Point Cloud Attribute Compression Method, Device and Storage Medium Based on Deep Entropy Coding

Abstract

The present invention discloses a point cloud attribute compression method, a computer device and a storage medium based on deep entropy coding, including obtaining the attribute information and geometric information of point cloud data, performing transform coding on the attribute information to obtain transform coefficients, recording the intermediate information generated during the transform coding process, performing point-by-point feature extraction on the geometric information and transform coefficients to obtain point cloud information features, performing point-by-point feature extraction on the intermediate information to obtain transform coding information features, fusing the point cloud information features and transform coding information features and then performing probability prediction to obtain a probability distribution, and performing entropy coding on the transform coefficients and the probability distribution to obtain a compressed file and other steps. By extracting point cloud information such as attribute information and geometric information and transform coding information, the present invention is beneficial to symbol probability distribution prediction, can process complex and diverse point cloud data, better realizes symbol probability prediction, and improves the compression performance. The present invention is widely applied to the technical field of point cloud data processing.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud data processing, and in particular to a point cloud attribute compression method, a computer device, and a storage medium based on depth entropy coding. Background Art

[0002] A point cloud is a three-dimensional data structure that is widely used in practical applications including virtual reality, autonomous driving, and high-precision mapping. In recent years, with the development of three-dimensional data acquisition technology, the amount of point cloud obtained has been increasing. Correspondingly, point cloud data compression has become the key to point cloud storage and transmission, and is also a necessary basis for many three-dimensional vision applications.

[0003] The related technologies of point cloud data compression include transform coding and entropy coding. In existing approximate implementation schemes, the entropy coding part is mainly completed by artificially estimating symbol probabilities or based on specific rules. In the method of artificially estimating symbol probabilities, it is assumed that the symbols follow a specific distribution (such as Laplace distribution, normal distribution, etc.), the parameters of the distribution (mean, variance, etc.) are obtained through statistics, the probability prediction is realized using this distribution, and the prediction probability is combined with an entropy encoder to realize entropy coding; in the method based on specific rules, the input symbols are transformed at multiple levels based on specific rules, such as first performing run-length coding on the symbols and then performing entropy coding, etc. Run-length coding refers to detecting repeated sequences of bits or characters and replacing them with their occurrence times. Existing technologies are usually only developed and tested on dense point cloud data, and cannot take into account the compression of multiple types of point clouds (such as dense point clouds, LiDAR point clouds, etc.). The performance is low when compressing complex point cloud data containing multiple types of point clouds. Summary of the Invention

[0004] Aiming at at least one technical problem such as the current point cloud data compression technology being unable to take into account the compression of multiple types of point clouds and having low compression performance, the purpose of the present invention is to provide a point cloud attribute compression method, a computer device, and a storage medium based on depth entropy coding.

[0005] On the one hand, an embodiment of the present invention includes a point cloud attribute compression method based on depth entropy coding, including:

[0006] Obtain point cloud data;

[0007] Obtain the attribute information and geometric information of the point cloud data;

[0008] Use a transform coding algorithm to perform transform coding on the attribute information to obtain transform coefficients;

[0009] Record the intermediate information generated during the process of the transform coding;

[0010] Perform point-by-point feature extraction on the geometric information and the transform coefficients to obtain point cloud information features;

[0011] Extract point-by-point features from the intermediate information to obtain transformed coding information features;

[0012] Perform probability prediction after fusing the point cloud information features and the transformed coding information features to obtain a probability distribution;

[0013] Perform entropy coding on the transformation coefficients and the probability distribution to obtain a compressed file.

[0014] Further, the transformation coding algorithm is a three-dimensional wavelet transform algorithm.

[0015] Further, the performing transformation coding on the attribute information includes:

[0016] Voxelize the attribute information to obtain a plurality of point cloud attribute voxels;

[0017] Arrange each of the point cloud attribute voxels;

[0018] Set a transformation direction;

[0019] Starting from the point cloud attribute voxel at the initial position in the transformation direction, perform multi-level transformations in sequence; in each level of transformation, for two adjacent point cloud attribute voxels, generate high-frequency components and low-frequency components according to the two adjacent point cloud attribute voxels, and transfer the high-frequency components and the low-frequency components to the next level of transformation as the point cloud attribute voxels in the next level of transformation. For the point cloud attribute voxel that is not adjacent to other point cloud attribute voxels, transfer the point cloud attribute voxel to the next level of transformation as the point cloud attribute voxel in the next level of transformation;

[0020] Record the high-frequency components and low-frequency components obtained from the last level of transformation as the transformation coefficients.

