A Point Cloud Attribute Compression Method Based on Inter-Block Prediction and Graph Fourier Transform

Abstract

The present invention proposes a point cloud attribute compression method based on inter-block prediction and graph Fourier transform to solve the problem of improving compression performance while maintaining the "bit rate-distortion" balance. First, clustering is performed on the point cloud to divide the point cloud into different point cloud blocks. The present invention will encode the point cloud attribute information based on the point cloud blocks. Specifically, the color attribute of the point cloud block after region division is split into two parts: the attribute mean and the attribute residual. According to the different characteristics of the two parts of information, they are processed separately. For the attribute residual information, the present invention proposes a graph Fourier transform method to further enhance the sparsity of the transform coefficients, thereby obtaining better compression performance. For the attribute mean information, according to the correlation between the attribute means of adjacent point cloud blocks, prediction is used to eliminate the inter-block redundancy. Then, the attribute mean and the attribute residual after transformation and prediction processing are encoded and decoded, and reconstructed based on inverse transformation and inverse prediction.

Description

Technical Field:

[0001] The present invention relates to the field of computer image processing, and more particularly to a clustering-based point cloud attribute compression method using inter-block prediction and graph Fourier transform. Background Art:

[0002] Nowadays, point clouds are widely used in many fields such as 3D immersive telepresence, 3D broadcasting, cultural and heritage reconstruction. With the improvement of scanning device technology, point clouds play an increasingly important role in the 3D field. However, the point cloud sets we usually obtain are digital 3D models containing millions of points, which occupy a large amount of storage space. Therefore, how to adopt an efficient compression algorithm for effective storage and transmission is an urgent problem to be solved. The existing point cloud attribute compression methods can be divided into the following categories: traditional point cloud attribute compression methods, prediction-based point cloud attribute compression methods, and transform-based point cloud attribute compression methods.

[0003] 1) Traditional point cloud attribute compression methods: Directly compressing point clouds has lower costs compared to converting point clouds into polygon meshes. In addition to compressing geometric information, Huang et al. (Y. Huang, J. Peng, C.-J. Kuo and M. Gopi, "A Generic Scheme for Progressive Point Cloud Coding," in IEEE Transactions on Visualization and Computer Graphics, vol. 14, no. 2, pp. 440-453, March-April 2008, doi: 10.1109 / TVCG.2007.70441) found that compressing the color attributes of point clouds can better improve performance and compression effects. To encode structured point clouds, in the recent work of (J. Kammerl, N. Blodow, R.B. Rusu, S. Gedikli, M. Beetz and E. Steinbach, "Real-time compression of point cloud streams," 2012 IEEE International Conference on Robotics and Automation, 2012, pp. 778-785, doi: 10.1109 / ICRA.2012.6224647), real-time encoding is carried out based on the extension of the octree representation of unstructured point clouds. However, in terms of attribute compression, it cannot make good use of the correlation between points (R. Mekuria, K. Blom and P. Cesar, "Design, Implementation, and Evaluation of a Point Cloud Codec for Tele-Immersive Video," in IEEE Transactions on Circuits and Systems for Video Technology, vol. 27, no. 4, pp. 828-842, April 2017, doi: 10.1109 / TCSVT.2016.2543039).

