the geometric information C of the point cloud data P to be encoded, the points in the point cloud are hierarchically and uniformly partitioned, including:

[0016] Based on the geometric information of the input point cloud P, a global feature representation of the input point cloud is obtained by using an h-layer shared multilayer perceptron mechanism and a max pooling layer to extract one D-channel global feature. The number of channels in this h-layer shared multilayer perceptron is set as follows: Both h and h are preset positive integers;

[0017] Subsequently, an m-layer shared multilayer perceptron mechanism is used to extract local features of the D channels point by point. The number of channels in this m-layer shared multilayer perceptron is set as follows: Both m and m are preset positive integers;

[0018] Copy the global geometric feature representation of the obtained input point cloud to the same dimension as the obtained per-point local geometric feature, and splice the two along the channel dimension to construct a joint geometric feature representation G;

[0019] Based on the joint geometric feature representation G, use a self-attention mechanism with a head H1, an n-layer shared multi-layer perceptron mechanism, and a linear projection layer to adaptively estimate the importance score ω of each point, and normalize it using the softmax function, where the number of channels of the n-layer shared multi-layer perceptron mechanism is set to Both H1 and n are preset positive integers;

[0020] Using the obtained importance score ω of each point, sort the input point cloud P from low to high according to the importance score, and select the half of the points with high scores as the set of points to be updated The half of the points with low scores is used as the set of points to be predicted For the obtained set of points to be updated Continue to perform uniform partitioning according to the importance score ω, and select the half of the points with high scores as the set of points to be updated at the next level The half of the points with low scores is used as the set of points to be predicted at the next level Repeat the uniform partitioning process until the preset level L is reached, and obtain the L-layer set of points to be updated And the L-layer set of points to be predicted

[0021] Preferably, for the attribute signals at each level after partitioning, design a cross-attention filtering mechanism using geometric attribute correlation to perform a first forward boosting operation to obtain a first boosted representation, including:

[0022] Use a sparse convolutional layer to extract the high-dimensional feature F of the input point cloud attribute signal X as the initial attribute feature of the first layer;

[0023] Based on the initial attribute feature, perform a forward spatial boosting operation on the L-layer point clouds obtained by the partitioning in ascending order, and finally obtain a multi-scale attribute feature representation of the point cloud, including the residual attribute feature sets {E (1) , F (2) , …, E (L)} of all L-layer point clouds with resolutions halved successively from the first layer to the L layer, and the updated attribute feature F (L+1) ;

[0024] Among them, the first forward boosting operation on the attribute signal of any l-th (1 ≤ l < L) layer point cloud is:

[0025] Based on the geometric coordinates of the two parts of the point cloud obtained from the division, the initial attribute features F of the l-th layer are... (l) The points are assigned to two parts of the point cloud to obtain the set of points to be predicted. Attributes and characteristics and points to be updated Attributes and characteristics

[0026] Using the set of points to be updated geometric coordinates and attribute features and the set of points to be predicted geometric coordinates Each and Using the key tensor (K), value tensor (V), and query tensor (Q) as the three components, a cross-attention filtering mechanism is constructed to predict the set of points to be predicted. Attributes and characteristics

[0027] from Subtract the predicted value of the attribute feature from the value to obtain the residual attribute feature. in This represents the prediction operation at layer l;

[0028] Using the obtained set of points to be predicted residual attribute feature E (l) Geometric coordinates and points to be updated geometric coordinates Each and Using the key tensor (K), value tensor (V), and query tensor (Q) as the three components, a cross-attention filtering mechanism is constructed to update the set of points to be updated. Attributes and characteristics

[0029] The updated value of the attribute feature is compared with Add them together to form the initial attribute features of the (l+1)th layer. in This represents the update operation at level l.

[0030] Preferably, the cross-attention filtering mechanism operates as follows:

[0031] Get the key tensor (K), value tensor (V), and query tensor (Q);

[0032] The initial filtered output is obtained through an H2 head attention mechanism.

[0033]

[0034] Where H2 is a preset positive integer, and W O Represents the learnable parameter matrix;

[0035] The initial filtered output is enhanced using two sparse convolutional layers and one ReLU nonlinear activation layer. Obtain enhanced filter output As the output of the cross-attention filtering mechanism.

[0036] Preferably, the first lifting representation is passed through a forward second lifting structure to obtain a second lifting representation, including:

[0037] The forward second lifting structure connects t channel dimensionality-up modules and t channel dimensionality-down modules based on the forward lifting structure in series, achieving channel scaling step by step; and embeds several optional channel dimension-invariant modules based on the lifting structure after the channel dimensionality-up modules, wherein...

[0038] Each module of the forward second lifting structure includes two sparse convolutional layers and a ReLU activation function in its prediction and update operations. After the prediction and update operations, the predicted / updated values ​​are added to the input signal of each branch.

[0039] The channel dimension-up module based on the forward lifting structure forces the second input branch to zero and merges the two output branches along the channel dimension;

[0040] The channel dimensionality reduction module based on the forward lifting structure divides the input into two parts evenly along the channel dimension and discards the second output branch;

[0041] The channel dimension invariant module based on the lifting structure keeps the two input branches unchanged.

[0042] According to a second aspect of the present invention, a point cloud attribute encoding method based on a lifting structure is provided for corresponding decoding of the encoding obtained in any one of the claims, comprising:

[0043] Obtain the point cloud attribute signal bitstream, and process the point cloud attribute signal bitstream to obtain a quantized hidden layer representation;

[0044] The quantized hidden layer representation is dequantized to obtain the reconstructed hidden layer representation;

[0045] The reconstructed hidden layer representation is passed through one or more sparse convolution-based upsampling layers to obtain a reconstructed second boosted representation;

[0046] The reconstructed second lifting representation is passed through a reverse second lifting structure to obtain a reconstructed first lifting representation. The reverse second lifting structure is symmetrical to the forward second lifting structure. It connects t channel dimensionality-up modules and t channel dimensionality-down modules based on the reverse lifting structure in series to achieve channel scaling step by step. Several optional channel dimension-invariant modules based on the lifting structure are embedded after the channel dimensionality-up modules. Each reverse lifting module operates similarly to the previous lifting module, except that the summation step after the prediction and update operations of each module is replaced with subtraction, that is, the input signal of each branch is subtracted from the predicted value / updated value.

