Three-dimensional volume data compression and reconstruction method for scientific visualization

The three-dimensional volume data is trained and reconstructed through neural network models, which solves the problem of high storage and transmission costs of three-dimensional volume data, and realizes efficient three-dimensional volume data reconstruction and training costs, ensuring high-quality consistency of the reconstruction data.

CN118941712APending Publication Date: 2024-11-12TIANJIN UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410965431.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-18
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The high storage and transmission cost of three-dimensional volume data limits its wide application in more fields, and the prior art is difficult to effectively improve the efficiency of three-dimensional volume data reconstruction and reduce training costs.

Method used

By training three-dimensional body data based on neural network model, the model is constructed using the original data of the sampling point, the prediction data is output for reconstruction, and the reconstruction quality is measured by the peak signal-to-noise ratio, and the Adamax optimizer and specific loss function are used to improve reconstruction efficiency .

Benefits of technology

Three-dimensional volume data reconstruction with high approximation is realized, which significantly improves the reconstruction efficiency and reduces training costs, ensuring high-quality consistency between the reconstruction data and the original data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941712A_ABST
    Figure CN118941712A_ABST
Patent Text Reader

Abstract

The invention discloses a three-dimensional volume data compression and reconstruction method for scientific visualization, and the method comprises the steps: obtaining the original voxel data of all leaf nodes based on a compressed three-dimensional volume data set, extracting a part of leaf nodes as sampling points, and taking the original three-dimensional volume data of the sampling points as sample three-dimensional volume data; and constructing a neural network model, training the neural network model based on the sample three-dimensional volume data, outputting predicted three-dimensional volume data, reconstructing the predicted three-dimensional volume data, and obtaining reconstructed three-dimensional volume data. The method provided by the invention can effectively improve the three-dimensional volume data reconstruction efficiency and reduce the training cost.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional volume data compression and reconstruction, and in particular to a three-dimensional volume data compression and reconstruction method for scientific visualization. Background Art

[0002] With the continuous development of computer graphics and computer vision, implicit neural representation has been widely used. Implicit neural representation can be used to achieve realistic visual effects in computer graphics processing, including: using implicit neural representation to generate or reconstruct complex three-dimensional objects, and using implicit neural representation methods in ray tracing to build and represent complex geometric shapes and material properties in the scene, thereby improving the realism of three-dimensional object rendering effects; in addition, implicit neural representation can also be used to compress volume scalar fields.

[0003] As a data type with high dimensionality and complexity, three-dimensional volume data has important uses in many fields such as computer graphics, medical imaging diagnosis, geology, space physics, chemistry, and materials science. A lot of valuable information can be obtained from it. As an effective method to extract meaningful information from volume data sets through interactive graphics or image technology, volume data visualization has become an increasingly important field in scientific data visualization. However, due to its unique properties, the storage and transmission costs of three-dimensional volume data are often high, which undoubtedly increases the difficulty of data processing and also limits its wide application in more fields. Summary of the invention

[0004] In view of this, an object of the present invention is to provide a three-dimensional volume data compression and reconstruction method for scientific visualization, wherein the method can effectively improve the efficiency of three-dimensional volume data reconstruction and reduce training costs.

[0005] The embodiment of the present invention provides a method for compressing and reconstructing three-dimensional volume data for scientific visualization, which is characterized by comprising:

[0006] S1: acquiring original voxel data of all leaf nodes based on the compressed three-dimensional volume data set, extracting some of the leaf nodes as sampling points, and using the original three-dimensional volume data of the sampling points as sample three-dimensional volume data;

[0007] S2: constructing a neural network model, training the neural network model based on the sample three-dimensional volume data, outputting predicted three-dimensional volume data, reconstructing the predicted three-dimensional volume data, and obtaining reconstructed three-dimensional volume data.

[0008] Preferably, S1 specifically includes the following steps:

[0009] S101: compressing a three-dimensional volume data set based on a compression level to obtain the compressed three-dimensional volume data set, where the compression level includes: high compression, medium compression, and low compression;

[0010] S102: Read the compressed three-dimensional volume data set to obtain the original three-dimensional volume data of all the leaf nodes;

[0011] S103: extracting part of the leaf nodes as the sampling points based on the randomly generated index, and using the original three-dimensional volume data of the sampling points as the sample voxel data.

[0012] Preferably, S2 specifically includes the following steps:

[0013] S201: constructing the neural network model based on the coefficient field, the basis field and the projection function;

[0014] S202: training the neural network model based on the sample three-dimensional volume data, and outputting predicted three-dimensional volume data;

[0015] S203: Reconstruct the predicted voxel values ​​and coordinate values ​​in the predicted three-dimensional volume data to obtain reconstructed three-dimensional volume data.

