A resolution-adaptive geometric lossy coding method, apparatus, and medium for point clouds based on deep residual compression and sparse representation.
By employing a resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation, the problems of adaptability and insufficient utilization of latent variables in large-scale point cloud coding are solved, achieving efficient point cloud compression and reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SUN YAT SEN UNIV
- Filing Date
- 2022-11-15
- Publication Date
- 2026-06-30
AI Technical Summary
Existing point cloud compression methods struggle to effectively encode large-scale point clouds, lack resolution adaptability, fail to fully utilize latent variable information representation, have limited receptive fields, and are unable to handle relationships between distant points.
A resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation is adopted. Through multiple rounds of encoding and decoding, a neural network module is used for feature residual and mask encoding, combined with an arithmetic coding algorithm, to achieve point cloud compression with a large receptive field and low computational complexity.
It achieves efficient compression and reconstruction of point clouds at any scale, eliminates the need to encode feature point coordinates, and improves the adaptability and efficiency of point cloud encoding.
Smart Images
Figure CN115883850B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of graphics processing technology, and in particular to a resolution-adaptive point cloud geometric lossy encoding method, computer device, and storage medium based on depth residual compression and sparse representation. Background Technology
[0002] Point clouds are collections of points with three-dimensional geometric positions and other attributes (e.g., color, reflectivity), which can be used to flexibly represent stereoscopic visual data such as 3D scenes and objects. Efficient point cloud compression methods can effectively promote the development of emerging fields such as virtual reality and autonomous driving.
[0003] Currently, representative traditional point cloud compression methods internationally include geometry-based point cloud compression (G-PCC) and video-based point cloud compression (V-PCC). G-PCC models and encodes point cloud information using appropriate data structures (such as octrees or triangular meshes) based on the coordinate information of the point cloud. The V-PCC method maps the 3D point cloud to a 2D plane and then encodes the point cloud using mature video coding methods.
[0004] Although traditional point cloud coding methods have achieved excellent compression performance, these methods rely heavily on manually designed coding strategies and are difficult to optimize end-to-end.
[0005] In recent years, several improved point cloud compression methods have emerged. For example, in the lossy geometric compression of dense point clouds, the PGCGv2 compression method based on 3D sparse convolution has achieved superior performance compared to other methods, while maintaining relatively low computational complexity. A simplified flowchart of this algorithm is shown below. Figure 1 As shown, the dark network dots represent occupied voxels. Disadvantages of this type of method include:
[0006] 1. For large-scale point clouds, the backbone network used in this method is relatively shallow, and its receptive field is relatively limited, making it difficult to model the relationship between points that are far apart.
[0007] 2. This method uses a backbone network with a fixed downsampling rate, making it lack adaptability to point clouds of different resolutions. Specifically, for larger-scale point clouds, this method requires additional point cloud encoding tools to encode the coordinates of feature points.
[0008] 3. This method maps the geometric information of the original point cloud to a single latent variable space, failing to fully utilize the latent variables generated at various resolutions during the downsampling process for information representation. Summary of the Invention
[0009] To address the technical problems existing in current related technologies, the purpose of this invention is to provide a resolution-adaptive point cloud geometric lossy coding method, computer device, and storage medium based on deep residual compression and sparse representation.
[0010] On one hand, embodiments of the present invention include a resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation, comprising:
[0011] Acquire point cloud data to be processed;
[0012] The point cloud data to be processed is subjected to lossy compression to obtain lossy compressed data;
[0013] The encoding module performs a multi-round encoding process. In each round of the encoding process, the encoding module encodes the input data corresponding to the round of the encoding process and outputs the corresponding encoded bitstream. The input data corresponding to the first round of the encoding process is the lossy compressed data, and the input data corresponding to any other round of the encoding process is the encoded bitstream output by the previous round of the encoding process.
[0014] The entire encoded bitstream is used as the encoding result for the point cloud data to be processed.
[0015] Further, the lossy compression of the point cloud data to be processed to obtain lossy compressed data includes:
[0016] The point cloud data to be processed is downsampled to obtain the lossy compressed data.
[0017] Furthermore, each of the encoding modules encodes its corresponding input data and outputs a corresponding encoded bitstream, including:
[0018] For any of the encoding modules, obtain the feature residuals and masks corresponding to the input data of the encoding module;
[0019] The feature residuals are sequentially quantized and processed by arithmetic coding algorithms to obtain the first bitstream corresponding to the coding module.