[0047] The reconstructed first lifting representation is passed through a reverse first lifting structure to obtain the reconstructed point cloud attribute signal.

[0048] According to a third aspect of the present invention, a point cloud attribute compression method based on lifting structures is provided, comprising:

[0049] Obtain the point cloud to be compressed, including the geometric coordinates and attribute signals of each point;

[0050] The point cloud attribute encoding method based on the lifting structure described above is used to encode the point cloud attribute signal into a point cloud attribute signal bitstream;

[0051] The point cloud attribute decoding method based on the above-mentioned lifting structure is used to decode the point cloud attribute signal stream to obtain the reconstructed point cloud attribute signal.

[0052] Preferably, the parameters of the forward and backward first lifting structure, the forward and backward second lifting structure, the sparse convolution-based downsampling and upsampling layers, the super-prior encoder, and the super-prior decoder are obtained through rate-distortion optimization, specifically through the following operations:

[0053] Step 1: Obtain N voxelized training point cloud data, divide them into slices, and select several slices as the training set;

[0054] Step 2: Based on the training set, use the point cloud attribute encoding method based on lifting structure to encode the point cloud attribute signals of the training set into a point cloud attribute signal bitstream.

[0055] Step 3: Using the point cloud attribute decoding method based on the lifting structure described above, decode the point cloud attribute signal stream to obtain the reconstructed training set point cloud attribute signal;

[0056] Step 4: For each training set point cloud slice, construct a loss function based on the attribute signal, the reconstructed attribute signal, the geometric coordinates, and the importance score ω of each point, and use it for end-to-end training to obtain the parameters of the point cloud compression method;

[0057] Step 5: Calculate the loss function for all Nr training set slices. in The loss function for the nth slice of the training set;

[0058] Step 6: Based on the loss function Alternately optimize the importance score ω of the partitioned modules and other learnable parameters in the first boosting structure;

[0059] Repeat steps two through six iteratively until convergence.

[0060] According to a fourth aspect of the present invention, a point cloud attribute signal compression system is provided, the point cloud attribute signal compression system comprising an encoder and a decoder, characterized in that:

[0061] Encoder: Using any of the point cloud attribute encoding methods based on lifting structures described above, the point cloud attribute signals are encoded into a point cloud attribute signal bitstream;

[0062] Decoder: The point cloud attribute signal stream is decoded using the point cloud attribute decoding method based on the lifting structure to obtain the reconstructed point cloud attribute signal.

[0063] According to a fifth aspect of the present invention, a terminal is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the program, it can be used to execute any of the point cloud attribute encoding methods based on lifting structures, or the point cloud attribute decoding methods based on lifting structures, or the point cloud attribute compression methods based on lifting structures described in the present invention.

[0064] According to a sixth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that, when executed by a processor, the program can be used to perform any of the point cloud attribute encoding methods based on lifting structures, or the point cloud attribute decoding methods based on lifting structures, or the point cloud attribute compression methods based on lifting structures described in any one of the present invention.

[0065] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0066] The point cloud attribute compression method based on lifting structure in this embodiment of the invention utilizes a lifting structure that can efficiently capture irregular geometric structures in point cloud data while eliminating redundant information in the spatial and channel dimensions of point cloud attributes. This method inherits the reversible properties of the lifting structure, generating compact multi-scale representations and achieving high-fidelity point cloud attribute reconstruction.

[0067] The point cloud attribute encoding / decoding method, compression method, and system based on lifting structure in this invention embodiment are applicable to lossless geometry and lossy attribute compression of point cloud signals. They make full use of the geometric topological correlation of point clouds and effectively improve the compression efficiency of point cloud attribute signals in an end-to-end manner.

[0068] The point cloud attribute encoding / decoding method, compression method, and system based on lifting structure in the embodiments of the present invention, compared with the variational autoencoder structure based on second-order point convolution, or sparse convolution, or a combination of sparse convolution and Transformer, can reversibly and adaptively utilize the correlation of irregular topological structure of point cloud signals to remove attribute redundancy. Therefore, it can effectively improve the reconstruction quality of point cloud attribute signals and save encoding overhead.

[0069] The point cloud attribute encoding / decoding method, compression method, and system based on lifting structure in this invention embodiment, compared with the MPEG G-PCC point cloud compression international standard, can adaptively mine geometric attribute correlations in an end-to-end optimized manner, while optimizing the transformation and entropy models, and obtain significant performance gains. Attached Figure Description

[0070] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0071] Figure 1 This is a flowchart of a point cloud attribute encoding method based on lifting structure according to an embodiment of the present invention;

[0072] Figure 2 This is a block diagram illustrating the principle of point cloud attribute encoding and decoding based on a lifting structure, according to a preferred embodiment of the present invention.

[0073] Figure 3 This is a schematic diagram of a lifting structure according to an embodiment of the present invention;

[0074] Figure 4 This is a flowchart of the module division in the first lifting structure according to an embodiment of the present invention;

[0075] Figure 5 This is a block diagram illustrating the principle of the cross-attention filtering mechanism in the first lifting structure according to an embodiment of the present invention.

[0076] Figure 6 This is a flowchart of a point cloud attribute decoding method based on a lifting structure according to an embodiment of the present invention. Detailed Implementation

[0077] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0078] This invention utilizes the inherent topological correlation of point cloud signals to obtain multi-scale representations of point cloud attributes in a reversible manner, effectively removing redundant information between signals. Furthermore, it performs end-to-end compression of the representation results based on a super-prior entropy coding model, achieving compression performance that surpasses the best existing deep learning-based coding schemes and the MPEG G-PCC international standard scheme.