[0016] Preferably, the neural network model uses Adamax optimizer during training and uses 2e -2 As the learning rate, the formula for the loss function is:

[0017]

[0018] Where, L MAE That is, the loss function is the loss calculated using the mean absolute error; f(x i ) is the predicted voxel value; yi is the original voxel value.

[0019] Preferably, the method further comprises: measuring the quality difference between the reconstructed three-dimensional volume data and the original three-dimensional volume data based on a peak signal-to-noise ratio.

[0020] Preferably, the formula for the peak signal-to-noise ratio is:

[0021]

[0022] Wherein, MAX is the maximum value of the voxel value; MSE is the mean square error, that is, the average value of the square of the difference between the reconstructed three-dimensional volume data and the sample three-dimensional volume data.

[0023] The present invention has the following beneficial effects: the method provided by the present invention can reconstruct three-dimensional volume data with a high degree of similarity to the original three-dimensional volume data, and the method has a fast training speed, which can effectively improve the efficiency of three-dimensional volume data reconstruction and reduce training costs.

[0024] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0025] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 A schematic diagram of a flow chart of a method for compressing and reconstructing three-dimensional volume data for scientific visualization provided by an embodiment of the present invention;

[0028] Figure 2a A schematic diagram of visualization of original three-dimensional volume data obtained from a 300×180×120 data set with a high compression level provided in an embodiment of the present invention;

[0029] Figure 2b A visualization diagram of reconstructed three-dimensional volume data obtained from a 300×180×120 data set with a high compression level provided in an embodiment of the present invention;

[0030] Figure 3a A schematic diagram of visualization of original three-dimensional volume data obtained from a 300×180×120 data set with a medium compression level provided in an embodiment of the present invention;

[0031] Figure 3b A visualization diagram of reconstructed three-dimensional volume data obtained from a 300×180×120 data set with a medium compression level provided in an embodiment of the present invention;

[0032] Figure 4a A schematic diagram of visualization of original three-dimensional volume data obtained from a 300×180×120 data set with a low compression level provided in an embodiment of the present invention;

[0033] Figure 4b A visualization diagram of reconstructed three-dimensional volume data obtained from a 300×180×120 data set with a low compression level provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] To facilitate understanding of this embodiment, Figure 1 A three-dimensional volume data compression and reconstruction method for scientific visualization disclosed in an embodiment of the present invention is introduced in detail.

[0036] Embodiment 1:

[0037] This embodiment provides a method for compressing and reconstructing three-dimensional volume data for scientific visualization, including the following steps:

[0038] S1: acquiring original voxel data of all leaf nodes based on the compressed 3D volume data set, extracting some leaf nodes as sampling points, and using the original 3D volume data of the sampling points as sample 3D volume data;

[0039] S2: construct a neural network model, train the neural network model based on the sample three-dimensional volume data, output predicted three-dimensional volume data, reconstruct the predicted three-dimensional volume data, and obtain reconstructed three-dimensional volume data.

[0040] The following is a detailed description of the above steps:

[0041] S1: Based on the compressed three-dimensional volume data set, the original voxel data of all leaf nodes are obtained, some leaf nodes are extracted as sampling points, and the original voxel data of the sampling points are used as sample voxel data.

[0042] Among them, S1 specifically includes the following steps:

[0043] S101: compressing the three-dimensional volume data set based on a compression level to obtain a compressed three-dimensional volume data set, where the compression level includes: high compression, medium compression, and low compression;

[0044] S102: Read the compressed 3D volume data set to obtain original 3D volume data of all leaf nodes;

[0045] The original three-dimensional volume data includes the coordinate values ​​of the leaf nodes and the original voxel values, and the original voxel values ​​can correspond to the voxel colors after data visualization.

[0046] S103: extracting some leaf nodes as sampling points based on the randomly generated index, and using the original three-dimensional volume data of the sampling points as sample voxel data.

[0047] The sample three-dimensional volume data includes the coordinate values ​​of the sampling points and the sample voxel values.

[0048] S2: construct a neural network model, train the neural network model based on the sample three-dimensional volume data, output predicted three-dimensional volume data, reconstruct the predicted three-dimensional volume data, and obtain reconstructed three-dimensional volume data.

[0049] Wherein, S2 includes the following steps:

[0050] S201: constructing a neural network model based on the coefficient field, the basis field and the projection function;

[0051] In this embodiment, the neural network model is a multilayer perceptron (MLP) constructed based on coefficient fields, basis fields and projection functions through two fully connected layers of three ReLU activation functions, which is used to capture complex patterns and features in sample three-dimensional volume data through a learning process.

