Three-dimensional point cloud geometric compression method and device based on diffusion model and medium

Abstract

The invention discloses a three-dimensional point cloud geometric compression method and device based on a diffusion model and a medium, and the method comprises the steps: enabling an input point cloud to pass through a dual-domain encoder, obtaining low-frequency features and high-frequency features, and extracting skeleton points for space guidance; probability distribution of the low-frequency features is modeled through a full decomposition entropy model, probability distribution of the high-frequency features is modeled under low-frequency guidance, arithmetic coding is carried out, and skeleton points are subjected to lossless compression through G-PCC to jointly form a final compressed code stream; and carrying out multi-round iterative denoising reconstruction through a diffusion model, taking low-frequency features, high-frequency features and skeleton points as conditional information, inputting the conditional information into a decoding process in a mode of affine transformation injection and attention mechanism fusion, and guiding gradual recovery of a point cloud structure and details. According to the method, more robust point cloud geometric feature extraction is realized through a high and low frequency feature double-path fusion mode, a diffusion reconstruction process based on affine transformation and an attention mechanism is further guided, and the compression efficiency and the reconstruction quality are effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional point cloud data compression, and in particular relates to a three-dimensional point cloud geometry compression method, device and medium based on a diffusion model. Background Art

[0002] With the widespread application of three-dimensional perception technology in fields such as microwave photon radar, autonomous driving, robotics, and virtual reality, point cloud data, as a basic representation, has become a research hotspot for compression and high-quality reconstruction. Traditional point cloud compression methods, such as octrees or prediction trees, rely on rule-based modeling and are unable to fully utilize the geometric feature expression capabilities of point clouds. In recent years, deep learning-based point cloud compression methods have made some progress in feature extraction and reconstruction quality, but the following problems still exist: 1) Most methods focus on modeling local neighborhood features and ignore the compression value of global shape information; 2) The commonly used variational autoencoder (VAE) structure has ambiguity in the reconstruction process and has difficulty capturing high-frequency details; 3) The reconstruction network suffers from information loss and missing details when dealing with complex geometric structures.

[0003] As an emerging generative paradigm, diffusion models have achieved breakthroughs in recent years in image, speech, and point cloud generation tasks, demonstrating excellent reconstruction accuracy and detail preservation. Furthermore, point clouds inherently possess multi-scale structural characteristics, including high- and low-frequency features. Effectively utilizing this information during compression can significantly improve point cloud compression ratios and reconstruction quality. Summary of the Invention

[0004] Purpose of the invention: The present invention provides a three-dimensional point cloud geometric compression method, device and medium based on a diffusion model, which can simultaneously preserve the overall structure and local detail information of the point cloud, and utilize the generation capability of the diffusion model to improve compression efficiency and reconstruction accuracy.

[0005] Technical solution: The present invention provides a method for compressing three-dimensional point cloud geometry based on a diffusion model, comprising the following steps:

[0006] (1) Processing the collected point cloud data;

[0007] (2) Construct a point cloud compression model DiffPCC, which includes a dual-domain encoder, a bottleneck layer module, and a diffusion decoder. The input point cloud is passed through the dual-domain encoder to obtain low-frequency features and high-frequency features, and skeleton points are extracted for spatial guidance. The bottleneck layer module adopts differentiated compression strategies for low-frequency features, high-frequency features, and skeleton points to form a complete compressed code stream. Multiple rounds of denoising and reconstruction are performed through the diffusion decoder. Low-frequency, high-frequency features, and skeleton points are used as conditional information, and fused through affine injection and attention mechanism to achieve the gradual recovery of point cloud structure and details.

[0008] (3) The constructed point cloud compression model DiffPCC is trained, and the model parameters are end-to-end optimized using a rate-distortion loss function that includes chamfer distance and encoding bit rate to achieve efficient compression and decompression of three-dimensional point cloud data.

[0009] Furthermore, the implementation process of step (1) is as follows:

[0010] The collected point cloud data is processed by centering and translation normalization, and uniformly mapped to the unit sphere; each point cloud is resampled to a fixed number.