[0079] Reference Figure 1 As shown, based on the above inventive concept, the present invention provides an embodiment of a point cloud attribute encoding method based on a lifting structure, comprising:

[0080] S1. Obtain the point cloud data to be encoded, including the geometric coordinates and attribute signals of each point;

[0081] S2. Using the forward first lifting structure, the first lifting representation of the attribute signal is obtained based on the geometric coordinates. The first lifting structure realizes the spatial adaptive partitioning by sensing the geometric information of the point cloud. A cross-attention filtering mechanism is designed to use the correlation of geometric attributes for attribute prediction and updating.

[0082] S3. Pass the first lifting representation through the forward second lifting structure to obtain the second lifting representation. The second lifting structure consists of one or more pairs of channel dimensionality increase and channel dimensionality decrease modules based on the forward lifting structure.

[0083] Of course, the second lifting structure can also include several optional channel dimension-invariant modules based on the forward lifting structure, set between the channel dimensionality-up module and the channel dimensionality-down module. If there are multiple pairs of channel dimensionality-up and channel dimensionality-down modules, then first t channel dimensionality-up modules are concatenated, then several channel dimension-invariant modules are concatenated, and finally t channel dimensionality-down modules are concatenated. S4. The second lifting representation is passed through one or more downsampling layers based on sparse convolution to obtain the hidden layer representation;

[0084] S5. Use a uniform scalar quantizer to quantize the hidden layer representation, and use the quantized hidden layer representation for encoding to obtain the point cloud attribute signal bitstream.

[0085] The above embodiments of the present invention, through the forward first lifting structure and the forward second lifting structure, can effectively utilize the topological correlation of point cloud signals, while eliminating redundant information of point cloud attributes in the spatial and channel dimensions, thereby achieving high-performance compression of point cloud attributes; inheriting the reversible characteristics of the lifting structure, it generates a compact multi-scale representation, providing strong support for high-fidelity reconstruction.

[0086] Reference Figure 2 As shown, in some embodiments of the present invention, based on the above steps S1-S5, wherein S5 uses a uniform scalar quantizer to quantize the hidden layer representation, and performs encoding operations using the quantized hidden layer representation, the detailed encoding operations may further include:

[0087] S51. Quantize the hidden layer representation using a uniform scalar quantizer;

[0088] S52. The quantized hidden layer representation is passed through a super-prior encoder to obtain super-prior information, wherein the super-prior encoder consists of one or more downsampling layers based on sparse convolution;

[0089] S53. Perform uniform scalar quantization on the aforementioned prior information to obtain the quantized prior information;

[0090] S54. Perform arithmetic encoding on the quantized prior information and output the prior information bitstream;

[0091] S55. Dequantize the quantized prior information to obtain the reconstructed prior information;

[0092] S56. The distribution of the quantized hidden layer representation is modeled as a Gaussian distribution, and a super-prior decoder is used to predict the mean and scale parameters of the Gaussian distribution based on the reconstructed super-prior information. The super-prior decoder consists of one or more upsampling layers based on sparse convolution.

[0093] S57. Based on the mean and scale parameters of the Gaussian mixture distribution, perform probability estimation on the quantized hidden layer representation and calculate the bit rate;

[0094] S58. Perform arithmetic coding based on the estimated probability to obtain the hidden layer representation bitstream;

[0095] S59. Merge the prior information code stream and the hidden layer representation code stream to obtain the point cloud attribute signal code stream.

[0096] In some embodiments of the present invention, step S2 above, which utilizes a forward first lifting structure to obtain a first lifted representation of the attribute signal based on the geometric coordinates, may include:

[0097] S21. Perceive the geometric information of the point cloud and realize adaptive spatial partitioning. Perform hierarchical uniform partitioning of the points in the point cloud according to the geometric information C of the point cloud P to be encoded.

[0098] S22. For the attribute signals of each level after partitioning, design a cross-attention filtering mechanism using geometric attribute correlation, perform a forward first boosting operation, and obtain the first boosted representation.

[0099] like Figure 3 As shown, in some preferred embodiments of the present invention, S21 senses the geometric information of the point cloud to achieve adaptive spatial partitioning, and performs hierarchical uniform partitioning of the points in the point cloud according to the geometric information C of the point cloud data P to be encoded, including:

[0100] S211, based on the geometric information of the input point cloud P, uses an h-layer shared multilayer perceptron mechanism and a max-pooling layer to extract global features from one D channel, obtaining a global geometric feature representation of the input point cloud. The number of channels in this h-layer shared multilayer perceptron is set as follows: Both h and h are preset positive integers;

[0101] S212 then uses an m-layer shared multilayer perceptron mechanism to extract local features of the D channels point by point. The number of channels in this m-layer shared multilayer perceptron is set as follows: Both m and m are preset positive integers;

[0102] S213 copies the global geometric feature representation of the obtained input point cloud to the same dimension as the obtained point-by-point local geometric features, and concatenates the two along the channel dimension to construct a joint geometric feature representation G;

[0103] S214, based on the joint geometric feature representation G, uses an H1 head self-attention mechanism, an n-layer shared multilayer perceptron mechanism, and a linear projection layer to adaptively estimate the importance score ω of each point, and normalizes it using a softmax function. The number of channels in the n-layer shared multilayer perceptron mechanism is set to... H1 and n are both preset positive integers;

[0104] S215 uses the importance score ω of each point to sort the input point cloud P from low to high importance score, and selects the half of the points with the highest scores as the point set to be updated. The half of the points with the lowest scores are used as the set of points to be predicted. For the obtained set of points to be updated Based on the importance score ω, the points are further divided evenly, and the half with the highest scores are selected as the set of points to be updated in the next level. The half of the points with the lowest scores are used as the set of points to be predicted in the next level. Repeat the uniform partitioning process until the preset level L is reached, and obtain the set of points to be updated at level L and the set of points to be predicted at level L

