A neural network-based bidirectional texture function image compression and editing method
By decoupling parameterization and neural networks, the light intensity and albedo of the bidirectional texture function are separated, solving the problem of inconvenient editing in traditional compression methods and achieving efficient semantic editing and memory optimization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING UNIV
- Filing Date
- 2026-03-23
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies struggle to independently edit semantics such as albedo and glossiness of bidirectional texture functions during compression, and traditional compression methods result in memory pressure and inconvenient editing.
The bidirectional texture function data is converted into mid-range vectors and difference angle vector indices through a parameterization stage. The light intensity and albedo distribution tensors are decoupled using a neural network. A loss function is designed for iterative optimization, and the albedo distribution map is modified for editing.
It enables controlled editing of specific semantics while maintaining high-quality compression, particularly the modification of albedo, reducing memory usage and increasing editing flexibility.
Smart Images
Figure CN122391771A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an image compression and editing method, and more particularly to an image compression and editing method based on a bidirectional texture function of a neural network. Background Technology
[0002] Traditional graphics pipelines rely on the simplified assumption of "diffuse reflection + specular reflection," which often results in "imperfections" on materials like fabrics that possess subsurface structures and spatially varying reflective properties. This is because the material's appearance is not solely determined by a single BRDF (Brake-Based Surface Function), but rather by the combined effects of microsurface normal distribution, subsurface scattering, shadow occlusion, and multiple bounces. Its reflection field varies in high dimension, nonlinearly, and asymmetricly with respect to the light source and viewpoint. Accurate calculation of nonlocal optical effects such as subsurface scattering, shadow occlusion, and multiple bounces typically relies on ray tracing or path integration methods, introducing significant time overhead that is insufficient for real-time rendering requirements. To avoid the complex light transmission process in online solutions, researchers introduced a bidirectional texture function: dense sampling of the material under test in the real world, pre-recording the radiance that already covers the above nonlocal effects as a six-dimensional lookup table: using the incident direction wi, the outgoing direction wo, and two-dimensional position information as indexes, the pixel information at a specific location under the corresponding light-view relationship is searched; during the rendering stage, only the stored image needs to be indexed and interpolated based on the instantaneous incident and outgoing directions to quickly reproduce the physically correct visual appearance at the pixel level.
[0003] However, since bidirectional texture functions are essentially large numbers of texture images, they often cause memory pressure when used. Currently, bidirectional texture functions typically employ traditional mathematical compression methods such as PCA and neural network compression methods; the latter, compared to the former, introduces non-linear operations during compression, achieving better fidelity at the same compression ratio.
[0004] When using bidirectional texture functions compressed by neural networks, the semantics of their parameters are often difficult to understand due to the introduction of nonlinear operations; therefore, editing them becomes challenging. Current methods mostly only achieve relatively controllable appearance transfer of bidirectional texture functions, such as making the material outline of the bidirectional texture function approximate the outline of the guide map, or weighting several bidirectional texture functions to generate new materials. The information that needs modification is coupled together, limiting the editing effect. It is impossible to edit more common and specific semantics, such as albedo and gloss, individually. Summary of the Invention
[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for bidirectional texture function image compression and editing based on neural networks, which addresses the shortcomings of the existing technology.
[0006] To address the aforementioned technical problems, this invention discloses a bidirectional texture function-based image compression and editing method, comprising the following steps:
[0007] Step 1, parameterization stage: The original bidirectional texture function image data is parameterized to obtain the mid-range vector. index Difference angle vector index Resolution index of texture images ;
[0008] Step 2, the decoupling stage, uses a neural network to decouple the indices obtained in Step 1, resulting in the light intensity distribution tensor. and albedo distribution tensor The light intensity distribution map and albedo distribution map were obtained by decoding respectively.
[0009] Step 3, compression stage: The tensors obtained in step 2 are further transformed through a neural network to obtain the compressed bidirectional texture function;
[0010] Step 4: Design the loss function and iteratively optimize the calculation process of Steps 2 and 3;
[0011] Step 5: Modify and replace the albedo distribution map in Step 2 and retrain to complete the editing.
[0012] Furthermore, the parameterization of the original bidirectional texture function image data described in step 1 specifically includes:
[0013] Step 1-1: Change the dependent variable in the original bidirectional texture function data from the incident direction. and launch direction Convert to mid-range vector Sum and difference angle vectors and according to the mid-range vector Sum and difference angle vectors Replacing the ascending order with index numbers yields the following result: index and index ;
[0014] Steps 1-2: Obtain the resolution of the texture image from the original bidirectional texture function data. index .