[0004] 2) Transform-based point cloud attribute compression method: Zhang et al. proposed an efficient point cloud compression method based on graph Fourier transform in (C. Zhang, D. Florêncio and C. Loop, "Point cloud attribute compression with graph transform," 2014 IEEE International Conference on Image Processing (ICIP), 2014, pp. 2066 - 2070, doi: 10.1109 / ICIP.2014.7025414), which effectively reduces the correlation of signals compared to the traditional discrete cosine transform. When the point cloud is sparse, the way of constructing the graph will create many isolated subgraphs. Using sparse representation to acquire, represent, and compress high-dimensional signals helps to write the signal as a linear combination of several atoms from a pre-specified basis or dictionary, which saves a lot of space. Moreover, the sparse principle also plays an important role in data modeling. Therefore, the technique of utilizing the sparsity of signals in the transform domain or dictionary to process signals is a current research hotspot, such as graph Fourier transform, discrete cosine transform, and wavelet transform to redundant dictionaries (R. Rubinstein, A. M. Bruckstein and M. Elad, "Dictionaries for Sparse Representation Modeling," in Proceedings of the IEEE, vol. 98, no. 6, pp. 1045 - 1057, June 2010, doi: 10.1109 / JPROC.2010.2040551). In (Y. Xu et al., "Predictive Generalized Graph Fourier Transform for Attribute Compression of Dynamic Point Clouds," in IEEE Transactions on Circuits and Systems for Video Technology, vol. 31, no. 5, pp. 1968 - 1982, May 2021, doi: 10.1109 / TCSVT.2020.3015901), the generalized graph Fourier transform is proven to be the optimal spatio-temporal decorrelation prediction transform.

[0005] 3) Prediction-based Point Cloud Attribute Compression Method: Compressing point cloud attributes through prediction is also a common compression method. R.A. Cohen et al. proposed their method using 3-D block-based prediction and transform coding to compress point cloud attributes in (R.A. Cohen, D. Tian and A. Vetro, "Attribute compression for sparse point clouds using graph transforms," 2016 IEEE International Conference on Image Processing (ICIP), 2016, pp. 1374-1378, doi: 10.1109 / ICIP.2016.7532583). S. Gu et al. proposed graph prediction in (S. Gu, J. Hou, H. Zeng, H. Yuan and K.-K. Ma, "3D Point Cloud Attribute Compression Using Geometry-Guided Sparse Representation," in IEEE Transactions on Image Processing, vol. 29, pp. 796-808, 2020, doi: 10.1109 / TIP.2019.2936738). They selected a small number of representative points from previously encoded point cloud blocks and constructed an underlying graph structure based on geometric information to predict the points to be encoded. In addition, the input voxelized 3D point cloud was divided into blocks of equal size. Between the two, an inter-block prediction scheme was applied to eliminate redundancy. C. Ma et al. proposed a prediction scheme based on fast recoloring technology for lossless and near-lossless compression of attributes that fully utilized the correlation with neighborhood attributes in (C. Ma, G. Li, Q. Zhang, Y. Shao, J. Wang and S. Liu, "Fast Recolor Prediction Scheme in Point Cloud Attribute Compression," 2020 IEEE International Conference on Visual Communications and Image Processing (VCIP), 2020, pp. 50-53, doi: 10.1109 / VCIP49819.2020.9301768). Their method not only utilized the geometric relationship based on Euclidean distance but also fully considered the potential geometric distribution.

[0006] In summary, traditional point cloud attribute compression methods have low cost but cannot make good use of the correlation between points during attribute compression; point cloud attribute compression methods based on transformation can effectively save the bit rate by utilizing the sparsity of sparse point clouds, but cause relatively serious distortion when dealing with dense point clouds; point cloud attribute compression methods based on prediction can effectively remove redundancy when dealing with dense point clouds but increase the bit rate when dealing with sparse point clouds. If the advantages of point cloud attribute compression methods based on transformation and those based on prediction can be combined when dealing with point clouds to better utilize the correlation between point cloud data, there is hope to achieve a better "bit rate - distortion" balance when compressing point clouds. Summary of the Invention:

[0007] The present invention proposes a point cloud attribute compression method based on inter - block prediction and graph Fourier transform. The aim is to improve the reconstruction quality while reducing the bits per pixel (bpp) required for encoding and decoding by utilizing the color attribute correlation between points, and to have better compression performance on the premise of maintaining the "bit rate - distortion" balance.