[0105] In this embodiment, the number of layers h of the shared multi-layer perceptron used to extract global features is set to 6, and the number of channels in each layer are 64, 64, 64, 128, 2048, and 512 respectively. The number of layers m of the shared multi-layer perceptron used to extract per-point local features is set to 3, and the number of channels in each layer are 32, 64, and 128 respectively; H₁ is set to 8, n is set to 3, and the number of channels in each layer of the shared multi-layer perceptron are 64, 64, and 64 respectively; the level L is set to 2 to balance performance and computational complexity; in other embodiments, other parameters can also be selected, and the present invention is not limited to the above parameters

[0106] In the above embodiments of the present invention, by adopting the detailed operation of S21, the hierarchical uniform partitioning is adaptively realized by perceiving the geometric information of the point cloud, and the problem of inconsistent dimensions caused by the odd-even partitioning strategy in the traditional lifting structure is solved

[0107] As Figure 4 shown, S22 perceives the geometric information of the point cloud, realizes the adaptive partitioning in space, and performs hierarchical uniform partitioning on the points in the point cloud according to the geometric information C of the point cloud data P to be encoded, including

[0108] S221 uses a sparse convolutional layer to extract the high-dimensional feature F of the input point cloud attribute signal X as the initial attribute feature of the first layer

[0109] S222 based on the initial attribute feature, performs forward spatial lifting operations on the L layers of point clouds obtained by partitioning in ascending order, and finally obtains the multi-scale attribute feature representation of the point cloud, including the set of residual attribute features {E (1) , E (2) , …, E (L)} of all L layers of point clouds with halved resolutions from the first layer to the L layer and the updated attribute feature F in the L layer point cloud (L+1) ;

[0110] S223 Among them, as Figure 4 shown, the forward first lifting operation on the point cloud attribute signal of any l-th (1 ≤ l < L) layer is

[0111] According to the geometric coordinates of the two parts of the point cloud obtained by partitioning, the initial attribute feature F (l) of the l-th layer is assigned to the two parts of the point cloud sets, and the set of points to be predicted is obtained Attributes and characteristics and points to be updated Attributes and characteristics

[0112] Utilizing the point set to be updated geometric coordinates and attribute features and the set of points to be predicted geometric coordinates Each and Using the key tensor (K), value tensor (V), and query tensor (Q) as the components, a geometrically perceptual cross-attention filtering mechanism is constructed to predict the set of points to be predicted. Attributes and characteristics

[0113] from Subtract the predicted values ​​of the attribute features from the original values ​​to obtain the residual attribute features. in This represents the prediction operation at layer l;

[0114] Using the obtained set of points to be predicted residual attribute feature E (l) Geometric coordinates and points to be updated geometric coordinates Each and Using the key tensor (K), value tensor (V), and query tensor (Q) as the components, a geometry-aware cross-attention filtering mechanism is constructed to update the set of points to be updated. Attributes and characteristics

[0115] Update the attribute value and Add them together to form the initial attribute features of the (l+1)th layer. in This represents the update operation at level l.

[0116] In the above embodiments of the present invention, the detailed operation of S22 is adopted to remove attribute redundancy by utilizing the geometric topological correlation of point clouds, thereby obtaining a compact multi-scale representation and improving coding performance.

[0117] Furthermore, such as Figure 4 As shown, the cross-attention filtering mechanism operates as follows:

[0118] Get the key tensor (K), value tensor (V), and query tensor (Q);

[0119] The initial filtered output is obtained through an H2 head attention mechanism.

[0120]

[0121] Where H2 is a preset positive integer, and W O Represents the learnable parameter matrix;

[0122] The initial filtered output is enhanced by using two sparse convolutional layers and one ReLU nonlinear activation layer. Obtain enhanced filter output As the output of the cross-attention filtering mechanism.

[0123] In this embodiment, H2 is set to 8 to balance performance and computational complexity. In other embodiments, other parameters may be selected. This invention is not limited to the parameters described above.

[0124] The cross-attention filtering mechanism described in the above embodiments of the present invention is specifically designed for point clouds. It utilizes the geometric correlation between the two parts to remove attribute redundancy. Unlike the classic cross-attention filtering mechanism, it uses the geometric coordinates of the two parts as the Q and K values ​​of the attention mechanism, while in typical self-attention mechanisms, Q, K, and V are all attribute values. This embodiment fully utilizes the geometric topological correlation of point clouds to eliminate redundant information of point cloud attributes in the spatial and channel dimensions, generating a compact multi-scale representation and achieving high-fidelity point cloud attribute reconstruction.

[0125] In some preferred embodiments of the present invention, in step S3, obtaining a second boosted representation by passing the first boosted representation through a forward second boosted structure includes:

[0126] The forward second lifting structure concatenates t channel dimensionality-up modules and t channel dimensionality-down modules based on the forward lifting structure, progressively achieving channel scaling. Several optional channel dimensionality-invariant modules based on the lifting structure are embedded after the channel dimensionality-up modules.

[0127] Each module of the forward second lifting structure includes two sparse convolutional layers and a ReLU activation function for prediction and update operations. After prediction and update operations, the predicted / updated values ​​are added to the input signal of each branch.

[0128] The channel dimension-up module based on the forward lifting structure forces the second input branch to zero and merges the two output branches along the channel dimension;

[0129] The channel dimensionality reduction module based on the forward lifting structure divides the input into two parts evenly along the channel dimension and discards the second output branch;

[0130] The channel dimension invariant module based on the lifting structure keeps the two input branches unchanged.