[0015] Furthermore, step 1-1 specifically includes:
[0016] Let the incident direction be... and launch direction for:
[0017]
[0018]
[0019] in, Let be the polar angle of the incident direction in spherical coordinates. Let be the azimuth angle of the incident direction in spherical coordinates. Let be the polar angle of the launch direction in spherical coordinates. The azimuth angle of the launch direction in spherical coordinates;
[0020] Then the mid-range vector The calculation method is as follows:
[0021]
[0022] Let the mid-range vector be When rotated to the North Pole axis, the rotation matrix is: Then the difference angle vector for:
[0023]
[0024] After variable substitution, the bidirectional reflection function can be written as a function of the midrange vector. Sum and difference angle vectors The expression formed :
[0025]
[0026] in, Let be the polar angle of the mid-range vector in spherical coordinates. Let be the azimuth angle of the mid-range vector in spherical coordinates. Let the difference angle vector be the polar angle in spherical coordinates. The azimuth angle of the difference angle vector in spherical coordinates;
[0027] The transformed midrange vector Sum and difference angle vectors The index numbers are then changed in ascending order to obtain the one-dimensional index information corresponding to each texture in the bidirectional texture function. index and index ;in,
[0028] To make the mid-range vector The value, according to First priority As the second priority, the data is converted into a one-dimensional coordinate representation in ascending order;
[0029] To make the difference angle vector The value, according to First priority As the second priority, it is converted into a one-dimensional coordinate representation in ascending order.
[0030] Furthermore, in steps 1-2, the resolution of the texture image obtained from the original bidirectional texture function data is... index Specifically, it includes:
[0031] The pixel information of the texture image resolution in the original bidirectional texture function data is converted into a one-dimensional coordinate representation with row number as the first priority and column number as the second priority.
[0032] Furthermore, the decoupling stage described in step 2 specifically includes:
[0033] Step 2-1, utilizing the embedding layer of the neural network and embedding layer , respectively index and index Transformed into an 8-dimensional feature tensor;
[0034] Using the embedding layer of a neural network ,Will index Transformed into a 16-dimensional feature tensor;
[0035] Step 2-2, using the linear layer Linear_xy1 with LEAKY_RELU as the activation function, ... index The 16-dimensional features are transformed into a 24-dimensional feature tensor;
[0036] Steps 2-3 involve splitting the 24-dimensional feature tensor into three equal parts, where 2 / 3 of the parts are... index and index By concatenating the characteristic tensors of the transformation, the light intensity distribution tensor is obtained. ;
[0037] The remaining 1 / 3 constitutes the albedo distribution tensor. ;
[0038] Steps 2-4, for the albedo distribution tensor and light intensity distribution tensor Using an independent MLP decoder and decoder Decoding yields an albedo distribution map and a light intensity distribution map represented in RGB.
[0039] Furthermore, the compression stage described in step 3 specifically includes:
[0040] Step 3-1, Light Intensity Distribution Tensor After passing through a linear layer with LEAKY_RELU as the activation function;
[0041] Step 3-2, Albedo Distribution Tensor After two linear layers with LEAKY_RELU as the activation function;
[0042] Step 3-3: Combine the outputs of Step 3-1 and Step 3-2 to obtain a 40-dimensional material feature tensor;
[0043] Steps 3-4: Apply an independent MLP decoder to the material feature tensor. Decoding yields a bidirectional texture function result image represented in RGB, which is the compressed bidirectional texture function.
[0044] Furthermore, the loss function described in step 4 is expressed as follows:
[0045]
[0046] in, For the total loss function, and This represents the actual albedo and light intensity distribution corresponding to the bidirectional texture function. and The albedo and light intensity distribution calculated in step 2 are used. This corresponds to the pixel value in the original bidirectional texture function. The pixel values in the compressed bidirectional texture function calculated in step 3, For MAE loss, This is the MSE loss.
[0047] Furthermore, step 5, which involves modifying and replacing the albedo distribution map in step 2 and retraining, specifically includes:
[0048] Step 5-1: Using the albedo distribution map from Step 2 as the base map, modify and replace local details of the base map using digital image processing or drawing tools to obtain the guide map. ;
[0049] Step 5-2, guide the image As an MLP decoder The result is that the remaining learning gradients are cleared to zero, retaining only those that are consistent with the previous learning gradients. index The gradients of the relevant parts are used for reverse training, and after a preset number of rounds, the modified and replaced bidirectional texture function is obtained.