[0008] The present invention proposes a point cloud attribute compression method based on inter - block prediction and graph Fourier transform. First, the present invention performs clustering on the point cloud and divides the point cloud into different point cloud blocks. The present invention encodes the point cloud attribute information based on the point cloud blocks. Specifically, the color attribute of the point cloud block after region division is split into two parts: the attribute mean and the attribute residual. According to the different characteristics of the two parts of information, they are processed separately. For the attribute residual information, the present invention proposes a graph Fourier transform method to further enhance the sparsity of the transform coefficients, thereby obtaining better compression performance. For the attribute mean information, according to the correlation between the attribute means of adjacent point cloud blocks, a prediction method is used to eliminate the redundancy between blocks.

[0009] The steps of the above - mentioned method are as follows:

[0010] Step 1: Perform region division on the point cloud, divide the point cloud into different point cloud blocks, and decompose the color attribute of each point cloud block into the attribute mean of each point cloud block and the attribute residual of each point.

[0011] Step 2: Perform geometric - guided graph Fourier transform on the attribute residual to obtain the attribute residual transform sequence and encode it.

[0012] Step 3: Obtain the reference point cloud attribute mean and the attribute mean difference sequence through inter - block prediction for the attribute mean and encode them.

[0013] Step 4: Decode and reconstruct the encoded attribute residual transform sequence, the reference point cloud attribute mean, and the attribute mean difference sequence.

[0014] The specific operations of each step are as follows:

[0015] Specific operations in Step 1: Use the K-means algorithm for clustering according to the geometric information of the points in the point cloud, and divide the point cloud into K point cloud blocks.

[0016] Specifically, select K points from the point cloud as the clustering centers, and calculate the distances from all points in the point cloud to these K clustering centers. For each point, compare its distances to the K clustering centers and divide it into the cluster where the nearest clustering center is located, so as to determine the composition of the K clusters.

[0017] Use the cosine similarity between the two normal vectors corresponding to the point and the clustering center to judge the distance between the point and the clustering center. For the normal vector of any point in the point cloud The normal vector of the selected clustering center And the included angle θ between the two normal vectors, calculate the cosine value cosθ of the two normal vectors. The closer the cosine value cosθ is to 1, the greater the cosine similarity between the two normal vectors, that is, the closer the distance between the points to which the two normal vectors belong and the clustering center.

[0018] Select all the points closest to a certain clustering center and the clustering center to jointly form the same cluster, and define the cluster as a point cloud block, that is, the point cloud block is composed of all points that meet the distance condition and the corresponding clustering center.

[0019] From the generation of K clusters, it can be seen that the point cloud is divided into K point cloud blocks.

[0020] For each point cloud block, calculate the mean values of its three color channels Y, U, and V to form the attribute mean; for each point, use the difference between its own color attribute and the attribute mean of the point cloud block to which it belongs to obtain the attribute residual of each point.

[0021] Specific operations in Step 2: For the attribute residuals of each point decomposed in Step 1, the present invention proposes a geometric-guided graph Fourier transform for the attribute residuals.

[0022] In order to utilize the correlation between points, the present invention uses the normal vectors corresponding to the points to describe the geometric similarity between two points. In order to calculate the geometric similarity between two points, we introduce the concept of a graph. For each point cloud block, it is defined as a weighted undirected graph G(V, E, W). Define the points in the point cloud as the nodes in the graph, and V is the set of all nodes in the weighted undirected graph G; connect any two points i and j in the point cloud, and define this connection as the edge between the corresponding nodes V i and V j in the graph, and E is the set of all edges in the weighted undirected graph G; define the edge weight connecting the nodes V i and V j as W i,jused to calculate the geometric similarity between points i and j, where W is a set of edge weights and also a symmetric adjacency matrix, and W i,j = W j,i .