[0020] The mask is processed by an arithmetic coding algorithm to obtain the second bitstream corresponding to the coding module;
[0021] The input data is downsampled to obtain the third bitstream corresponding to the encoding module;
[0022] The first bitstream, the second bitstream, and the third bitstream are used as the encoded bitstream output by the encoding module.
[0023] Furthermore, the arithmetic coding algorithm is the rANS coding algorithm.
[0024] Furthermore, the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation also includes:
[0025] Storing and / or transmitting part or all of the encoded bitstream.
[0026] Furthermore, the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation also includes:
[0027] The decoding module performs a multi-round decoding process. In each round of the decoding process, the decoding module decodes the input data corresponding to the round of decoding process and outputs the corresponding decoding result of the round of decoding process. The input data corresponding to the first round of decoding process is the last encoded bitstream, and the input data corresponding to any other round of decoding process is the encoded bitstream in reverse order and the decoding result output by the previous round of decoding process.
[0028] Obtain the decoding result output by the decoding module in the last round of the decoding process, and use it as lossless decoding data;
[0029] The lossless decoded data is decompressed using lossy methods to obtain reconstructed point cloud data.
[0030] Further, the lossy decompression of the lossless decoded data to obtain reconstructed point cloud data includes:
[0031] The lossless decoded data is upsampled to obtain the reconstructed point cloud data.
[0032] Furthermore, the encoding module is a neural network encoding module, and the decoding module is a neural network decoding module.
[0033] On the other hand, embodiments of the present invention also include a computer device, including a memory and a processor, the memory for storing at least one program, and the processor for loading the at least one program to execute the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in the embodiments.
[0034] On the other hand, embodiments of the present invention also include a storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in the embodiments.
[0035] The beneficial effects of the present invention are as follows: The resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in the embodiments has a large receptive field and relatively small computational load. It can compress and reconstruct point clouds and aggregate point cloud data of any scale into a single feature point, thereby eliminating the need to encode feature point coordinates while ensuring the large receptive field of the backbone network. Attached Figure Description
[0036] Figure 1 This is a schematic diagram illustrating the principle of PGCGv2;
[0037] Figure 2 This is a schematic diagram illustrating the principle of the point cloud compression scheme based on depth residuals in the embodiment.
[0038] Figure 3 This is a schematic diagram illustrating the principle of a lossy geometric compression scheme based on partially lossless compression in the embodiments.
[0039] Figure 4 This is a flowchart of a resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in the embodiment.
[0040] Figure 5 and Figure 6 This is a schematic diagram illustrating the principle of the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in the embodiment.
[0041] Figure 7 This is a schematic diagram of the neural network encoding module in the embodiment;
[0042] Figure 8 This is a schematic diagram of the neural network decoding module in the embodiment;
[0043] Figure 9 This is a schematic diagram of the classification module in the embodiment. Detailed Implementation
[0044] In this embodiment, we can first consider the following two schemes, A and B.
[0045] A. Point cloud compression scheme based on deep residual
[0046] The simplified process of Option A is as follows: Figure 2 As shown, Figure 2 The following example illustrates a two-dimensional point cloud compression model based on convolution, where dark network points represent occupied voxels.
[0047] Reference Figure 2At the encoding end, Scheme A first voxels the input point cloud, and then uses a series of sparse convolution-based downsampling modules to lossily map the geometric information of the voxelized point cloud to the feature space. The resulting latent variables are then distributed and entropy encoded to generate the transmission bitstream. At the decoding end, the bitstream is first decoded based on global priors to obtain latent variables. Then, a series of upsampling modules are used to upsample the latent variables to the original resolution to achieve lossy reconstruction of the point cloud.
[0048] Compared to conventional residual-based compression methods, Scheme A does not generate residuals in the original data space, but rather generates feature residuals in an optimizable feature space, thus offering greater flexibility in latent variable generation. Compared to other existing deep learning-based point cloud compression methods, Scheme A fully utilizes latent variables at various resolutions for information representation, and through a recurrent downsampling subnetwork, ensures adaptive latent variable generation for point clouds of arbitrary scales.
[0049] Scheme A does not impose specific restrictions on the basic network structure or whether it is a lossless geometric codec. Figure 2 The convolutional network shown is for demonstration purposes only. Scheme A is applicable to various data compression models based on feature extraction and latent variable encoding. For models based on lossy point cloud encoding using sparse representations, the geometric information of the feature point clouds at various resolutions generated at the decoder is not entirely consistent with the corresponding geometric information at the encoder. In this case, the encoder generates matching feature residuals based on the geometric information reconstructed at the decoder.