[0021] Further, the formula used for generating high-frequency components and low-frequency components according to two adjacent point cloud attribute voxels, and transferring the high-frequency components and the low-frequency components to the next level of transformation as the point cloud attribute voxels in the next level of transformation is:

[0022]

[0023] where d represents the level of transformation, x, y, z represent the transformation direction, l d,2x,y,z represents the low-frequency component of one of the two adjacent point cloud attribute voxels, l d,2x+1,y,z represents the low-frequency component of the other of the two adjacent point cloud attribute voxels, l d+1,x,y,z represents the low-frequency component of the point cloud attribute voxel in the next level of transformation, h d+1,x,y,zRepresents the high-frequency component in the voxel of the point cloud attribute in the next-level transformation, w 1 Represents the weight corresponding to the transformation direction 2x, y, z in the d-level transformation, w 2 Represents the weight corresponding to the transformation direction 2x + 1, y, z in the d-level transformation.

[0024] Further, the point cloud attribute compression method based on depth entropy coding further includes:

[0025] Before using the transform coding algorithm to perform transform coding on the attribute information, perform denoising, chunking, and color space transformation processing on the attribute information;

[0026] After obtaining the transform coefficients, perform structuring processing on the transform coefficients.

[0027] Further, the obtaining of the transform coding information feature by performing per-point feature extraction on the intermediate information includes:

[0028] Obtain the weight w and the low-frequency component l in each level of transformation as the intermediate information;

[0029] Use a multi-layer fully connected neural network MLP to perform per-point feature extraction on the weight w and the low-frequency component l to obtain the first per-point feature f Pt = MLPs(w, l);

[0030] Use a three-dimensional neural network 3DConv to perform feature aggregation of neighboring points on the weight w and the low-frequency component l to obtain the first spatial feature f St = 3DConv(w, l);

[0031] The first per-point feature f Pt and the first spatial feature f St are concatenated along the channel dimension to obtain the transform coding information feature f trans = MLPs(f Pt , f St ).

[0032] Further, the obtaining of the point cloud information feature by performing per-point feature extraction on the geometric information and the transform coefficients includes:

[0033] Use a multi-layer fully connected neural network MLP to perform per-point feature extraction on the geometric information xyz and the transform coefficient a to obtain the second per-point feature f P = MLPs(xyz, a);

[0034] Use a three-dimensional neural network 3DConv to perform feature aggregation of neighboring points on the geometric information xyz and the transform coefficient a to obtain the second spatial feature f S= 3DConv(xyz,a);

[0035] Concatenate the second point - wise feature f P with the second spatial feature f S along the channel dimension to obtain the point - cloud information feature f pc = MLPs(f P , f S ).

[0036] Furthermore, after fusing the point - cloud information feature and the transformed coding information feature, perform probability prediction to obtain a probability distribution, including:

[0037] Use a multi - layer fully - connected neural network MLP to fuse the point - cloud information feature f pc and the transformed coding information f trans to obtain feature information f = MLPs(f pc , f trans );

[0038] Use a multi - layer fully - connected neural network MLP to regress the probability distribution P = MLPs(f) from the feature information f.

[0039] On the other hand, an embodiment of the present invention further includes a computer device, including a memory and a processor. The memory is used to store at least one program, and the processor is used to load the at least one program to execute the point - cloud attribute compression method based on depth entropy coding in the embodiment.

[0040] On the other hand, an embodiment of the present invention further includes a storage medium, in which a program executable by a processor is stored. The program executable by the processor is used to execute the point - cloud attribute compression method based on depth entropy coding in the embodiment when executed by the processor.