[0023] Define the angle between the two normal vectors corresponding to point i and point j as θ i,j , and its cosine value is cosθ i,j , e is the base of the natural logarithm function, and σ is a weight parameter. Define W i,j 's calculation formula as:

[0024]

[0025] Further define the degree matrix D to record the degree of each node in the weighted undirected graph G. Therefore, the degree matrix D is a diagonal matrix, where the p-th diagonal element D(p,p) is the sum of all elements in the p-th row of the adjacency matrix W:

[0026]

[0027] Then use the degree matrix D and the adjacency matrix W to calculate the graph Laplacian matrix L:

[0028] L = D - W (3)

[0029] Perform eigenvalue decomposition on the graph Laplacian matrix L to obtain the eigenvector matrix Φ and the eigenvalue matrix A:

[0030] L = ΦAΦ T (4)

[0031] Take the eigenvector matrix Φ as the graph Fourier transform basis, and project the attribute residuals of each point in the Y, U, and V color channels into the graph Fourier transform domain. Take the attribute residual Y Y of the point in the Y color channel as an example to perform graph Fourier transform and calculate the transform coefficient There is:

[0032]

[0033] Perform the same projection transformation on the attribute residuals of each point in the U and V color channels and the attribute residuals of the Y color channel to obtain the transform coefficients. The transform coefficients of the three channels form an attribute residual transform sequence.

[0034] Use an arithmetic encoder to encode the attribute residual transform sequence.

[0035] Specific operations in Step 3: Predict the attribute means of the K point cloud blocks decomposed in Step 1.

[0036] Arrange the attribute means of the K point cloud blocks into an R K×1A column vector, where the t-th element of the column vector is the attribute mean of the t-th point cloud block.

[0037] Select the attribute mean of the first point cloud block as the reference point cloud attribute mean, that is, select the first element of the column vector as the reference element. Starting from the first point cloud block, calculate the difference between all adjacent elements in the column vector as the mean difference of the color attributes between two adjacent point cloud blocks. All the differences form a sequence of attribute mean differences.

[0038] Taking a point cloud containing 100 point cloud blocks as an example, the attribute means are arranged into an R 100×1 column vector. Select the attribute mean of the first point cloud block as the reference point cloud attribute mean, that is, select the first element of the column vector as the reference element. Starting from the first point cloud block, calculate the difference between all adjacent elements in the column vector as the mean difference of the color attributes between two adjacent point cloud blocks, that is, use the first element in the column vector of attribute means to subtract the second element, the second element subtract the third element... and so on until the ninety-ninth element subtracts the one-hundredth element, obtaining a total of ninety-nine differences, that is, the mean differences of the color attributes between ninety-nine adjacent point cloud blocks. All the differences form a sequence of attribute mean differences containing ninety-nine elements.

[0039] Use an arithmetic encoder to encode the reference point cloud attribute mean and the sequence of attribute mean differences.

[0040] Specific operations in Step 4: At the decoder, decode the encoded attribute residual transformation sequence obtained in Step 2 and the encoded reference point cloud attribute mean and the sequence of attribute mean differences obtained in Step 3 respectively.

[0041] For the decoded attribute residual transformation sequence, use the feature vector matrix Φ as the inverse graph Fourier transform basis again for inverse graph Fourier transform to reconstruct the attribute residuals of each point from the attribute residual transformation sequence. Taking the transformed coefficient of the point in the Y color channel after decoding as an example, calculate the attribute residual Y' Y :

[0042]

[0043] For the decoded reference point cloud attribute mean and the sequence of attribute mean differences, add the reference point cloud attribute mean to the corresponding elements in the sequence of attribute mean differences to predict the attribute means of K point cloud blocks.

[0044] Since each point has its corresponding point cloud block, the attribute mean of the point cloud block to which a certain point belongs is used as the attribute mean of that point, and the color attribute of each point is obtained by adding the attribute residual of each point to the attribute mean.