[0050] B. Lossy geometric compression scheme based on partially lossless compression
[0051] Scheme B works by using mask-based partially lossless geometric compression to assist high-quality lossy geometric compression, such as... Figure 3 As shown (using a 2D point cloud as an example). Specifically, for Figure 3 The voxelized feature point cloud, obtained after several lossy downsampling operations, is used by the encoder to generate a binary mask that losslessly represents the geometric information of the original feature point cloud, based on the voxel occupancy of the expanded point cloud generated by the decoder and the original feature point cloud. This mask is then encoded based on the distribution predicted by the neural network. The decoder decodes this mask to achieve lossless upsampling of the point cloud's geometric information at the current scale. During training, the optimization objective is to minimize the cross-entropy between the predicted distribution and the actual mask.
[0052] Based on the principles of schemes A and B, this embodiment designs a resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation. (Refer to...) Figure 4 The resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation includes the following steps:
[0053] S1. Obtain the point cloud data to be processed;
[0054] S2. Perform lossy compression on the point cloud data to be processed to obtain lossy compressed data;
[0055] S3. Use an encoding module to perform a multi-round encoding process. In each round of encoding, the encoding module encodes the input data corresponding to that round of encoding and outputs the corresponding encoded bitstream. The input data corresponding to the first round of encoding is lossy compressed data, and the input data corresponding to any other round of encoding is the encoded bitstream output by the previous round of encoding.
[0056] S4. Use the entire encoded bitstream as the encoding result for the point cloud data to be processed.
[0057] In this embodiment, the principle of steps S1-S4 is as follows: Figure 5 As shown. Due to Figure 5 The encoding process performed by multiple encoders (Enc) in a given context is similar, and similarly, the encoding process performed by multiple decoders (Dec) is similar. Therefore, it can be simplified as follows: Figure 6 The style shown.
[0058] In step S1, point cloud data to be processed can be acquired through methods such as LiDAR imaging.
[0059] In step S2, refer to Figure 5 or Figure 6 The point cloud data to be processed is subjected to lossy compression, and the result of lossy compression is lossy compressed data.
[0060] In this embodiment, the point cloud data to be processed can be downsampled, and the result of the downsampling can be used as lossy compressed data. That is, the "lossy compression" in step S2 can specifically be "downsampling".
[0061] In step S3, the encoding module performs a multi-round encoding process. In each round of encoding, the encoding module encodes the input data corresponding to that round of encoding and outputs the corresponding encoded bitstream. The input data corresponding to the first round of encoding is lossy compressed data, and the input data corresponding to any other round of encoding is the encoded bitstream output by the previous round of encoding.
[0062] In step S3, the encoding module (or encoder) used can be a neural network encoding module. In this embodiment, the structure of the neural network encoding module used is as follows: Figure 7 As shown. Wherein, "Convc*n 3"" indicates a sparse convolution with output channel c and kernel size n×n×n, "s↓" and "s↑" indicate upsampling and downsampling with factor s, and "ReLU" indicates a corrected linear unit.
[0063] When executing step S3, either one encoding module can be used to execute each round of encoding processing, or multiple encoding modules can be used, with each encoding module executing its corresponding round of encoding processing.
[0064] When performing step S3, refer to Figure 5 In each round of encoding processing, the encoding module encodes the input data corresponding to that round and outputs the corresponding encoded bitstream. Taking the use of multiple encoding modules as an example... Figure 5 In the sequence, the first encoding module is Figure 5 In the "lossless compression module", the leftmost encoder Enc encodes the lossy compressed data. The bitstream obtained by the first encoding module can be saved and used as input data for the second encoding module (the encoding module to the right of the first encoding module), which then encodes it. This process continues until multiple bitstreams are obtained.
[0065] In this embodiment, the multiple bitstreams obtained by executing step S3 constitute the execution result of the resolution-adaptive point cloud geometric lossy encoding method based on deep residual compression and sparse representation, i.e., the encoded bitstream. The encoded bitstream contains information from the point cloud data to be processed, can serve as the compression result of the point cloud data to be processed, and can be stored or transmitted, thereby realizing the storage and transmission of information from the point cloud data to be processed.