[0131] In the above embodiments of the present invention, the channel dimensionality upscaling module, channel dimensionality reduction module, and channel dimensionality invariant module are all lifting structures that require prediction and updating. Their inputs include two branches, each representing a feature. The features of one branch are used to predict the other branch, and then the predicted branch is used to update the other branch. Here, the two branches refer to the representation features of the two input parts of each module. The first input branch is the first input part of the lifting structure, i.e., the upper branch; the second input branch is the second input part of the lifting structure, i.e., the lower branch. The channel dimensionality upscaling module, channel dimensionality reduction module, and possibly the channel dimensionality invariant module are essentially lifting structures that operate at the channel level. Two sparse convolutional layers and a ReLU activation function constitute the prediction and update operators.

[0132] In this embodiment, the number of blocks t for both the channel upsampling enhancement module and the channel downsampling enhancement module is set to 2. In other embodiments, other parameters can also be selected, and the present invention is not limited to the parameters described above.

[0133] The embodiments of the present invention employ the detailed operation of S3 described above, and remove channel redundancy by progressively scaling the first enhanced multi-scale spatial representation in the channel dimension.

[0134] The lifting structure constructed in the above embodiments of the present invention can efficiently capture irregular geometric structures in point cloud data, while eliminating redundant information of point cloud attributes in spatial and channel dimensions, generating compact multi-scale representations, and realizing high-fidelity point cloud attribute reconstruction.

[0135] Reference Figure 6 As shown, based on the above inventive concept, the present invention provides an embodiment of a point cloud attribute decoding method based on a lifting structure, comprising:

[0136] M1. Obtain the point cloud attribute signal bitstream, process the point cloud attribute signal bitstream to obtain the quantized hidden layer representation;

[0137] M2. Dequantize the quantized hidden layer representation to obtain the reconstructed hidden layer representation;

[0138] M3. The reconstructed hidden layer representation is passed through one or more sparse convolution-based upsampling layers to obtain the reconstructed second boosted representation;

[0139] M4. The reconstructed second lifting representation is passed through a reverse second lifting structure to obtain the reconstructed first lifting representation. The reverse second lifting structure is symmetrical to the forward second lifting structure. t channel dimensionality-up modules and t channel dimensionality-down modules based on the reverse lifting structure are connected in series to achieve channel scaling step by step. Several optional channel dimension-invariant modules based on the lifting structure are embedded after the channel dimensionality-up modules. Each reverse lifting module operates similarly to the preceding lifting module, except that the summation step after the prediction and update operations of each module is replaced with subtraction, that is, the input signal of each branch is subtracted from the predicted value / updated value; here, the predicted value / updated value is the predicted value or the updated value.

[0140] M5. The reconstructed first lift representation is passed through the inverse first lift structure to obtain the reconstructed point cloud attribute signal.

[0141] Specifically, in one embodiment, based on the above steps M1-M5, wherein M1 acquires the point cloud attribute signal bitstream, and the point cloud attribute signal bitstream is processed to obtain a quantized hidden layer representation, wherein the processing operation may include:

[0142] M11. Obtain the point cloud attribute signal bitstream and split the point cloud attribute signal bitstream into a priori information bitstream and a hidden layer representation bitstream;

[0143] M12. Decode the prior information bitstream to obtain the quantized prior information;

[0144] M13. Dequantize the quantized prior information to obtain the reconstructed prior information;

[0145] M14. Quantized hidden layer representation obtained from reconstructed prior information via a priori decoder. The mean μ and scale parameter σ of the Gaussian distribution;

[0146] M15. Entropy decoding of the hidden layer representation bitstream is performed using the mean μ and the scaling parameter σ to obtain the quantized hidden layer representation;

[0147] Then execute M2-M5 until the reconstructed point cloud attribute signal is obtained, thus completing the decoding.

[0148] In some embodiments of the present invention, in the above-described M4, the inverse first lifting structure senses the geometric information of the point cloud, and performs hierarchical adaptive uniform partitioning on the reconstructed first geometric sense lifting representation based on the geometric information C of the point cloud data P to be encoded, thereby obtaining the reconstructed multi-scale first lifting representation. and Following a hierarchical order from high to low, reverse update and reverse prediction operations are performed sequentially to reconstruct the high-dimensional attribute features of the point cloud. And reconstruct the original properties through another sparse convolutional layer.

[0149] Among them, as Figure 4 shown, the reverse spatial enhancement operation for the point cloud attribute signal of any l-th (1 ≤ l < L) layer is as follows:

[0150] Utilize the reconstructed point set 's geometric coordinates residual attribute features and the point set 's geometric coordinates respectively as the key tensor (K), value tensor (V), and query tensor (Q), and adopt the cross-attention filtering mechanism to update the attribute features of the point set

[0151] Subtract the updated value of the attribute features from the initial attribute features of the reconstructed (l + 1)-th layer as the reconstructed attribute features of the l-th layer point set Among them represents the update operation of the l-th layer;

[0152] Utilize the geometric coordinates of the obtained point set reconstructed attribute features and the point set 's geometric coordinates respectively as the key tensor (K), value tensor (V), and query tensor (Q), and adopt the cross-attention filtering mechanism to predict the attribute features of the point set

[0153] Add to the predicted value of the attribute features to obtain the reconstructed attribute features Among them represents the prediction operation of the l-th layer.

[0154] Based on the above inventive concept, as Figure 2 shown, an embodiment of the present invention further provides a point cloud attribute compression method based on a boosting structure, including:

[0155] Obtain the point cloud to be compressed, including the geometric coordinates and attribute signals of each point;

[0156] Adopt the point cloud attribute encoding method based on the boosting structure of any one of the above embodiments to encode the point cloud attribute signal into a point cloud attribute signal bitstream;

[0157] Adopt the point cloud attribute decoding method based on the boosting structure of the above embodiment to decode the point cloud attribute signal bitstream to obtain the reconstructed point cloud attribute signal. ​​​​​

[0158] Furthermore, the parameters of the forward and backward first lifting structures, the forward and backward second lifting structures, the sparse convolution-based downsampling and upsampling layers, and the super-prior encoder and super-prior decoder are obtained through rate-distortion optimization. Specifically, the operations are as follows:

[0159] Step 1: Obtain N voxelized training point cloud data, divide them into slices, and select several slices as the training set;

[0160] The slice division involves selecting r points from all voxelized training set point cloud data using the farthest point sampling algorithm, and using the r points as centers, obtaining k-1 neighbors of the r points through nearest neighbor search to form r slices, where r and k are pre-set positive integers;

[0161] Step 2: Based on the training set, use the point cloud attribute encoding method based on lifting structure from any of the above embodiments to encode the point cloud attribute signals of the training set into a point cloud attribute signal bitstream.