[0045] Compared with the existing technologies, the present invention proposes a dual-branch model, which divides the color attributes of the point cloud into two parts: attribute residuals and attribute means. Through the designed Weighted Graph Fourier Transform (NWGFT), the sparsity of the point cloud is utilized to transform the attribute residuals, and the inter-block prediction of the attribute mean relationship between point cloud blocks is performed by using the color attribute correlation of points. On the premise of reducing the distortion of color attributes, the bit rate required for encoding and decoding is reduced, and point cloud compression with better performance is carried out. Description of the Drawings:

[0046] Figure 1 Point cloud attribute compression framework diagram based on inter-block prediction and graph Fourier transform;

[0047] Figure 2(a) RD curves of the method proposed in the present invention, NWGFT, and RAHT on the Andrew dataset;

[0048] Figure 2(b) RD curves of the method proposed in the present invention, NWGFT, and RAHT on the Ricardo dataset

[0049] Figure 3(a) Subjective rendering result diagram of the original point cloud;

[0050] Figure 3(b) Comparative diagram of subjective rendering results of the point cloud reconstructed using the NWGFT method under similar bpp;

[0051] Figure 3(c) Comparative diagram of subjective rendering results of the point cloud reconstructed using the RAHT method under similar bpp;

[0052] Figure 3(d) Comparative diagram of subjective rendering results of the point cloud reconstructed using the method of the present invention under similar bpp. Detailed Implementation Manner:

[0053] In order to be able to more clearly describe the technical content of the present invention, the following will be further described in combination with specific examples:

[0054] The framework diagram of the present invention is as Figure 1 , and the specific implementation process is divided into two stages, the encoding stage and the decoding stage.

[0055] I. Encoding Stage

[0056] The encoding stage is divided into three steps: dividing the point cloud into multiple point cloud blocks and decomposing its color attributes into the attribute mean of each point cloud block and the attribute residual of each point, performing geometric-guided graph Fourier transform on the attribute residuals to obtain the attribute residual transform sequence and encoding it, and obtaining the reference point cloud attribute mean and the attribute mean difference sequence through inter-block prediction of the attribute means and encoding them.

[0057] 1. Divide the point cloud into multiple point cloud blocks and decompose its color attributes into the attribute mean of each point cloud block and the attribute residual of each point

[0058] The K-means algorithm is used to cluster the points according to their geometric information and divide the point cloud into K point cloud blocks.

[0059] Specifically, K points are selected from the point cloud as cluster centers, and the distances from all points in the point cloud to the K cluster centers are calculated. For each point, its distance to the K cluster centers is compared and the point is divided into the cluster with the closest cluster center, thereby determining the composition of the K clusters.

[0060] The distance between a point and the cluster center is determined by the cosine similarity between the two normal vectors corresponding to the point and the cluster center. Normal vector of the selected cluster center And the angle θ between the two normal vectors, calculate the cosine value cosθ of the two normal vectors. The closer the cosine value cosθ is to 1, the greater the cosine similarity between the two normal vectors, that is, the closer the distance between the point to which the two normal vectors belong and the cluster center is.

[0061] All points closest to a cluster center are selected to form the same cluster together with the cluster center, and the cluster is defined as a point cloud block, that is, the point cloud block consists of all points that meet the distance conditions and the corresponding cluster center.

[0062] From the generation of K clusters, it can be seen that the point cloud is divided into K point cloud blocks.

[0063] For each point cloud block, the mean of its three color channels, Y, U, and V, is calculated to form the attribute mean. For each point, the attribute residual of each point is obtained by subtracting its own color attribute from the attribute mean of the point cloud block to which it belongs.

[0064] 2. Perform a geometric guided graph Fourier transform on the attribute residual to obtain the attribute residual transformation sequence and encode it

[0065] For the attribute residual of each point decomposed in step 1, the present invention proposes a geometric guidance graph Fourier transform for the attribute residual.

[0066] In order to utilize the correlation between points, the present invention uses the normal vector corresponding to the point to describe the geometric similarity between two points. In order to calculate the geometric similarity between two points, we introduce the concept of graph. For each point cloud block, it is defined as a weighted undirected graph G(V,E,W). The points in the point cloud are defined as nodes in the graph, V is the set of all nodes in the weighted undirected graph G; connect any two points i and j in the point cloud, and define this line as the corresponding node V of these two points i and V jThe edges in the graph, where E is the set of all edges in the weighted undirected graph G; it is defined that the edge weight connecting nodes V i and V j is W i,j which is used to calculate the geometric similarity between points i and j. W is also a set of edge weights and a symmetric adjacency matrix, where W i,j = W j,i .