[0011] Furthermore, the dual-domain encoder includes a low-frequency feature extraction branch and a high-frequency feature extraction branch: the low-frequency branch is based on the improved point cloud neural network PointNet, and extracts the overall structural information of the point cloud through multi-layer feature distillation and average pooling operations, generating low-frequency features with global geometric perception capabilities; the high-frequency branch adopts a high-frequency arithmetic coding module, combined with the point-level attention mechanism and the block-level attention mechanism, to model the local change areas and inter-block structures in the point cloud, and extract fine-grained high-frequency features.

[0012] Furthermore, the bottleneck layer module includes a low-frequency feature entropy modeling module, a high-frequency feature joint modeling module and a skeleton point compression module; the low-frequency feature entropy modeling module performs probability modeling on the low-frequency features through a full decomposition entropy model, and uses arithmetic coding to achieve compression; the high-frequency feature joint modeling module constructs a low-frequency-guided high-frequency entropy model under the guidance of low-frequency features, which is used to model its symbol probability distribution and complete arithmetic coding; the skeleton point compression module performs lossless compression through the G-PCC method based on the octree structure; the encoding results output by the three modules together constitute the final compressed code stream for storage and transmission, and after inverse reconstruction, it is provided to the diffusion decoder for conditional reconstruction.

[0013] Furthermore, the diffusion decoder is constructed based on the conditional diffusion model, including a word segmenter module, a conditional information injection module and a point rearrangement module; the word segmenter module divides the initial noise point cloud into local areas through K-nearest neighbor clustering guided by skeleton points, and extracts regional-level Token representations; the conditional information injection module is used to inject the low-frequency features, high-frequency features and skeleton points generated by the encoder into the diffusion process as conditional information, wherein the low-frequency features are injected into the Token representation through affine transformation, and the high-frequency features are fused with the Token through the self-attention mechanism; the updated Token is feature propagated and the residual updated through graph convolution to form the input of the next iteration; the point rearrangement module is used to map the Token generated by the final iteration into three-dimensional point cloud coordinates to achieve gradual restoration of structure and details.

[0014] Furthermore, the implementation process of step (3) is as follows:

[0015] By simultaneously optimizing the dual-domain encoder, bottleneck layer module, and diffusion decoder module, unified modeling of compression coding and point cloud reconstruction is achieved;

[0016] The rate-distortion loss function composed of chamfer distance and feature bit rate is used:

[0017]

[0018] Among them, P is the original point cloud, is the reconstructed point cloud, CD is the chamfer distance, R(f l ) and R(f h ) represent the bit rates of low-frequency features and high-frequency features respectively, and λ is used to balance distortion and compression rate.

[0019] A device according to the present invention includes a memory and a processor, wherein:

[0020] a memory for storing computer programs capable of running on the processor;

[0021] The processor is configured to execute the steps of the three-dimensional point cloud geometry compression method based on the diffusion model when running the computer program.

[0022] The present invention provides a storage medium having a computer program stored thereon, which, when executed by at least one processor, implements the steps of the three-dimensional point cloud geometry compression method based on the diffusion model as described above.

[0023] Beneficial effects: Compared with the existing technology, the beneficial effects of the present invention are: the present invention proposes a point cloud compression modeling method that integrates high- and low-frequency features, and introduces the conditional diffusion model into the point cloud reconstruction process. Compared with traditional methods, this method can better maintain the structural consistency and detail restoration ability of the point cloud; the present invention uses a high- and low-frequency feature separation structure at the encoding end to more accurately characterize the multi-scale features of the point cloud; in the compression stage, a low-frequency-guided high-frequency entropy model is introduced to effectively compress high-frequency detail information; in the decoding stage, a diffusion reconstruction process is proposed, combined with affine injection and attention fusion mechanism, which effectively improves the geometric fidelity and boundary clarity of the generated point cloud; the overall method shows better reconstruction quality and compression efficiency than existing methods on multiple point cloud data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the DiffPCC framework of the point cloud compression model;

[0025] Figure 2 Schematic diagram of the dual-domain encoder module structure in the network framework of the present invention;

[0026] Figure 3It is a structural diagram of a diffusion decoder;

[0027] Figure 4 Schematic diagram of the compression effect of DiffPCC of the present invention and other methods. DETAILED DESCRIPTION

[0028] The present invention will be further described in detail below with reference to the accompanying drawings.

[0029] The present invention proposes a three-dimensional point cloud geometry compression method based on a diffusion model, comprising the following steps:

[0030] Step 1: Collection of point cloud data.