[0066] In this embodiment, the input data of each encoding module is different, but the principle of the data processing process performed by each encoding module is similar. The steps performed by one of the encoding modules can be used as an example for explanation.
[0067] In this embodiment, when any encoding module performs the step of "encoding its corresponding input data and outputting the corresponding encoded bitstream", it can specifically perform the following steps:
[0068] P1. Obtain the feature residuals and mask corresponding to the input data of this encoding module;
[0069] P2. The feature residuals are sequentially quantized and processed by arithmetic coding algorithms to obtain the first bitstream corresponding to the coding module;
[0070] P3. Perform arithmetic coding algorithm processing on the mask to obtain the second bitstream corresponding to the coding module;
[0071] P4. Perform downsampling on the input data to obtain the third bitstream corresponding to this encoding module;
[0072] P5. The first bitstream, the second bitstream, and the third bitstream are used as the encoded bitstreams output by this encoding module.
[0073] The encoding process of any encoding module is as follows: Figure 6 As shown. In step P1, the encoding module obtains the feature residuals and masks corresponding to the input data (if it is the first encoding module, the input data is lossy compressed data; otherwise, the input data is the bitstream output by the previous encoding module).
[0074] In step P2, after quantizing the feature residuals, the encoding module processes them using an arithmetic coding algorithm to obtain the first bitstream corresponding to the encoding module. Figure 6 (The stream located on the far left).
[0075] In step P3, the encoding module uses an arithmetic coding algorithm to process the mask to obtain the second bitstream corresponding to the encoding module. Figure 6 (The bitstream located in the middle).
[0076] In step P4, the encoding module downsamples the input data to obtain the third bitstream corresponding to the encoding module. Figure 6 (The stream located on the far right).
[0077] The first, second, and third bitstreams obtained in steps P2-P4 constitute the encoded bitstream output by the encoding module.
[0078] In this embodiment, the arithmetic encoding algorithm used by the encoding module in steps P1-P5 is the rANS encoding algorithm.
[0079] The rANS encoding algorithm compresses and encodes the latent features output by the neural network and the actual point cloud coordinates, making it a feasible arithmetic entropy encoding algorithm.
[0080] The rANS entropy model encoding formula is as follows:
[0081]
[0082] In the formula, x is the encoded value, s is the character to be encoded, CDF is the cumulative distribution function, and f s The number of times s appears.
[0083] Let's explain with an example:
[0084] For the character set ['a', 'b', 'c'], with quantization n = 3, its statistical distribution is [f a ,f b ,f c]=[5,2,1], corresponding to CDF[s]=[0,5,7,8], the encoding process of the string "abc" is as follows: Given an initial value x0=8, then
[0085]
[0086]
[0087]
[0088] The encoded 375 is used for transmission and storage.
[0089] The rANS encoding algorithm requires two elements: first, the character set to be encoded (which can be the characters s to be encoded); and second, the probability of occurrence of each character, i.e., the distribution of the characters (which can be the cumulative distribution function CDF). Currently, most methods assume that the character distribution follows a uniform or Gaussian distribution. In this method, the character set to be encoded is known during the encoding process, while the character distribution can be estimated by modeling. The estimation process in this embodiment is as follows:
[0090] First, assuming y is the latent feature of the point cloud to be encoded, assign a prior z to each point cloud image.
[0091] According to Shannon's cross-entropy formula, the final encoded code length is:
[0092]
[0093] Specifically, the conditional entropy model in this embodiment is defined as follows:
[0094]
[0095]
[0096] Where f K and f k From the DeepFactorized entropy model:
[0097] g k (y i )=y i +tanh(a (k) ⊙tanh(y i ),
[0098] f k (y i ) = g k (softplus(H (k) )y i +b (k) ), 1≤k≤K,
[0099] fK (y i ) = sigmoid(softplus(H (K) )y i +b (K) )
[0100] Where Q represents the truncation quantization function.
[0101] The resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in this embodiment uses long-distance residual connections to assist in constructing a backbone network with a large receptive field and relatively low computational cost. Specifically, the principle of the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in this embodiment is as follows:
[0102] 1. Use a cyclic double downsampling and upsampling module to compress and reconstruct the point cloud; for input point clouds of any scale, aggregate them into a single feature point through cyclic downsampling, thereby eliminating the need to encode feature point coordinates while ensuring the large receptive field of the backbone network;
[0103] 2. Constrain the number of channels in deep networks to maintain low computational complexity in the backbone network;
[0104] 3. Use residual connections with information entropy constraints to compensate for the lack of semantic information of individual latent variables.