[0162] Step 3: Using the point cloud attribute decoding method based on the lifting structure described above, the point cloud attribute signal stream is decoded and reconstructed into the training set point cloud attribute signal;

[0163] Step 4: For each training set point cloud slice, construct a loss function based on the attribute signal, the reconstructed attribute signal, the geometric coordinates, and the importance score ω of each point. This function is used for end-to-end training to obtain the parameters of the point cloud compression method. The specific form is as follows:

[0164]

[0165] in, It is the loss function used to train the first lifting structure partitioning module. It is the rate-distortion loss function, and γ is used to balance these two loss functions. Specifically:

[0166]

[0167] Here, μ represents the regularization parameter, and c i and c represents the geometric coordinates and reconstructed attribute signal corresponding to the i-th point, respectively; j and ω j These represent the geometric coordinates and importance score of the j-th point, respectively. It is a rate-distortion loss function used to balance bit rate and distortion:

[0168]

[0169] here, Represents the original attribute signal x and the reconstructed attribute signal Distortion between; and These represent the quantized hidden layer representations respectively. and prior information The bit rate; λ is the Lagrange multiplier used to balance the code rate and distortion;

[0170] Step 5: Calculate the loss function for all Nr training set slices. in The loss function for the nth slice of the training set;

[0171] Step 6: Based on the loss function Alternate optimization is used to optimize the importance score ω of the partitioned modules in the first lifting structure and other learnable parameters. Specifically: first, the importance score ω of the partitioned modules is minimized based on the geometric coordinates and the reconstructed attribute signals. Update the importance score ω; secondly, given the significance score ω, minimize the rate-distortion loss function. and Medium attribute fitting loss part The sum of these parameters is used to jointly optimize other learnable parameters.

[0172] Repeat steps two through six iteratively until convergence.

[0173] In this embodiment, r is set to 55, the total number of point cloud training sets N is 3632, k is set to 2048, and a total of 199760 slices are extracted. The regularization parameters γ and μ are set to 50 and 1, respectively, and the Lagrange multipliers λ are set to 30000, 25000, 20000, 15000, 10000, 4000, 3000, 2000, and 1000, respectively, to obtain models with different bit rates. In other embodiments, other parameters can also be selected. This invention is not limited to the parameters mentioned above.

[0174] The techniques used in each step of the above compression method can be referred to the implementation techniques corresponding to the steps in the above embodiment of the point cloud attribute encoding and decoding method for lifting structures, and will not be repeated here.

[0175] Based on the same inventive concept, this embodiment of the invention also provides a point cloud attribute signal compression system, comprising:

[0176] Encoder: The point cloud attribute signal is encoded into a point cloud attribute signal bitstream using the point cloud attribute encoding method based on the lifting structure in any of the above embodiments;

[0177] Decoder: The point cloud attribute signal stream is decoded using the point cloud attribute decoding method based on the lifting structure in any of the above embodiments to obtain the reconstructed point cloud attribute signal.

[0178] The parts of the present invention not specifically described in the above embodiments can be implemented using existing technologies.

[0179] Based on the 3D point cloud compression system described in the above embodiments, the following is a detailed explanation using specific application examples: In this embodiment, the key parameters are set as follows: Following the general test conditions for MPEG G-PCC point cloud compression, the test point cloud data used in the experiment includes 15 human point cloud datasets from MUVB, MPEG CAT1, and Owlii datasets, respectively. The training data is obtained from a large dataset, and there is no overlap between the training data and the test dataset. Since the human eye is more sensitive to changes in image brightness, the original point cloud is converted from the RGB color space to the YUV color space. This embodiment selects four bitrates comparable to the baseline method for fair comparison. The basic model with λ = 30000 is trained for 50 epochs, with a learning rate of 5 × 10⁻⁶ for the first 30 epochs. -5 Adjusted to 10 after the 30th epoch. -5 After the 40th epoch, it was further adjusted to 5×10 -6 The models corresponding to the remaining λ values ​​were obtained by fine-tuning the base model with λ = 30000, and the learning rate during the fine-tuning phase was uniformly set to 10. -6 Each fine-tuning training session lasts for 5 epochs. In this embodiment, the Adam optimizer is used during training, with its exponential decay parameter betas set to 0.9 and 0.999 respectively, and the batch size is 16.

[0180] The system described in the above embodiment of the present invention is used to compress test point cloud data, and the average BD-PSNR and BD-BR are calculated as indicators of compression performance. The larger the BD-PSNR, the better the reconstruction quality of the compression system, and the smaller the BD-BR, the more bitrate the compression system saves.

[0181] Compared to Sparse-PCAC, the system in this embodiment achieves an average BD-PSNR gain of 1.45dB and an average BD-BR of -25.20% across all test data, representing a bit rate saving of 25.20%. Compared to TSC-PCAC, the system in this embodiment achieves an average BD-PSNR gain of 0.73dB and an average BD-BR of -12.60% across all test data, representing a bit rate saving of 12.60%. Compared to the current best MPEG international point cloud compression standard, G-PCC Lift, the system in this embodiment achieves an average BD-PSNR gain of 1.31dB and an average BD-BR of -10.49% across all test data, representing a bit rate saving of 10.49%. Compared to the current best MPEG international point cloud compression standard, G-PCC Lift... Compared to RAHT, the system in this embodiment achieves an average BD-PSNR gain of 0.62dB and an average BD-Rate of -10.49% across all test data, representing a 10.49% reduction in bit rate.