[0050] Beneficial effects:
[0051] 1. By introducing a parameterized and directional feature training process, this invention can achieve the technical effect of editing specific semantics separately while ensuring relatively high-quality compression.
[0052] 2. This invention utilizes a neural network method to decouple the optical semantics of materials during the compression of bidirectional texture functions. While compressing, it retains parameters that are easier to modify and understand, making modifications to them more controllable. Attached Figure Description
[0053] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.
[0054] Figure 1 This is a schematic diagram of the neural network framework in an embodiment of the present invention.
[0055] Figure 2 This is a schematic diagram showing the result of Rusin parameterization in an embodiment of the present invention.
[0056] Figure 3 This is a schematic diagram of the editing process in an embodiment of the present invention.
[0057] Figure 4 This is a schematic diagram showing the comparison between the modified BTF texture and the original texture from different angles in an embodiment of the present invention.
[0058] Figure 5 These are renderings of various modified BTF materials in embodiments of the present invention. Detailed Implementation
[0059] This invention provides a compression method for bidirectional texture functions. While compressing, it performs directional feature decoupling of specific semantic parameters to achieve more controllable editing of bidirectional texture functions, mainly focusing on the modification of albedo.
[0060] This invention proposes a method for compressing and editing bidirectional texture function images based on neural networks, which consists of a data parameterization process, a compression process, and an editing process. Details are as follows:
[0061] Step 1, parameterization stage.
[0062] Using the Rusinkiewicz method, the dependent variable in the original bidirectional texture function data is changed from the incident direction. , emission direction Convert to mid-range vector Difference angle vector The converted h and d are then converted into index numbers in ascending order.
[0063] The specific meanings of the parameters h and d are as follows: h is... (incident direction) and The mid-range vector (of the viewpoint direction); d is the incident direction when h is rotated to the north pole axis. The position after rotation.
[0064] At the same time, the resolution index information of the texture image is obtained. Specifically, when the resolution of the texture image is w*h, for the pixel at position (i,j) (i and j start from 0, j increases from left to right, and j increases from top to bottom), the converted index is i*w+j.
[0065] like Figure 2 The image shown is a schematic diagram of the light intensity distribution before and after parameterization at various angles. It can be seen that after parameterization, the light intensity distribution exhibits the characteristics of concentrated signal energy, gradual change, and dominant symmetry. Such a signal is easier for neural networks to fit.
[0066] Step 2, Decoupling Phase.
[0067] Using the embedding layers Embh and Embd of the neural network, the indices of h and d are each transformed into 8-dimensional feature tensors. The resolution index xy of the texture image is then transformed into a 16-dimensional feature tensor using the Embxy layer. Subsequently, a linear layer Linear_xy1 with LEAKY_RELU activation function is used to transform the 16-dimensional features of the resolution index into a 24-dimensional feature tensor. This 24-dimensional feature tensor is then split into three equal parts. Two-thirds are combined with the feature tensors transformed from h and d indices to obtain the light intensity distribution tensor Ti. The remaining one-third forms the albedo distribution tensor Ta independently. Independent MLP decoders are then applied to the albedo distribution tensor and the light intensity distribution tensor. , Decode the data to obtain the albedo distribution map and light intensity distribution map represented in RGB.
[0068] Step 3, compression stage.
[0069] After applying one and two linear layers of LEAKY_RELU as activation functions respectively to the light intensity distribution tensor and albedo distribution tensor without changing their dimensions, they are concatenated to obtain a 40-dimensional feature tensor representing the material.
[0070] A separate MLP decoder is applied to the material feature tensor. Decoding is performed to obtain the bidirectional texture function result image represented in RGB. Since the memory occupied by the light intensity distribution tensor, albedo distribution tensor, and subsequent framework at this stage is 1:1000 compared to the original bidirectional texture function data, compression is effectively achieved.
[0071] Step 4, iterative optimization.
[0072] The loss function to be optimized is trained by using the sum of the albedo, bidirectional texture function map and the L1 loss of the actual label generated each time the dependent variable changes, and the L2 loss of the light intensity and the actual label.
[0073] Step 5, the editing process.