[0105] In this embodiment, in addition to performing steps S1-S4, the following steps may also be performed:
[0106] S5. Use the decoding module to perform a multi-round decoding process. In each round of decoding, the decoding module decodes the input data corresponding to the round of decoding and outputs the corresponding decoding result. The input data corresponding to the first round of decoding is the last encoded bitstream, and the input data corresponding to any other round of decoding is the encoded bitstream in reverse order and the decoding result output by the previous round of decoding.
[0107] S6. Obtain the decoding result output by the decoding module in the last round of decoding processing, and use it as lossless decoding data;
[0108] S7. Perform lossy decompression on the lossless decoded data to obtain reconstructed point cloud data.
[0109] Steps S5-S7 are equivalent to the reverse process of steps S1-S4.
[0110] In step S5, the decoding module performs a multi-round decoding process. In each round of decoding, the decoding module decodes the input data corresponding to the encoding process and outputs the corresponding decoding result of the decoding process.
[0111] In step S5, the decoding module (or decoder) used can be a neural network decoding module. In this embodiment, the structure of the neural network decoding module used is as follows: Figure 8 As shown. Wherein, "Convc*n 3 "" indicates a sparse convolution with output channels c and kernel size n×n×n, "s↓" and "s↑" indicate upsampling and downsampling with factors of s, and "ReLU" indicates a corrected linear unit. (See reference...) Figure 5 The decoder also includes a classification module (classifier). The structure of the classifier used in this embodiment is as follows: Figure 9 As shown.
[0112] When performing step S5, either one decoding module can be used to execute each round of decoding process separately, or multiple decoding modules can be used, with each decoding module executing its corresponding round of decoding process separately.
[0113] When performing step S5, refer to Figure 5 In each round of decoding, the decoding module encodes the input data corresponding to that round and outputs the corresponding decoding result. Taking the use of multiple decoding modules as an example... Figure 5 In the sequence, the first decoding module is... Figure 5 The rightmost decoder, Dec, in the "lossless compression module" processes the last encoded bitstream. Figure 5 The rightmost bitstream is decoded to obtain the corresponding decoding result; the decoding result obtained by the first decoding module can be saved and used as the input data for the second decoding module (the encoding module to the left of the first encoding module), while the second to last encoded bitstream ( Figure 5 The second bitstream from the right (its reverse order in the entire encoded bitstream is the same as the order of the second decoding module in the entire decoding module) is also used as input data for the second decoding module, which performs decoding processing. This process is repeated until the decoding result of the last decoding module is obtained, which is the lossless decoded data.
[0114] In this embodiment, the lossless decoded data can be upsampled, and the result obtained from the upsampling can be used as the reconstructed point cloud data. That is, the "lossy decompression" in step S7 can specifically be "upsampling".
[0115] By executing steps S5-S7, the encoded bitstream obtained from steps S1-S4 can be decoded, thereby reconstructing the information in the point cloud data to be processed and obtaining reconstructed point cloud data with the same information as the original point cloud data to be processed.
[0116] A computer program can be written to execute the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in this embodiment. This computer program can be written into a computer device or storage medium. When the computer program is read out and run, the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in this embodiment can be executed, thereby achieving the same technical effect as the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation in the embodiment.
[0117] It should be noted that, unless otherwise specified, when a 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. Furthermore, the descriptions of "upper," "lower," "left," and "right" used in this disclosure are only relative to the relative positional relationships of the various components of this disclosure in the accompanying drawings. The singular forms "a," "described," and "the" used in this disclosure are also intended to include the plural forms, unless the context clearly indicates otherwise. Moreover, unless otherwise defined, all technical and scientific terms used in this embodiment have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this embodiment specification is only for describing particular embodiments and is not intended to limit the invention. The term "and / or" as used in this embodiment includes any combination of one or more of the associated listed items.
[0118] It should be understood that although the terms first, second, third, etc., may be used to describe various elements in this disclosure, 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, a first element may also be referred to as a second element without departing from the scope of this disclosure, and similarly, a second element may also be referred to as a first element. The use of any and all instances or exemplary language (“e.g.,” “such as,” etc.) provided in this embodiment is intended only to better illustrate embodiments of the invention and, unless otherwise required, does not impose a limitation on the scope of the invention.