[0041] The beneficial effects of the present invention are as follows: The point - cloud attribute compression method based on depth entropy coding in the embodiment can fit the symbol probability distribution through neural network deep learning. By extracting point - cloud information such as attribute information and geometric information as well as transformed coding information, it is beneficial for symbol probability distribution prediction. Compared with the current related technologies, the point - cloud attribute compression method based on depth entropy coding in this embodiment makes better use of the information implicitly contained in the data itself, can process complex and diverse point - cloud data, reduces the expert design cost, better realizes symbol probability prediction, and improves the compression performance. Brief Description of the Drawings

[0042] Figure 1 It is a flowchart of the point - cloud attribute compression method based on depth entropy coding in the embodiment;

[0043] Figure 2Schematic diagram of the point cloud attribute compression method based on depth entropy coding in the embodiment;

[0044] Figure 3 Schematic diagram of the 3D wavelet transform algorithm used in the embodiment. Specific implementation manner

[0045] In this embodiment, referring to Figure 1 , the point cloud attribute compression method based on depth entropy coding includes the following steps:

[0046] S1. Obtain point cloud data;

[0047] S2. Obtain the attribute information and geometric information of the point cloud data;

[0048] S3. Use a transform coding algorithm to perform transform coding on the attribute information to obtain transform coefficients;

[0049] S4. Record the intermediate information generated during the transform coding process;

[0050] S5. Perform point-by-point feature extraction on the geometric information and transform coefficients to obtain point cloud information features;

[0051] S6. Perform point-by-point feature extraction on the intermediate information to obtain transform coding information features;

[0052] S7. Perform probability prediction after fusing the point cloud information features and transform coding information features to obtain a probability distribution;

[0053] S8. Perform entropy coding on the transform coefficients and probability distribution to obtain a compressed file.

[0054] The principles of steps S1 - S8 are as shown in Figure 2 . Figure 2 The dashed boxes in it represent the data to be processed, the data in the intermediate process, or the data finally processed, and the solid boxes represent the modules for processing data, and these modules can be software modules or hardware modules with corresponding functions.

[0055] In step S2, by parsing the point cloud data, the attribute information and geometric information of the point cloud data are obtained. Among them, the attribute information of the point cloud data includes information such as the color and reflectivity of the point cloud for describing its optical characteristics, and the geometric information of the point cloud data includes information such as the x, y, and z coordinates of the point cloud.

[0056] In step S3, referring to Figure 2 , the transform coding algorithm is executed through the transform coding module to perform transform coding on the attribute information to obtain transform coefficients. Specifically, an autoencoder constructed based on a deep neural network can be used to execute the transform coding algorithm, or the 3D wavelet transform algorithm (RAHT) can be used as the transform coding algorithm.

[0057] In this embodiment, before performing step S3, that is, using a transform coding algorithm to perform transform coding on the attribute information, the attribute information can be preprocessed first, where the preprocessing includes at least one processing process such as denoising, chunking, and color space transformation processing. After preprocessing the attribute information, step S3 is performed, that is, using a transform coding algorithm to perform transform coding on the preprocessed attribute information.

[0058] When using the three-dimensional wavelet transform algorithm as the transform coding algorithm, when performing the step of performing transform coding on the attribute information in step S3, the following steps can be specifically performed:

[0059] S301. Voxelize the attribute information to obtain multiple point cloud attribute voxels;

[0060] S302. Arrange each point cloud attribute voxel;

[0061] S303. Set the transform direction;

[0062] S304. Starting from the point cloud attribute voxel at the initial position in the transform direction, perform multiple-level transforms in sequence; in each level of transform, for two adjacent point cloud attribute voxels, generate high-frequency components and low-frequency components according to the two adjacent point cloud attribute voxels, and transfer the high-frequency components and low-frequency components to the next level of transform as the point cloud attribute voxels in the next level of transform. For a point cloud attribute voxel that is not adjacent to other point cloud attribute voxels, transfer the point cloud attribute voxel to the next level of transform as the point cloud attribute voxel in the next level of transform;

[0063] S305. Record the high-frequency components and low-frequency components obtained from the last level of transform as transform coefficients.

[0064] The principles of steps S301 - S305 are as Figure 3 shown.

[0065] In step S301, the attribute information can be constructed into a binary tree according to the spatial partitioning relationship of RAHT. For easy display, Figure 3 a simple two-dimensional example is used to illustrate the construction method of RAHT and its corresponding binary tree, and this method is also applicable to the three-dimensional case. Referring to Figure 3 (a), RAHT first voxelizes the attribute information to obtain multiple point cloud attribute voxels such as l 1 、l 2 and l 3 etc.

[0066] In step S302, for l 1 、l 2 and l 3The voxels of the point cloud attributes are arranged so that they have a certain positional relationship, such as Figure 3 The leftmost side of (a) shows l 1 , l 2 With l 3 A positional relationship between voxels with equal point cloud attributes.