[0067] Define the angle between the two normal vectors corresponding to point i and point j as θ i,j , and its cosine value is cosθ i,j . e is the base of the natural logarithm function, and σ is a weight parameter. Define the calculation formula of W i,j as:

[0068]

[0069] Furthermore, define the degree matrix D to record the degree of each node in the weighted undirected graph G. Therefore, the degree matrix D is a diagonal matrix, where the p-th diagonal element D(p, p) is the sum of all elements in the p-th row of the adjacency matrix W:

[0070]

[0071] Then use the degree matrix D and the adjacency matrix W to calculate the graph Laplacian matrix L:

[0072] L = D - W (3)

[0073] Perform eigenvalue decomposition on the graph Laplacian matrix L to obtain the eigenvector matrix Φ and the eigenvalue matrix A:

[0074] L = ΦAΦ T (4)

[0075] Take the eigenvector matrix Φ as the graph Fourier transform basis, and project the attribute residuals of each point in the Y, U, and V color channels into the graph Fourier transform domain. Taking the attribute residual Y Y of the point in the Y color channel as an example for graph Fourier transform, calculate the transformation coefficient There is:

[0076]

[0077] For the attribute residuals of each point in the U and V color channels and the attribute residuals of the Y color channel, perform the same projection transformation to obtain the transformation coefficients. The transformation coefficients of the three channels form the attribute residual transformation sequence.

[0078] Use an arithmetic encoder to encode the attribute residual transformation sequence.

[0079] 3. Predict the attribute means of the K point cloud blocks decomposed in Step 1 to obtain the reference point cloud attribute mean and the attribute mean difference sequence, and encode them.

[0080] Predict the attribute means of the K point cloud blocks decomposed in the first step.

[0081] Arrange the attribute means of the K point cloud blocks into an R K×1 column vector, where the t-th element of the column vector is the attribute mean of the t-th point cloud block.

[0082] Randomly select the attribute mean of the m-th point cloud block as the reference point cloud attribute mean, that is, select the m-th element of the column vector as the reference element, and subtract it from the remaining (K - 1) elements to obtain the mean difference of the color attributes between the K point cloud blocks. All the mean differences form the attribute mean difference sequence.

[0083] Use an arithmetic encoder to encode the reference point cloud attribute mean and the attribute mean difference sequence.

[0084] II. Decoding stage

[0085] At the decoder, decode the encoded attribute residual transform sequence obtained in Step 2 and the encoded reference point cloud attribute mean and attribute mean difference sequence obtained in Step 3 respectively.

[0086] For the decoded attribute residual transform sequence, use the feature vector matrix Φ as the inverse graph Fourier transform basis for the inverse graph Fourier transform again to reconstruct the attribute residual of each point from the attribute residual transform sequence. Taking the decoded transform coefficient of the point in the Y color channel as an example, calculate the attribute residual Y' R :

[0087]

[0088] For the decoded reference point cloud attribute mean and attribute mean difference sequence, add the reference point cloud attribute mean to the corresponding elements in the attribute mean difference sequence to predict the attribute means of the K point cloud blocks.

[0089] Since each point belongs to a specific point cloud block, use the attribute mean of the point cloud block to which a point belongs as the attribute mean of that point, and add the attribute residual of each point to the attribute mean to obtain the color attribute of each point.

[0090] Performance evaluation:

[0091] Now, the present invention is compared with two state-of-the-art transform-based point cloud attribute compression methods, namely NWGFT (Y. Xu et al., "Cluster-Based Point Cloud Coding with Normal Weighted Graph Fourier Transform," 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018, pp. 1753-1757, doi: 10.1109 / ICASSP.2018.8462684) and RAHT (G. Sandri, R. De Queiroz and P. A. Chou, "Compression of Plenoptic Point Clouds Using the Region-Adaptive Hierarchical Transform," 2018 25th IEEE International Conference on Image Processing (ICIP), 2018, pp. 1153-1157, doi: 10.1109 / ICIP.2018.8451367). The method of Y. Xu is a method for compressing attributes using the normal NWGFT. The method of G. Sandri is a wavelet-based attribute compression method.