[0119] 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 storage medium. The method can be implemented using standard programming techniques—including a non-transitory computer-readable storage medium configured with a computer program, wherein such a storage medium 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. Furthermore, for this purpose, the program can run on a programmed application-specific integrated circuit (ASIC).
[0120] Furthermore, the procedures described in this embodiment can be performed in any suitable order unless otherwise indicated by this embodiment or clearly contradicted by the context. The procedures (or variations and / or combinations thereof) described in this embodiment can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.
[0121] Furthermore, the method can be implemented in any suitable type of computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices. Aspects of the invention can be implemented as machine-readable code stored on a non-transitory storage medium or device, whether removable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it is readable by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein. Furthermore, the machine-readable code, or portions thereof, can be transmitted via wired or wireless networks. The invention described in this embodiment includes these and other different types of non-transitory computer-readable storage media when such media comprises instructions or programs that implement the steps described above in conjunction with a microprocessor or other data processor. When programmed according to the methods and techniques described in the invention, the invention also includes the computer itself.
[0122] 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 invention, the transformed data represents physical and tangible objects, including specific visual depictions of physical and tangible objects generated on the display.
[0123] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention, as long as they achieve the technical effects of the present invention by the same means, should be included within the scope of protection of the present invention. Within the scope of protection of the present invention, the technical solutions and / or implementation methods can have various modifications and variations.
Claims
1. A resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation, characterized in that, The resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation includes: Acquire point cloud data to be processed; The point cloud data to be processed is subjected to lossy compression to obtain lossy compressed data; The encoding module performs a multi-round encoding process. In each round of the encoding process, the encoding module encodes the input data corresponding to the round of the encoding process and outputs the corresponding encoded bitstream. The input data corresponding to the first round of the encoding process is the lossy compressed data, and the input data corresponding to any other round of the encoding process is the encoded bitstream output by the previous round of the encoding process. The entire encoded bitstream is used as the encoding result for the point cloud data to be processed; Each of the aforementioned encoding modules encodes its corresponding input data and outputs a corresponding encoded bitstream, specifically including: For any of the encoding modules, obtain the feature residuals and masks corresponding to the input data of the encoding module; The feature residuals are sequentially quantized and processed by arithmetic coding algorithms to obtain the first bitstream corresponding to the coding module. The mask is processed by an arithmetic coding algorithm to obtain the second bitstream corresponding to the coding module; The input data is downsampled to obtain the third bitstream corresponding to the encoding module; The first bitstream, the second bitstream, and the third bitstream are used as the encoded bitstream output by the encoding module.
2. The resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation according to claim 1, characterized in that, The step of performing lossy compression on the point cloud data to be processed to obtain lossy compressed data includes: The point cloud data to be processed is downsampled to obtain the lossy compressed data.
3. The resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation according to claim 1, characterized in that, The arithmetic coding algorithm is the rANS coding algorithm.
4. The resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation according to claim 1, characterized in that, The resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation also includes: Storing and / or transmitting part or all of the encoded bitstream.
5. The resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation according to any one of claims 1-4, characterized in that, The resolution-adaptive geometric lossy coding method for point clouds based on deep residual compression and sparse representation also includes: The decoding module performs a multi-round decoding process. In each round of the decoding process, the decoding module decodes the input data corresponding to the round of decoding process and outputs the corresponding decoding result of the round of decoding process. The input data corresponding to the first round of decoding process is the last encoded bitstream, and the input data corresponding to any other round of decoding process is the encoded bitstream in reverse order and the decoding result output by the previous round of decoding process. Obtain the decoding result output by the decoding module in the last round of the decoding process, and use it as lossless decoding data; The lossless decoded data is decompressed using lossy methods to obtain reconstructed point cloud data.
6. The resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation according to claim 5, characterized in that, The step of performing lossy decompression on the lossless decoded data to obtain reconstructed point cloud data includes: The lossless decoded data is upsampled to obtain the reconstructed point cloud data.
7. The resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation according to claim 5, characterized in that, The encoding module is a neural network encoding module, and the decoding module is a neural network decoding module.
8. A computer device, characterized in that, The method includes a memory and a processor, the memory being used to store at least one program, and the processor being used to load the at least one program to execute the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation as described in any one of claims 1-7.
9. A computer-readable storage medium storing a processor-executable program, characterized in that, The processor-executable program, when executed by the processor, is used to perform the resolution-adaptive point cloud geometric lossy coding method based on deep residual compression and sparse representation as described in any one of claims 1-7.