[0052] Among them, the coefficient field is composed of coefficient domain factors with identity transformation, and the coefficient domain factors are used to express local spatial variation characteristics; the basis field is composed of basis function factors with periodic transformation, and the basis function is used to simulate the commonality of shared patterns in the entire sample three-dimensional volume data. When the coefficient domain factors are combined with the basis function factors, they can represent global and local characteristics, thereby obtaining similar features in space and across scales, and regressing the target signal from the factor product between the coefficient domain factors and the basis function factors through the learned projection function.

[0053] Further, the same basis function is applied at multiple positions and scales using the periodic coordinate transformation γ, and the basis function is a sawtooth function. The sawtooth function has periodicity and continuity and can well describe the periodic structure in the signal. When the same basis function is applied at multiple positions and scales using the periodic coordinate transformation γ, since the coefficients are freely changed in the signal domain, γi(x)=x is used to perform the corresponding coordinate transformation, and the resulting product field is passed to the projection function, which maps the product field to the output of the signal.

[0054] In addition, the neural network model is provided with the following parameters: the number of neurons in the hidden layer of the network hidden_dim, the number of fully connected layers in the network num_layers, and different compression levels correspond to different parameter settings.

[0055] Among them, when the compression level is high compression, num_layers is set to 2, indicating that each MLP contains two hidden layers, hidden_dim is set to 64, that is, each hidden layer of the MLP contains 64 hidden units; when the compression level is medium compression, num_layers is still set to 2, hidden_dim is set to 256, that is, the number of hidden units in each hidden layer is increased to 256; when the compression level is low compression, num_layers is still set to 2, hidden_dim is set to 512, that is, the number of hidden units in each hidden layer reaches 512.

[0056] S202: training a neural network model based on the sample three-dimensional volume data, and outputting predicted three-dimensional volume data;

[0057] The predicted three-dimensional volume data includes the coordinate values ​​of the sampling points and the predicted voxel values, and the predicted voxel values ​​are the values ​​of the voxel colors of the corresponding data obtained through prediction after visualization.

[0058] In this embodiment, S202 specifically includes: training the neural network model by inputting sample three-dimensional volume data into the neural network model, and the neural network model outputs predicted three-dimensional volume data of sampling points corresponding to the sample three-dimensional volume data according to coordinate values ​​in the sample three-dimensional volume data.

[0059] Preferably, the neural network model uses Adamax optimizer during training and uses 2e -2 As the learning rate, the formula for the loss function is:

[0060]

[0061] Where, L MAE That is, the loss function is the loss calculated using the mean absolute error; f(x i ) is the predicted voxel value; y i is the original voxel value.

[0062] In this embodiment, the neural network model is iterated 10,000 times during the training process. Each iteration calculates the value of the loss function based on the current predicted voxel value and the sample voxel value. As the iteration proceeds, the Adamax optimizer adjusts the parameters of the neural network model according to the gradient of the loss function, so that the value of the loss function gradually decreases. MAEIt keeps decreasing until convergence, and the training of the neural network model is completed.

[0063] S203: Reconstruct the predicted voxel values ​​and coordinate values ​​in the predicted three-dimensional volume data to obtain reconstructed three-dimensional volume data.

[0064] Preferably, the method further comprises: measuring the quality difference between the reconstructed three-dimensional volume data and the original three-dimensional volume data based on a peak signal-to-noise ratio.

[0065] Among them, Peak Signal to Noise Ratio (PSNR) is a commonly used evaluation index for measuring the quality of digital images or videos.

[0066] Furthermore, the formula for the peak signal-to-noise ratio is:

[0067]

[0068] Where MAX is the maximum value of the sample 3D volume voxel value; MSE is the mean square error, that is, the average value of the square of the difference between the reconstructed 3D volume data and the sample 3D volume data.

[0069] In this embodiment, the reconstructed three-dimensional volume data retains key information in the sample three-dimensional volume data and is very similar to the sample three-dimensional volume data in terms of visual effect.

[0070] Embodiment 2:

[0071] Based on a 3D volume data compression and reconstruction method for scientific visualization provided in Example 1, the Research Vessel Tangaroa dataset is used as a compressed 3D volume dataset, and the compressed 3D volume dataset is compressed and reconstructed.

[0072] The Research Vessel Tangaroa data set includes incompressible three-dimensional fluid simulation data of the CAD model of the research vessel Tangaroa. This embodiment selects the 150th time step data in the data set for compression and reconstruction.