[0031] The training data used in this implementation was collected from ShapeNet, a large-scale, publicly available 3D model dataset. Specifically, representative 3D object models were selected from the ShapeNet dataset, and their 3D mesh surfaces were uniformly sampled using standard point cloud conversion tools to generate a dense point cloud representation. To ensure spatial consistency and scale controllability of the point cloud during training, all sampled point cloud data was centered and normalized by translation, and uniformly mapped to the unit sphere. For uneven point density in the model, the point cloud was resampled using a fixed number (8192 points) to ensure consistency in the training sample dimensions and stability of the network input.

[0032] Step 2: Construct the point cloud compression model DiffPCC, such as Figure 1 As shown, it includes a dual-domain encoder, a bottleneck layer module and a diffusion decoder, and adopts a unified end-to-end training method.

[0033] like Figure 2 As shown in the figure, at the encoding end, the dual-domain encoder consists of a low-frequency feature extraction branch and a high-frequency feature extraction branch. The low-frequency branch, based on an improved point cloud neural network (PointNet), extracts the overall structural information of the point cloud through multi-layer feature distillation and average pooling operations, generating low-frequency features with global geometric awareness. The high-frequency branch uses a high-frequency arithmetic coding module, combined with point-level and block-level attention mechanisms, to model local variation regions and inter-block structures in the point cloud and extract fine-grained high-frequency features. The low-frequency and high-frequency features are input as conditional information into the diffusion decoder, guiding the structural recovery and detail reconstruction of the point cloud during the generation process.

[0034] The bottleneck layer module includes a low-frequency feature entropy modeling module, a high-frequency feature joint modeling module and a skeleton point compression module. In order to fully explore the multi-scale structural characteristics of the point cloud in the frequency domain, the present invention adopts differentiated compression strategies for low-frequency features, high-frequency features and skeleton points respectively: low-frequency features are probabilistically modeled by completely decomposing the entropy model, and compressed into a bit stream using arithmetic coding to express the overall shape of the point cloud in a compact form; high-frequency features establish a conditional Gaussian probability model under the guidance of low-frequency features, effectively capturing the distribution characteristics of local detail changes, thereby achieving more accurate joint entropy coding; skeleton points are used as key anchor points for structural information, and the G-PCC module is used for lossless compression to provide spatial reference support for subsequent decoding. The three types of information each play different roles and work together to form the final compressed code stream, taking into account both compression efficiency and reconstruction quality.

[0035] In the decoding stage, the diffusion model is initialized with Gaussian noise, and the point cloud is restored through multiple rounds of denoising. Figure 3 As shown in the figure, the diffusion decoder is constructed based on the conditional diffusion model, including a word segmenter module, a conditional information injection module and a point rearrangement module; the word segmenter module divides the initial noise point cloud into local regions through K-nearest neighbor clustering guided by skeleton points, and extracts regional-level Token representations; the conditional information injection module is used to inject the low-frequency features, high-frequency features and skeleton points generated by the encoding end into the diffusion process as conditional information, wherein the low-frequency features are injected into the Token representation through affine transformation, and the high-frequency features are fused with the Token through the self-attention mechanism; the updated Token is feature propagated and the residual updated through graph convolution to form the input of the next iteration; the point rearrangement module is used to map the Token generated by the final iteration into three-dimensional point cloud coordinates to achieve gradual restoration of structure and details.

[0036] Step 3: Train the constructed point cloud compression model DiffPCC and use the rate-distortion loss function including chamfer distance and encoding bit rate to perform end-to-end optimization on the model parameters to achieve efficient compression and decompression of 3D point cloud data.

[0037] During training, a rate-distortion joint optimization strategy is used to simultaneously optimize the quality of structural reconstruction and compression efficiency. The loss function consists of two parts: the reconstruction error term, which uses the Chamfer distance to measure the spatial error between the original point cloud and the reconstructed point cloud; and the compression rate constraint term, which penalizes the encoding bits of low-frequency and high-frequency features respectively. The loss function is as follows:

[0038]

[0039] R(f)=E x～p(x) [-log2 q(f(x))]

[0040]

[0041] in, Represents the chamfer distance between the original point cloud and the reconstructed point cloud, R(f h ) and R(f l ) represent the bit rate estimation of high-frequency features and low-frequency features respectively.