[0067] In step S303, the transformation directions such as x, y, and z are set. Figure 3 (a) shows the two transformation directions, x and y.

[0068] In step S304, transformations are performed along the transformation directions of x, y, z, etc., starting from the point cloud attribute voxel located at the initial position in the transformation direction, and multi-level transformations are performed in sequence. In each level of transformation, if there are adjacent points, low-frequency and high-frequency components are generated. If there are no adjacent points, the low-frequency components are directly transmitted to the next level.

[0069] For example, refer to Figure 3 (a) The first level of transformation is performed along the x direction. First, find the point cloud attribute voxel l at the initial position in the x direction. 1 , point cloud attribute voxel l 1 No adjacent points, that is, point cloud attribute voxel l 1 is not adjacent to other point cloud attribute voxels, then the point cloud attribute voxel l 1 Passed to the next level of transformation; and the point cloud attribute voxel l 2 With point cloud attribute voxel l 3 Adjacent, then according to the point cloud attribute voxel l 2 With point cloud attribute voxel l 3 Generate low frequency component l 4 With high frequency component h 1 , the low frequency component l 4 With high frequency component h 1 Passed to the next level of transformation as the point cloud attribute voxel in the next level of transformation; the second level of transformation is performed along the y direction. In the second level of transformation, the point cloud attribute voxel l 1 With low frequency component l 4 Adjacent, according to the point cloud attribute voxel l 1 With low frequency component l 4 Generate low frequency component l 5 With high frequency component h 2 . Correspondingly, Figure 3 As shown in (b), a binary tree can be constructed according to the structure of hierarchical transformation. If there are adjacent points, the nodes are aggregated and the parent node is generated. If there are no adjacent points, the parent node is directly generated. Transform step by step in each direction until all points are aggregated and the root node is generated.

[0070] When performing the step of generating high-frequency components and low-frequency components based on two adjacent voxelized point cloud attributes in step S304, and transmitting the high-frequency components and low-frequency components to the next-level transformation as the voxelized point cloud attributes in the next-level transformation, the following formula can be specifically used for calculation.

[0071] In this formula, d represents the level of transformation, that is, l d,2x,y,z and l d,2x+1,y,z represent the data obtained from the d-th level of transformation, l d+1,x,y,z and h d+1,x,y,z represent the data obtained from the (d + 1)-th level of transformation; x, y, z represent the transformation directions, l d,2x,y,z represents the low-frequency component of one of the two adjacent voxelized point cloud attributes, l d,2x+1,y,z represents the low-frequency component of the other of the two adjacent voxelized point cloud attributes, l d+1,x,y,z represents the low-frequency component of the voxelized point cloud attributes in the next-level transformation, h d+1,x,y,z represents the high-frequency component of the voxelized point cloud attributes in the next-level transformation, w 1 represents the weight corresponding to the transformation direction 2x, y, z in the d-th level of transformation, w 2 represents the weight corresponding to the transformation direction 2x + 1, y, z in the d-th level of transformation.

[0072] When the binary tree shown in Figure 3 (b) is established, w 1 can also represent the weight (number of leaf nodes) of the node corresponding to l d,2x,y,z in the binary tree, and w 2 can also represent the weight (number of leaf nodes) of the node corresponding to l d,2x+1,y,z in the binary tree.

[0073] After all levels of transformation are completed in step S304, in step S305, the high-frequency components and low-frequency components obtained from the last level of transformation are recorded as the transformation coefficients obtained by performing step S3.

[0074] In this embodiment, after obtaining the transformation coefficients by performing step S3, the transformation coefficients can also be structurally processed to organize the disordered transformation coefficients into structured data, which is conducive to the extraction of subsequent transform coding information.

[0075] In step S4, the intermediate information generated during the process of transform coding is recorded. Specifically, the intermediate information generated during the process of transform coding includes the node low-frequency components, the structured information of the binary tree (such as the weight of each node, that is, the number of leaf nodes), the voxelized point cloud attributes obtained by reconstructing the transform coding, and the node spatial positions, etc.

[0076] When performing step S5, that is, extracting point-by-point features from geometric information and transformation coefficients to obtain point cloud information features, the following steps can be specifically performed:

[0077] S501. Use a multi-layer fully connected neural network MLP to perform point-by-point feature extraction on geometric information xyz and transformation coefficient a to obtain a second point-by-point feature f P = MLPs(xyz, a);

[0078] S502. Use a three-dimensional neural network 3DConv to perform feature aggregation of neighborhood points on geometric information xyz and transformation coefficient a to obtain a second spatial feature f S = 3DConv(xyz, a);

[0079] S503. Concatenate the second point-by-point feature f P and the second spatial feature f S along the channel dimension to obtain point cloud information feature f pc = MLPs(f P , f S ).