[0182] Experiments show that the compression efficiency of the system in the above embodiments of the present invention is significantly better than that of Sparse-PCAC, TSC-PCAC and the international point cloud compression standard G-PCC.

[0183] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A point cloud attribute encoding method based on lifting structure, characterized in that, include: Obtain the point cloud to be encoded, including the geometric coordinates and attribute signals of each point; Using a forward first lifting structure, a first lifted representation of the attribute signal is obtained based on the geometric coordinates. The first lifting structure achieves spatial adaptive partitioning by sensing the geometric information of the point cloud, and performs attribute prediction and updating by using the correlation of geometric attributes through a cross-attention filtering mechanism. The first boosted representation is passed through the forward boosted structure to obtain the second boosted representation. The second boosted structure consists of one or more pairs of channel dimensionality upscaling and channel dimensionality downscaling modules based on the forward boosted structure. When there are multiple pairs, multiple channel dimensionality upscaling modules are first connected in series, and then multiple connected channel dimensionality downscaling modules are connected in series. The second boosted representation is passed through one or more sparse convolution-based downsampling layers to obtain the hidden layer representation; The hidden layer representation is quantized using a uniform scalar quantizer, and the quantized hidden layer representation is used for encoding to obtain the point cloud attribute signal bitstream. The method of obtaining a first lifted representation of the attribute signal based on the geometric coordinates using a forward first lifting structure includes: The geometric information of the perceived point cloud is used to achieve adaptive spatial partitioning based on the point cloud data to be encoded. Geometric information The points in the point cloud are divided into hierarchical and uniform segments; For the attribute signals of each level after the division, the cross-attention filtering mechanism is designed using geometric attribute correlation, and a forward first lifting operation is performed to obtain the first lifted representation; The geometric information of the perceived point cloud is used to achieve adaptive spatial partitioning, based on the point cloud to be encoded. Geometric information The points in the point cloud are divided into hierarchical, uniform subdivisions, including: Based on input point cloud Geometric information, using The layer shares a multilayer perceptron mechanism and a max pooling layer to extract a The global features of the channels are used to obtain a global geometric feature representation of the input point cloud. The number of channels of the shared multilayer sensor is set to 1. ,in and All are preset positive integers; Then use Layer-sharing multilayer perceptron mechanism for point-by-point extraction The local features of the channel are used to obtain point-by-point local geometric features. The number of channels of the shared multilayer sensor is set to 1. ,in, and All are preset positive integers; The global geometric feature representation of the input point cloud is copied to the same dimension as the point-by-point local geometric feature, and the two are concatenated along the channel dimension to construct a joint geometric feature representation. ; Based on the joint geometric feature representation , use one Head self-attention mechanism The layers share a multilayer perceptron mechanism and an adaptive linear projection layer to estimate the importance score of each point. And it is normalized using the softmax function, where The number of channels in the multilayer shared multilayer perceptron mechanism is set to ,in , and All are preset positive integers; Use the importance score for each point Input point cloud Sort the points by importance score from lowest to highest, and select the top half of the points as the set to be updated. The half of the points with the lowest scores are used as the set of points to be predicted. For the set of points to be updated Based on importance scores, the points are further divided evenly, and the half with the highest scores are selected as the set of points to be updated in the next level. The half of the points with the lowest scores are used as the set of points to be predicted in the next level. Repeat this uniform division process until the preset level is reached. ,get Layer to be updated point set ,as well as Layer of points to be predicted ; The method involves using geometric attribute correlation to design a cross-attention filtering mechanism for the attribute signals at each level after the partitioning, performing a forward first boosting operation to obtain a first boosted representation, including: Extracting input point cloud attribute signals using a sparse convolutional layer High-dimensional features , as the initial attribute features of the first layer; Based on the initial attribute features, the partitioning obtained The point cloud layers are sequentially subjected to forward spatial lifting operations from low to high scale to obtain multi-scale attribute feature representations of the point cloud, including those from layer 1 to layer 2. All layers with resolution halved sequentially Residual attribute feature set of layered point clouds and the Updated attribute features in layered point clouds ; Among them, for any number of The first forward lift operation for the layer point cloud attribute signal is as follows: Based on the geometric coordinates of the two point clouds obtained from the division, the first... Initial attribute characteristics of the layer The points are assigned to two parts of the point cloud to obtain the set of points to be predicted. Attributes and characteristics and points to be updated Attributes and characteristics ;in ; Using the set of points to be updated geometric coordinates and attribute features and the set of points to be predicted geometric coordinates , respectively , and As a key tensor Value tensor and query tensors Construct a cross-attention filtering mechanism to predict the set of points to be predicted. Attributes and characteristics ; from Subtract the predicted value of the attribute feature from the value to obtain the residual attribute feature. ,in Indicates the first Layer prediction operations; Using the obtained set of points to be predicted residual attribute characteristics Geometric coordinates and points to be updated geometric coordinates , respectively , and As a key tensor Value tensor and query tensors Construct a cross-attention filtering mechanism to update the set of points to be updated. Attributes and characteristics ; The updated value of the attribute feature is compared with Add them together, as the first... Initial attribute characteristics of the layer ,in Indicates the first Layer update operations.

2. The point cloud attribute encoding method based on lifting structure according to claim 1, characterized in that, The cross-attention filtering mechanism operates as follows: Get key tensor Value tensor and query tensors ; Through a The head attention mechanism obtains the initial filtered output. : ； in, It is a preset positive integer. , , and Represents the learnable parameter matrix; The initial filtered output is enhanced using two sparse convolutional layers and one ReLU nonlinear activation layer. To obtain enhanced filter output This serves as the output of the cross-attention filtering mechanism.