[0042] The present invention achieves better compression rate and higher-fidelity point cloud reconstruction effects in various scenarios by collaboratively encoding the high- and low-frequency features of the point cloud and introducing a decoding structure based on the conditional diffusion model.

[0043] The proposed DiffPCC method uses ShapeNetCoreV2 generated point cloud data as the training set, covering 51,300 point cloud samples across 55 categories, each containing 8,192 points. Point clouds in the training set are uniformly sampled from a mesh, simulating the actual point cloud acquisition process. The training, test, and validation sets are divided in an 8:1:1 ratio. To enhance the model's robustness to degradation, the point cloud degradation process employs a combination of Gaussian blurring with random parameters, bicubic downsampling, and Gaussian noise superposition to synthesize compressed input.

[0044] Regarding network parameter settings, DiffPCC training was implemented using the PyTorch 2.1 and pytorch3d 0.75 frameworks. All models were trained on the same experimental platform. The Adam optimizer was used during training, with an initial learning rate of 1e-4, which was decayed by a factor of 10 each time the model converged. A rate-distortion loss function was used, consisting of a Chamfer Distance and a compression bitrate term, to jointly optimize reconstruction accuracy and compression efficiency. The trade-off coefficient λ in the loss function was set to 1e-4. The model was trained for a total of 50 epochs.

[0045] The dual-domain encoder, a key encoding module, is responsible for extracting low-frequency features, high-frequency features, and skeleton points from the point cloud. The low-frequency channel extraction module uses average pooling instead of maximum pooling to enhance the representation of smooth regions. The high-frequency channel combines click attention with block-level attention to extract detailed block- and point-level features. The resulting features are compressed using independent entropy modeling or conditional modeling, and then fed into an arithmetic encoder. Skeleton points are losslessly encoded using the G-PCC module. On the decoding side, the diffusion model, the core of the conditional diffusion reconstruction network, consists of a tokenizer module, a conditional information injection module, and a point reordering module. The initial noisy point cloud is constructed using a KNN neighborhood. After aggregation to form tokens, the affine transformation of the low-frequency features and the attention representation of the high-frequency features are sequentially fused, and then structural enhancement is performed using a graph convolution module. After multiple rounds of iterations, the output tokens are upsampled by the point reordering module to restore the final reconstructed point cloud.

[0046] To verify the effectiveness of the present invention, the examples selected multiple typical point cloud samples and, under the same compression rate, reconstructed them using existing representative methods (including G-PCC, IPDAE, and PointSoup) and the DiffPCC model proposed in this invention. A comparative analysis was then conducted from a visual perspective. As shown in the figure, each row shows the reconstruction results of the same type of point cloud sample using different methods: the original point cloud (GT), a traditional method (G-PCC), a deep learning method (IPDAE, PointSoup), and the method proposed in this invention (DiffPCC). The corresponding bit rate (Bpp) and reconstruction accuracy (PSNR) are also marked in the figure.

[0047] from Figure 4 From the results shown, it can be observed that G-PCC can achieve a certain reconstruction quality at high bit rates, but there are problems such as geometric distortion and blurred boundaries; IPDAE has a low compression rate, but the reconstruction results are not good in terms of edge details and structural integrity; PointSoup has improved in some areas, but the overall reconstruction still has details missing. In contrast, the DiffPCC proposed in the present invention has shown better structure restoration and detail recovery capabilities in multiple samples, especially at low bit rates. It can still effectively preserve the point cloud contours and key features, showing a better rate-distortion balance. Taking the chair point cloud in the figure as an example, DiffPCC achieves a PSNR of 39.888dB using only a compression rate of 0.38Bpp, which is significantly better than other methods; it can also maintain high-quality reconstruction in samples with complex geometric structures such as benches and armchairs, reduce information loss and structural breaks, and fully demonstrate the robustness and practicality of the present invention in complex point cloud reconstruction.