[0080] When performing step S6, that is, extracting point-by-point features from intermediate information to obtain transformation coding information features, the following steps can be specifically performed:

[0081] S601. Obtain the weight w and low-frequency component l in each level of transformation as intermediate information;

[0082] S602. Use a multi-layer fully connected neural network MLP to perform point-by-point feature extraction on weight w and low-frequency component l to obtain a first point-by-point feature f Pt = MLPs(w, l);

[0083] S603. Use a three-dimensional neural network 3DConv to perform feature aggregation of neighborhood points on weight w and low-frequency component l to obtain a first spatial feature f St = 3DConv(w, l);

[0084] S604. Concatenate the first point-by-point feature f Pt and the first spatial feature f St along the channel dimension to obtain transformation coding information feature f trans = MLPs(f Pt , f St ).

[0085] In step S601, only two kinds of information, weight w and low-frequency component l, are used as intermediate information. More intermediate information generated during the transformation coding process (such as other intermediate information represented as m) can also be used to join the first point-by-point feature f PtWith the first spatial feature f St in the calculation process, for example, through f Pt = MLPs(w, l, m), the first pointwise feature f is calculated Pt , and through f St = 3DConv(w, l, m), the first spatial feature f is calculated St .

[0086] When performing step S7, that is, after fusing the point cloud information feature and the transformed encoding information feature and then performing probability prediction to obtain the probability distribution, the following steps can be specifically executed:

[0087] S701. Use a multi-layer fully connected neural network MLP to fuse the point cloud information feature f pc and the transformed encoding information f trans to obtain the feature information f = MLPs(f pc , f trans );

[0088] S702. Use a multi-layer fully connected neural network MLP to regress the probability distribution P = MLPs(f) for the feature information f

[0089] In this embodiment, in addition to directly regressing the probability distribution of symbols by performing the above steps S701 - S702, that is, using a multi-layer fully connected neural network, it is also possible to use a neural network to regress the probability density function of symbols. Specifically, for a finite number of input symbols, that is, the transform coefficients h, the multi-layer fully connected neural network MLPs can be used to directly regress their probability values P, obtaining P = MLPs(f).

[0090] In step S8, referring to Figure 2 , use an entropy encoder to perform entropy encoding on the transform coefficients and the probability distribution to obtain a compressed file, and this compressed file contains the compressed information related to the point cloud data

[0091] In this embodiment, through deep learning, the fitting of the symbol probability distribution is realized. Corresponding to different test scenarios, such as dense human point clouds or sparse LiDAR point clouds, etc., in the training stage, corresponding data sets can be designed and collected for training. For example, for dense human point clouds, a corresponding human point cloud training set can be constructed first, and for sparse LiDAR point clouds, a LiDAR point cloud training set can be constructed first. During the training process, since the probability distribution of the transform coefficients (input symbols) in the training set is known, loss functions including but not limited to cross-entropy can be used to fit the predicted probability distribution and the true probability distribution. Compared with the prior art, this strategy enables the point cloud attribute compression method based on deep entropy encoding in this embodiment to process complex and diverse point cloud data such as dense point clouds and LiDAR point clouds existing simultaneously

[0092] The point cloud attribute compression method based on deep entropy coding in this embodiment can fit the symbol probability distribution through the neural network deep learning method, while the current related technologies generally use the method based on manual rules to predict the probability. Compared with the current related technologies, the point cloud attribute compression method based on deep entropy coding in this embodiment makes better use of the information implicitly contained in the data itself, can process complex and diverse point cloud data, reduces the expert design cost, and improves the compression performance. The point cloud attribute compression method based on deep entropy coding in this embodiment extracts point cloud information such as attribute information and geometric information, which is beneficial to the prediction of the symbol probability distribution. However, the current related technologies do not effectively utilize the point cloud information. Compared with the current related technologies, the point cloud attribute compression method based on deep entropy coding in this embodiment better realizes the symbol probability prediction, thus improving the compression performance. The point cloud attribute compression method based on deep entropy coding in this embodiment extracts the transform coding information, which is beneficial to the prediction of the symbol probability distribution. However, the current related technologies do not effectively utilize this information. Compared with the current related technologies, the point cloud attribute compression method based on deep entropy coding in this embodiment better realizes the symbol probability prediction, thus improving the compression performance.