3. The point cloud attribute encoding method based on lifting structure according to claim 1, characterized in that, The first lifting representation is passed through a forward second lifting structure to obtain a second lifting representation, including: The forward second lifting structure includes a series of sequentially connected components. A channel dimensionality-upgrading module based on a forward boosting structure and A channel dimensionality reduction module based on a forward lifting structure is used to progressively scale channels. Several optional channel dimension-invariant modules based on lifting structures are embedded between the channel dimensionality increase and reduction modules. Each lifting structure contains two branches. Each module of the forward second lifting structure includes two sparse convolutional layers and a ReLU activation function in its prediction and update operations. After the prediction and update operations, the predicted / updated values ​​are added to the input signal of each branch. The channel dimension-up module based on the forward lifting structure forces the second input branch to zero and merges the two output branches along the channel dimension; The channel dimensionality reduction module based on the forward lifting structure divides the input into two parts evenly along the channel dimension and discards the second output branch; The channel dimension invariant module based on the lifting structure keeps the two input branches unchanged.

4. A point cloud attribute decoding method based on lifting structure, used to decode the encoding obtained by any one of claims 1-3, characterized in that, include: Obtain the point cloud attribute signal bitstream, and process the point cloud attribute signal bitstream to obtain a quantized hidden layer representation; The quantized hidden layer representation is dequantized to obtain the reconstructed hidden layer representation; The reconstructed hidden layer representation is passed through one or more sparse convolution-based upsampling layers to obtain a reconstructed second boosted representation; The reconstructed second lifting representation is passed through a reverse second lifting structure to obtain the reconstructed first lifting representation. The reverse second lifting structure is symmetrical to the forward second lifting structure and is connected in series. A channel-level upscaling module based on a reverse boosting structure and A channel dimensionality reduction module based on an inverse lifting structure is used to progressively scale the channels, and several optional channel dimension-invariant modules based on lifting structures are embedded after the channel dimensionality increase module. Each reverse boosting module operates similarly to the preceding boosting module, except that the summation step after the prediction and update operations of each module is replaced with subtraction, that is, the input signal of each branch is subtracted from the predicted value / updated value; The reconstructed first lifting representation is passed through a reverse first lifting structure to obtain the reconstructed point cloud attribute signal.

5. A point cloud attribute compression method based on lifting structure, characterized in that, include: Obtain the point cloud to be compressed, including the geometric coordinates and attribute signals of each point; The point cloud attribute encoding method based on lifting structure as described in any one of claims 1-3 is used to encode the point cloud attribute signal into a point cloud attribute signal bitstream. The point cloud attribute decoding method based on the lifting structure described in claim 4 is used to decode the point cloud attribute signal stream to obtain the reconstructed point cloud attribute signal.

6. The point cloud attribute compression method based on lifting structure according to claim 5, characterized in that, The point cloud attribute encoding method and the point cloud attribute decoding method based on the lifting structure are obtained by optimizing the pass-through distortion of the forward and backward first lifting structures, the forward and backward second lifting structures, the downsampling and upsampling layers based on sparse convolution, and the parameters of the super-prior encoder and super-prior decoder. The specific operations are as follows: Step 1: Obtain The training set point cloud data is voxelized and then sliced, with several slices selected as the training set. Slicing involves selecting the farthest point from all voxelized training set point cloud data using a farthest point sampling algorithm. points, with the aforementioned Centered on a point, the nearest neighbor search is used to obtain the... points Neighbors, forming A slice, and A pre-defined positive integer; Step 2: Based on the training set, use the point cloud attribute encoding method based on lifting structure as described in any one of claims 1-3 to encode the point cloud attribute signals of the training set into a point cloud attribute signal bitstream; Step 3: Using the point cloud attribute decoding method based on lifting structure as described in claim 4, the point cloud attribute signal code stream is decoded and reconstructed into the training set point cloud attribute signal; Step 4: For each training set point cloud slice, perform analysis based on the attribute signal, the reconstructed attribute signal, the geometric coordinates, and the importance score of each point. Construct a loss function to obtain parameters for the point cloud compression method during end-to-end training. The specific form is as follows: ； in, It is the loss function used to train the first lifting structure partitioning module. It is the rate-distortion loss function. To balance these two loss functions, specifically: ； here, Represents the regularization parameter. and They respectively represent the corresponding number The geometric coordinates of each point and the reconstructed attribute signals; It is a rate-distortion loss function used to balance bit rate and distortion: ， here, Represents the original attribute signal With reconstructed attribute signals Distortion between; and These represent the quantized hidden layer representations respectively. and prior information bit rate; It is a Lagrange multiplier used to balance bit rate and distortion; Step 5: Calculate all Loss function for each slice of the training set ,in For the first Loss function for each slice of the training set; Step 6: Based on the loss function Alternate optimization of the importance score of the partitioned modules in the first boost structure Other learnable parameters; Repeat steps two through six iteratively until convergence.

7. A point cloud attribute signal compression system, comprising an encoder and a decoder, characterized in that: Encoder: The point cloud attribute encoding method based on lifting structure as described in any one of claims 1-3 is used to encode the point cloud attribute signal into a point cloud attribute signal bitstream; Decoder: The point cloud attribute signal stream is decoded using the point cloud attribute decoding method based on lifting structure as described in claim 4 to obtain the reconstructed point cloud attribute signal.

8. A terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it can be used to execute the point cloud attribute encoding method based on lifting structure according to any one of claims 1-3, or the point cloud attribute decoding method based on lifting structure according to claim 4, or the point cloud attribute compression method based on lifting structure according to any one of claims 5-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, this program can be used to perform the point cloud attribute encoding method based on lifting structure as described in any one of claims 1-3, or the point cloud attribute decoding method based on lifting structure as described in claim 4, or the point cloud attribute compression method based on lifting structure as described in any one of claims 5-6.

Images