[0048] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A three-dimensional point cloud geometry compression method based on a diffusion model, characterized in that: The steps include: (1) Processing the collected point cloud data; (2) Construct a point cloud compression model DiffPCC, which includes a dual-domain encoder, a bottleneck layer module, and a diffusion decoder. The input point cloud is passed through the dual-domain encoder to obtain low-frequency features and high-frequency features, and skeleton points are extracted for spatial guidance. The bottleneck layer module adopts differentiated compression strategies for low-frequency features, high-frequency features, and skeleton points to form a complete compressed code stream. Multiple rounds of denoising and reconstruction are performed through the diffusion decoder. Low-frequency, high-frequency features, and skeleton points are used as conditional information, and fused through affine injection and attention mechanism to achieve the gradual recovery of point cloud structure and details. (3) The constructed point cloud compression model DiffPCC is trained, and the model parameters are end-to-end optimized using a rate-distortion loss function that includes chamfer distance and encoding bit rate to achieve efficient compression and decompression of three-dimensional point cloud data.

2. The method for 3D point cloud geometry compression based on a diffusion model according to claim 1, characterized in that: The implementation process of step (1) is as follows: The collected point cloud data is processed by centering and translation normalization, and uniformly mapped to the unit sphere; each point cloud is resampled to a fixed number.

3. The method for 3D point cloud geometry compression based on a diffusion model according to claim 1, characterized in that: The dual-domain encoder includes a low-frequency feature extraction branch and a high-frequency feature extraction branch: the low-frequency branch is based on an improved point cloud neural network PointNet, and extracts the overall structural information of the point cloud through multi-layer feature distillation and average pooling operations, generating low-frequency features with global geometric perception capabilities; the high-frequency branch uses a high-frequency arithmetic coding module, combined with point-level attention mechanism and block-level attention mechanism, to model the local change areas and inter-block structures in the point cloud, and extract fine-grained high-frequency features.

4. The method for 3D point cloud geometry compression based on a diffusion model according to claim 1, characterized in that: The bottleneck layer module includes a low-frequency feature entropy modeling module, a high-frequency feature joint modeling module and a skeleton point compression module; the low-frequency feature entropy modeling module performs probabilistic modeling on the low-frequency features through a full decomposition entropy model, and uses arithmetic coding to achieve compression; the high-frequency feature joint modeling module constructs a low-frequency-guided high-frequency entropy model under the guidance of low-frequency features, which is used to model its symbol probability distribution and complete arithmetic coding; the skeleton point compression module performs lossless compression through the G-PCC method based on the octree structure; the encoding results output by the three modules together constitute the final compressed code stream for storage and transmission, and after inverse process reconstruction, it is provided to the diffusion decoder for conditional reconstruction.

5. The method for 3D point cloud geometry compression based on a diffusion model according to claim 1, characterized in that: The diffusion decoder is constructed based on the conditional diffusion model, including a word segmenter module, a conditional information injection module and a point rearrangement module; the word segmenter module divides the initial noise point cloud into local regions through K-nearest neighbor clustering guided by skeleton points, and extracts regional-level Token representations; the conditional information injection module is used to inject the low-frequency features, high-frequency features and skeleton points generated by the encoder into the diffusion process as conditional information, wherein the low-frequency features are injected into the Token representation through affine transformation, and the high-frequency features are fused with the Token through the self-attention mechanism; the updated Token is subjected to feature propagation and residual update through graph convolution to form the input of the next iteration; the point rearrangement module is used to map the Token generated by the final iteration into three-dimensional point cloud coordinates to achieve the gradual restoration of structure and details.

6. The method for 3D point cloud geometry compression based on a diffusion model according to claim 1, characterized in that: The implementation process of step (3) is as follows: By simultaneously optimizing the dual-domain encoder, bottleneck layer module, and diffusion decoder module, unified modeling of compression coding and point cloud reconstruction is achieved; The rate-distortion loss function composed of chamfer distance and feature bit rate is used: Among them, P is the original point cloud, is the reconstructed point cloud, CD is the chamfer distance, R(f l ) and R(f h ) represent the bit rates of low-frequency features and high-frequency features respectively, and λ is used to balance distortion and compression rate.

7. A device, characterized in that: comprising a memory and a processor, wherein: a memory for storing computer programs capable of running on the processor; A processor is configured to execute the steps of the three-dimensional point cloud geometry compression method based on a diffusion model as described in any one of claims 1 to 6 when running the computer program.

8. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by at least one processor, implements the steps of the three-dimensional point cloud geometry compression method based on the diffusion model according to any one of claims 1 to 6.

Images