[0092] As shown in Figs. 2(a) and 2(b), the rate-distortion (RD) curves of the present invention and the two methods of NWGFT and RAHT on the datasets Andrew and Ricardo are shown. As shown in Fig. 2(a), the method proposed by the present invention reduces the bit rate (Bits Per Pixel, bpp) by 6.49% compared with RAHT and by 2.31% compared with NWGFT. These figures are calculated using BD-BR (Gisle "Improvement of BD-PSNR model," VCEG-AI11, July 2008.), which quantifies the difference between the two RD curves. Since RAHT is a competitive method with good performance, the performance improvement data compared with RAHT are further listed in Table 1.

[0093] In addition, we show the subjective results of the original point cloud and the reconstructed point clouds of the three methods of the present invention, NWGFT, and RAHT under similar bpp in the form of effect diagrams in FIGS. 3(a), 3(b), 3(c), and 3(d), and attach the bpp and peak signal-to-noise ratio (PSNR) used to compress the point cloud under each subjective result. PSNR is an objective standard for evaluating images, and a higher value means better image quality. From the intuitive comparison of the effects, it can be seen that there is an obvious mosaic effect in FIG. 3(c), while the algorithm proposed by the present invention retains more details in the data, and FIG. 3(d) has a better effect. Therefore, it can be concluded that the present invention can reduce the distortion caused by compression to a greater extent under similar bpp. There is no visible difference in the subjective images between FIG. 3(b) and FIG. 3(d), but from the data below the subjective results, it can be seen that for the same data set, this method obtains a higher PSNR while using a smaller bpp. It can be concluded that this method has better compression performance.

[0094] The experimental results confirm that the present invention makes more effective use of the correlation between points, reduces the bpp required for encoding and decoding without causing distortion of color attributes, has better reconstruction quality, achieves better compression performance on the premise of maintaining the "bit rate-distortion" balance, and has significant improvements compared with the state-of-the-art point cloud compression methods.

[0095] Table 1 Performance of the present invention under the partitioning step sizes {100, 200, 500} (compared with RAHT)

[0096]

[0097]

Claims

1. A point cloud attribute compression method based on inter-block prediction and graph Fourier transform, characterized in that The following steps are involved: Step 1: Divide the point cloud into different point cloud blocks and decompose its color attributes into the attribute mean of each point cloud block and the attribute residual of each point; Step 2: Perform geometric guided graph Fourier transform on the attribute residual to obtain the attribute residual transformation sequence and encode it; Step 3: Obtain the attribute mean and attribute mean difference sequence of the reference point cloud through inter-block prediction of the attribute mean, and encode them; Step 4: Decode and reconstruct the encoded attribute residual transformation sequence, the reference point cloud attribute mean, and the attribute mean difference sequence; Step 1 further includes: clustering the point cloud using a K-means algorithm according to geometric information of points in the point cloud, dividing the point cloud into K point cloud blocks; Specifically, K points are selected from the point cloud as cluster centers, and the distances from all points in the point cloud to the K cluster centers are calculated. For each point, the distance between the point and the K cluster centers is compared and the point is divided into the cluster where the nearest cluster center is located, thereby determining the composition of the K clusters; wherein the distance is the cosine similarity between the two normal vectors corresponding to the point and the cluster center, and the greater the cosine similarity between the two normal vectors, the closer the distance; each cluster is defined as a point cloud block; For each point cloud block, the mean of its three color channels, Y, U, and V, is calculated to form the attribute mean. For each point, the attribute residual of each point is obtained by subtracting its own color attribute from the attribute mean of the point cloud block to which it belongs.