[0074] By modifying and replacing the albedo map obtained from the albedo feature decoding as a new label, and then retraining by retaining only the gradients of the neural layers related to the albedo feature encoding process, the albedo of the bidirectional texture function can be modified.
[0075] Example 1:
[0076] like Figure 1 As shown, the embodiments of this application can be more specifically divided into parameterization process, decoupling process, compression process, and modification process. The specific technical solutions are as follows:
[0077] Step 1, perform the following variable substitutions:
[0078] Let the incident direction be The direction of launch is ,in, Let be the polar angle of the incident direction in spherical coordinates. Let be the azimuth angle of the incident direction in spherical coordinates. Let be the polar angle of the launch direction in spherical coordinates. The azimuth angle of the launch direction in spherical coordinates;
[0079] Then the mid-range vector for:
[0080]
[0081] Assuming When rotated to the North Pole axis, the rotation matrix is: Then the difference angle vector for:
[0082]
[0083] After variable substitution, the bidirectional reflection function can be written as a function of the midrange vector. Sum and difference angle vectors The expression formed:
[0084]
[0085] in, Let be the polar angle of the mid-range vector in spherical coordinates. Let be the azimuth angle of the mid-range vector in spherical coordinates. Let the difference angle vector be the polar angle in spherical coordinates. The azimuth angle of the difference angle vector in spherical coordinates.
[0086] After processing the original bidirectional texture function, the transformed midrange vector Difference angle vector By renaming them from smallest to largest index number, we can obtain the one-dimensional index information corresponding to each texture in the bidirectional texture function: , , .in,
[0087] To represent the h value from two-dimensional spherical coordinates ( )according to First priority, The second priority is to flatten and continuous the order from smallest to largest into a one-dimensional coordinate representation.
[0088] To be Represented from two-dimensional spherical coordinates ( )according to First priority, The second priority is to flatten and continuous the order from smallest to largest into a one-dimensional coordinate representation.
[0089] To flatten and continuousen the pixel information of the resolution into a one-dimensional coordinate representation, with row number as the first priority and column number as the second priority.
[0090] Step 2: Convert this one-dimensional information into a feature tensor through the embedding layer.
[0091] Since the light intensity distribution of a bidirectional texture function is not only related to angle changes but also varies at different locations, a fully connected layer is used to... Index, after The embedding layer yields a feature tensor, which is then converted to 24 dimensions and divided into three equal blocks, with two blocks respectively intersecting with... and go through and The obtained feature tensors are then concatenated to obtain a tensor representing the light intensity distribution. The remaining block represents the tensor of the albedo distribution. ; Characteristic tensor of light intensity distribution A single fully connected layer decoder Decoding yields a three-channel output representing the light intensity distribution. ,Right now Figure 1 The light intensity distribution map in the image; similarly, for the feature tensor... Also through a separate decoder Decoding yields the albedo value ,Right now Figure 1 Albedo distribution map.
[0092] Step 3, for , After processing with two layers and one layer of fully connected layers that do not change the dimensions, the results are concatenated to obtain the feature tensor representing the material. Then, a fully connected layer decoder is used. Decoding this feature tensor yields three-channel color values, which are the final bidirectional texture function result image represented in RGB.
[0093] Step 4, for each batch, the total loss function at this point. for:
[0094]
[0095] in, and This represents the actual albedo and light intensity distribution corresponding to the bidirectional texture function. and The corresponding value obtained in step 2. This corresponds to the pixel value in the original bidirectional texture function. The pixel values in the compressed bidirectional texture function calculated in step 3 are... For MAE loss, This is the MSE loss.
[0096] Based on the total loss function The calculation process in steps 2 and 3 is optimized.
[0097] Step 5: Regarding the material editing process, taking albedo modification as an example, after compression, utilize the albedo... Obtain the albedo distribution map as the base map. Use digital image processing algorithms or drawing tools to modify the local details of the base map to obtain a guide map with the desired effect. .
[0098] Finally, the guide map As The result is that the remaining learning gradients are cleared to zero, retaining only those that are related to the previous learning gradient. The gradient of the relevant part, using To perform reverse training on this part, the gradient preservation part and the reverse training part are as follows: Figure 3 As shown, after a certain number of rounds, the modified bidirectional texture function can be obtained. After modification, if you view the BTF texture at various angles, you can find... Figure 4 The results shown (this is the texture map obtained by neural BTF after adding a fuzzy albedo effect to the UBO2014 fabric04 dataset) show that only the albedo distribution has been modified, while the light intensity distribution remains consistent. When applying a series of modifications to fabric09 and stone04 in the UBO2014 dataset and rendering, the following results can be achieved: Figure 5 The results are shown. It is important to emphasize that currently, there is no industry practice to separately modify BTF-specific optical information, such as albedo distribution and light intensity distribution, at the neural BTF level, so it is not easy to make cross-sectional comparisons.