[0093] A computer program implementing the point cloud attribute compression method based on deep entropy coding in this embodiment can be written, and this computer program can be written into a computer device or a storage medium. When the computer program is read and run, it executes the point cloud attribute compression method based on deep entropy coding in this embodiment, thereby achieving the same technical effects as the point cloud attribute compression method based on deep entropy coding in the embodiment.

[0094] It should be noted that, unless otherwise specified, when a certain feature is referred to as "fixed" or "connected" to another feature, it can be directly fixed or connected to the other feature, or indirectly fixed or connected to the other feature. In addition, the up, down, left, right, etc. descriptions used in this disclosure are only relative to the mutual positional relationship of the various components of this disclosure in the drawings. The singular forms "a", "the" and "said" used in this disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. In addition, unless otherwise defined, all the technical and scientific terms used in this embodiment have the same meanings as those commonly understood by those skilled in the technical field of this technology. The terms used in the specification of this embodiment are only for describing specific embodiments, rather than for limiting the present invention. The term "and / or" used in this embodiment includes any combination of one or more of the related listed items.

[0095] It should be understood that although the terms first, second, third, etc. may be used in this disclosure to describe various elements, these elements should not be limited to these terms. These terms are only used to distinguish elements of the same type from each other. For example, without departing from the scope of this disclosure, the first element may also be referred to as the second element, and similarly, the second element may also be referred to as the first element. The use of any and all examples or exemplary language ("for example", "such as", etc.) provided in this embodiment is only intended to better illustrate the embodiments of the present invention and will not impose a limitation on the scope of the present invention unless otherwise required.

[0096] It should be recognized that embodiments of the present invention can be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable memory. The methods can be implemented in a computer program using standard programming techniques - including a non-transitory computer-readable storage medium configured with the computer program, wherein the storage medium so configured causes the computer to operate in a specific and predefined manner - according to the methods and drawings described in the specific embodiments. Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system. However, if desired, the program can be implemented in assembly or machine language. In any case, the language can be a compiled or interpreted language. In addition, for this purpose the program is capable of running on a programmed application-specific integrated circuit.

[0097] Furthermore, the operations of the processes described in this embodiment can be performed in any suitable order, unless this embodiment otherwise indicates or is otherwise clearly inconsistent with the context. The processes described in this embodiment (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions and can be implemented as code (e.g., executable instructions, one or more computer programs, or one or more applications) executed commonly on one or more processors, by hardware, or a combination thereof. The computer program includes a plurality of instructions executable by one or more processors.

[0098] Further, the method can be implemented in any type of computing platform operably connected, including but not limited to personal computers, minicomputers, mainframes, workstations, network or distributed computing environments, separate or integrated computer platforms, or communicating with charged particle tools or other imaging devices, etc. Aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into the computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer and used to configure and operate the computer to perform the processes described herein when the storage medium or device is read by the computer. Additionally, the machine-readable code, or portions thereof, can be transmitted via a wired or wireless network. When such media includes instructions or programs that implement the above-described steps in conjunction with a microprocessor or other data processor, the invention as described in this embodiment includes these and other different types of non-transitory computer-readable storage media. When programmed according to the methods and techniques of the present invention, the present invention also includes the computer itself.

[0099] A computer program can be applied to input data to perform the functions described in this embodiment, thereby transforming the input data to generate output data stored in non-volatile memory. The output information can also be applied to one or more output devices such as a display. In a preferred embodiment of the present invention, the transformed data represents physical and tangible objects, including a specific visual depiction of the physical and tangible objects generated on the display.

[0100] As described above, these are only the preferred embodiments of the present invention, and the present invention is not limited to the above-described embodiments. As long as the same means are used to achieve the technical effects of the present invention, any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention. Within the scope of protection of the present invention, its technical solutions and / or implementation manners can have various different modifications and changes.