[0073] FIG2 is a visualization comparison of the reconstructed 3D volume data and the original 3D volume data of the Research Vessel Tangaroa dataset of 300×180×120 with high compression level; specifically, Figure 2a Some of them are visualization diagrams of the original 3D volume data. Figure 2b A visualization diagram of reconstructed 3D volume data.

[0074] FIG3 is a visualization comparison of the reconstructed 3D volume data and the original 3D volume data of the Research Vessel Tangaroa dataset of 300×180×120 at medium compression level; specifically, Figure 3a This is a visualization diagram of the original 3D volume data. Figure 3b A visualization diagram of reconstructed 3D volume data.

[0075] FIG4 is a visualization comparison diagram of the reconstructed 3D volume data and the original 3D volume data of the Research Vessel Tangaroa dataset of 300×180×120 with low compression level; specifically, Figure 4a This is a visualization diagram of the original 3D volume data. Figure 4b A visualization diagram of reconstructed 3D volume data.

[0076] Based on the above visualization diagram, when the loss function of the neural network decreases in value until convergence, the corresponding PSNR value gradually increases to the maximum value, indicating that the higher the similarity between the reconstructed 3D volume data and the original 3D volume data, the better the compression effect. Among them, the visualization diagrams in this embodiment are all obtained through ParaView software. The better the compression effect, the better the visualization effect of the 3D volume data, so that the reconstructed 3D volume data is almost consistent with the original 3D volume data in visual effect.

[0077] Further, based on the method provided in Example 1, the quantitative analysis results in this example are as follows

[0078] As shown in Table 1:

[0079]

[0080] Based on the quantitative analysis results shown in Table 1, if high compression is selected to compress the 3D volume data set, the compression degree is high, but the compression effect is poor; if medium compression is selected to compress the 3D volume data set, the compression effect is good, and the quality of the reconstructed 3D volume data is significantly improved; if low compression is selected to compress the 3D volume data set, the compression effect is better, that is, the key information of the original 3D volume data can be better preserved, so that the quality of the reconstructed 3D volume data is higher, and the reconstructed 3D volume data is closer to the original 3D volume data, or even almost the same. It can be seen that this result verifies the flexibility and effectiveness of the method provided by this implementation under different parameter settings.

[0081] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A method for compressing and reconstructing three-dimensional volume data for scientific visualization, characterized in that: include: S1: acquiring original voxel data of all leaf nodes based on the compressed three-dimensional volume data set, extracting some of the leaf nodes as sampling points, and using the original three-dimensional volume data of the sampling points as sample three-dimensional volume data; S2: constructing a neural network model, training the neural network model based on the sample three-dimensional volume data, outputting predicted three-dimensional volume data, reconstructing the predicted three-dimensional volume data, and obtaining reconstructed three-dimensional volume data.

2. A method for compressing and reconstructing three-dimensional volume data for scientific visualization according to claim 1, characterized in that: S1 specifically includes the following steps: S101: compressing a three-dimensional volume data set based on a compression level to obtain the compressed three-dimensional volume data set, where the compression level includes: high compression, medium compression, and low compression; S102: Read the compressed three-dimensional volume data set to obtain the original three-dimensional volume data of all the leaf nodes; S103: extracting part of the leaf nodes as the sampling points based on the randomly generated index, and using the original three-dimensional volume data of the sampling points as the sample voxel data.

3. The method for compressing and reconstructing three-dimensional volume data for scientific visualization according to claim 1, characterized in that: S2 specifically includes the following steps: S201: constructing the neural network model based on the coefficient field, the basis field and the projection function; S202: training the neural network model based on the sample three-dimensional volume data, and outputting predicted three-dimensional volume data; S203: Reconstruct the predicted voxel values ​​and coordinate values ​​in the predicted three-dimensional volume data to obtain reconstructed three-dimensional volume data.

4. The method for compressing and reconstructing three-dimensional volume data for scientific visualization according to claim 1, characterized in that: The neural network model uses Adamax optimizer during training and uses 2e -2 As the learning rate, the formula for the loss function is: Where, L MAE That is, the loss function is the loss calculated using the mean absolute error; f(x i ) is the predicted voxel value; yi is the original voxel value.

5. The method for compressing and reconstructing three-dimensional volume data for scientific visualization according to claim 1, characterized in that: Also includes: The quality difference between the reconstructed three-dimensional volume data and the original three-dimensional volume data is measured based on a peak signal-to-noise ratio.

6. The method for compressing and reconstructing three-dimensional volume data for scientific visualization according to claim 5, characterized in that: The formula for the peak signal-to-noise ratio is: Wherein, MAX is the maximum value of the voxel value; MSE is the mean square error, that is, the average value of the square of the difference between the reconstructed three-dimensional volume data and the sample three-dimensional volume data.