2. The point cloud attribute compression method based on inter-block prediction and graph Fourier transform according to claim 1, wherein The following steps are involved: Step 2 further includes, For each point cloud block, it is defined as a weighted undirected graph \(G(V, E, W)\), where \(V=\{V i \}\) is the set of all nodes in the weighted undirected graph \(G\). The points in the point cloud block are defined as the nodes in the graph, and \(V i \) represents the \(i\)-th node in the graph \(G\), that is, the \(i\)-th point in the point cloud block; \(E = \{e i,j \}\) is the set of all edges in the weighted undirected graph \(G\), and \(e i,j \) refers to the edge between the nodes \(V i \) and \(V j \) in the graph \(G\), that is, the connection line between point \(i\) and point \(j\) in the point cloud; \(W=\{W i,j \}\) is the set of edge weights, which is a symmetric adjacency matrix. Among them, \(W i,j \) represents the edge weight connecting the nodes \(V i \) and \(V j \), and is used to calculate the geometric similarity between point \(i\) and point \(j\) in the point cloud. The specific calculation formula is as follows where θ i,j is the angle between two normal vectors corresponding to points i and j in the point cloud, e is the base of the natural logarithm function, and σ is a weight parameter; Furthermore, the degree matrix D is calculated to record the degree of each node in the weighted undirected graph G. The degree matrix D is a diagonal matrix, where the p-th diagonal element D(p,p) is the sum of all elements in the p-th row of the adjacency matrix W: Then, the graph Laplacian matrix L is calculated using the degree matrix D and the adjacency matrix W: L=DW (3) Perform eigendecomposition on the graph Laplace matrix L to obtain the eigenvector matrix Φ and eigenvalue matrix A: L = ΦAΦ T (4) The eigenvector matrix Φ is used as the graph Fourier transform basis, and the attribute residuals of the three color channels Y, U, and V of each point are projected into the graph Fourier transform domain to calculate the transformation coefficients. The calculation formula is uniformly expressed as follows: Among them, H represents a color channel, H ∈ {Y, U, V}, and Y H represents the attribute residual corresponding to the color channel. The transformation coefficients of the three channels form an attribute residual transformation sequence The attribute residual transform sequence is encoded using an arithmetic encoder.

3. The point cloud attribute compression method based on inter-block prediction and graph Fourier transform according to claim 2, characterized in that: Step three further includes: Predict the attribute mean of the K point cloud blocks decomposed in step 1; Arrange the attribute means of K point cloud blocks into a column vector R K×1 , where the t-th element of the column vector is the attribute mean of the t-th point cloud block; The attribute mean of the first point cloud block is selected as the attribute mean of the reference point cloud, that is, the first element of the column vector is selected as the reference element; Starting from the first point cloud block, the difference between all adjacent elements in the column vector is calculated as the mean difference of the color attribute between two adjacent point cloud blocks, and all the differences constitute the attribute mean difference sequence; The base point cloud attribute means and attribute mean difference sequences are encoded using an arithmetic encoder.

4. The point cloud attribute compression method based on inter-block prediction and graph Fourier transform according to claim 3, characterized in that: Step 4 further includes: at the decoder, decoding the encoded attribute residual transformation sequence obtained in Step 2 and the encoded reference point cloud attribute mean and attribute mean difference sequence obtained in Step 3 respectively; For the decoded attribute residual transformation sequence, the feature vector matrix Φ is used again as the graph Fourier transform basis for the inverse graph Fourier transform to reconstruct the attribute residual transformation sequence to obtain the attribute residual of each point. The calculation formula is uniformly expressed as follows: H represents a color channel, H ∈ {Y, U, V}, Y′ H represents the reconstructed attribute residual corresponding to the color channel; For the decoded reference point cloud attribute mean and attribute mean difference sequence, add the corresponding elements in the reference point cloud attribute mean and the attribute mean difference sequence, that is, predict the attribute means of K point cloud blocks; Since each point has its own belonging point cloud block, the predicted attribute mean of the point cloud block to which a certain point belongs is used as the attribute mean of the point. The reconstructed attribute residual and the predicted attribute mean of each point are the color attributes of each point after compression.

Images