Claims

1. A point cloud attribute compression method based on depth entropy coding, characterized in that, the point cloud attribute compression method based on depth entropy coding includes: Obtain point cloud data; Obtain the attribute information and geometric information of the point cloud data; Use a transform coding algorithm to perform transform coding on the attribute information to obtain transform coefficients; Record the intermediate information generated during the transform coding process; Perform point-by-point feature extraction on the geometric information and the transform coefficients to obtain point cloud information features; Perform point-by-point feature extraction on the intermediate information to obtain transform coding information features; Perform probability prediction after fusing the point cloud information features and the transform coding information features to obtain a probability distribution; Perform entropy coding on the transform coefficients and the probability distribution to obtain a compressed file; The performing transform coding on the attribute information includes: Voxelize the attribute information to obtain a plurality of point cloud attribute voxels; Arrange each of the point cloud attribute voxels; Set a transform direction; Starting from the point cloud attribute voxel at the initial position in the transform direction, perform multi-level transforms in sequence; in each level of transform, for two adjacent point cloud attribute voxels, generate high-frequency components and low-frequency components according to the two adjacent point cloud attribute voxels, and transfer the high-frequency components and the low-frequency components to the next level of transform as the point cloud attribute voxels in the next level of transform. For the point cloud attribute voxel that is not adjacent to other point cloud attribute voxels, transfer the point cloud attribute voxel to the next level of transform as the point cloud attribute voxel in the next level of transform; Record the high-frequency components and low-frequency components obtained from the last level of transform as the transform coefficients; Generate high-frequency components and low-frequency components based on two adjacent point cloud attribute voxels, and transfer the high-frequency components and the low-frequency components to the next-level transformation as the point cloud attribute voxels in the next-level transformation. The formula used is: ; Among them, represents the order of the transformation, represents the transformation direction, represents the low-frequency component of one of the two adjacent point cloud attribute voxels, represents the low-frequency component of the other of the two adjacent point cloud attribute voxels, represents the low-frequency component in the point cloud attribute voxel in the next-level transformation, represents the high-frequency component in the point cloud attribute voxel in the next-level transformation, represents the weight corresponding to the transformation direction in the -th level transformation, represents the weight corresponding to the transformation direction in the The performing point-by-point feature extraction on the intermediate information to obtain transform coding information features includes: Obtain the weights in each level of transformation and the low-frequency components as the intermediate information; Using a multi-layer fully-connected neural network For the said weights and the said low-frequency components Perform point-by-point feature extraction to obtain the first point-by-point feature ; Using a three-dimensional neural network For the said weights and the said low-frequency components perform feature aggregation of domain points to obtain a first spatial feature ; Concatenate the first point-by-point feature with the first spatial feature along the channel dimension to obtain the transformed coding information feature .

2. The point cloud attribute compression method based on depth entropy coding according to claim 1, characterized in that, the transform coding algorithm is a three-dimensional wavelet transform algorithm.

3. The point cloud attribute compression method based on depth entropy coding according to claim 1, characterized in that, the point cloud attribute compression method based on depth entropy coding further includes: Before using the transform coding algorithm to perform transform coding on the attribute information, perform denoising, chunking, and color space transformation processing on the attribute information; After obtaining the transform coefficients, perform structuring processing on the transform coefficients.

4. The point cloud attribute compression method based on depth entropy coding according to any one of claims 1-3, characterized in that, the performing point-by-point feature extraction on the geometric information and the transform coefficients to obtain point cloud information features includes: Use a multi-layer fully connected neural network For the geometric information and the transformation coefficients Perform point-by-point feature extraction to obtain the second point-by-point feature ; Using a three-dimensional neural network For the geometric information and the transformation coefficients perform feature aggregation of domain points to obtain a second spatial feature ; Combine the second point-by-point feature with the second spatial feature and splice them along the channel dimension to obtain the point cloud information feature .

5. The point cloud attribute compression method based on depth entropy coding according to any one of claims 1-3, characterized in that, the performing probability prediction after fusing the point cloud information features and the transform coding information features to obtain a probability distribution includes: Use a multi-layer fully connected neural network For the feature of the point cloud information And the transform coding information Perform feature fusion to obtain feature information ; Use a multi-layer fully connected neural network For the feature information Regress the probability distribution .

6. A computer device, characterized in that, it includes a memory and a processor. The memory is used to store at least one program, and the processor is used to load the at least one program to execute the point cloud attribute compression method based on depth entropy coding according to any one of claims 1-5.

7. A storage medium storing a program executable by a processor, characterized in that, the program executable by the processor, when executed by the processor, is used to execute the point cloud attribute compression method based on depth entropy coding according to any one of claims 1-5.

Images