[0099] In its specific implementation, this application provides a computer storage medium and a corresponding data processing unit. The computer storage medium is capable of storing a computer program, which, when executed by the data processing unit, can run the invention's content regarding a bidirectional texture function image compression and editing method based on a neural network, as well as some or all of the steps in various embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0100] Those skilled in the art will clearly understand that the technical solutions in the embodiments of the present invention can be implemented using computer programs and their corresponding general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of computer programs, i.e., software products. These computer program software products can be stored in a storage medium and include several instructions to cause a device containing a data processing unit (which may be a personal computer, server, microcontroller, MCU, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.
[0101] This invention provides an idea and method for bidirectional texture function image compression and editing based on neural networks. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.
Claims
1. A method for image compression and editing based on a bidirectional texture function using neural networks, characterized in that, Includes the following steps: Step 1, parameterization stage: The original bidirectional texture function image data is parameterized to obtain the mid-range vector. index Difference angle vector index Resolution index of texture images ; Step 2, the decoupling stage, uses a neural network to decouple the indices obtained in Step 1, resulting in the light intensity distribution tensor. and albedo distribution tensor The light intensity distribution map and albedo distribution map were obtained by decoding respectively. Step 3, compression stage: The tensors obtained in step 2 are further transformed through a neural network to obtain the compressed bidirectional texture function; Step 4: Design the loss function and iteratively optimize the calculation process of Steps 2 and 3; Step 5: Modify and replace the albedo distribution map in Step 2 and retrain to complete the editing.
2. The method for image compression and editing based on a neural network bidirectional texture function according to claim 1, characterized in that, Step 1, which involves parameterizing the original bidirectional texture function image data, specifically includes: Step 1-1: Obtain the dependent variable from the original bidirectional texture function data. index and index ; Steps 1-2: Obtain the resolution of the texture image from the original bidirectional texture function data. index .
3. The method for bidirectional texture function image compression and editing based on neural networks according to claim 2, characterized in that, The step 1-1 described above involves obtaining the dependent variable from the original bidirectional texture function data. index and index That is, the dependent variable in the original bidirectional texture function data is changed from the incident direction. and launch direction Convert to mid-range vector Sum and difference angle vectors and according to the mid-range vector Sum and difference angle vectors Replacing the ascending order with index numbers yields the following result: index and index ; Specifically, it includes: Let the incident direction be... and launch direction for: ； ； in, Let be the polar angle of the incident direction in spherical coordinates. Let be the azimuth angle of the incident direction in spherical coordinates. Let be the polar angle of the launch direction in spherical coordinates. The azimuth angle of the launch direction in spherical coordinates; Then the mid-range vector The calculation method is as follows: ； Let the mid-range vector be When rotated to the North Pole axis, the rotation matrix is: Then the difference angle vector for: ； After variable substitution, the bidirectional reflection function can be written as a function of the midrange vector. Sum and difference angle vectors The expression formed : ； in, Let be the polar angle of the mid-range vector in spherical coordinates. Let be the azimuth angle of the mid-range vector in spherical coordinates. Let the difference angle vector be the polar angle in spherical coordinates. The azimuth angle of the difference angle vector in spherical coordinates; The transformed midrange vector Sum and difference angle vectors The index numbers are then converted to ascending order to obtain the one-dimensional index information corresponding to each texture in the bidirectional texture function. index and index ;in, To make the mid-range vector The value, according to First priority As the second priority, the data is converted into a one-dimensional coordinate representation in ascending order; To make the difference angle vector The value, according to First priority As the second priority, it is converted into a one-dimensional coordinate representation in ascending order.
4. The method for image compression and editing based on a neural network bidirectional texture function according to claim 3, characterized in that, Steps 1-2 describe obtaining the resolution of the texture image based on the original bidirectional texture function data. index Specifically, it includes: The pixel information of the texture image resolution in the original bidirectional texture function data is converted into a one-dimensional coordinate representation with row number as the first priority and column number as the second priority.
5. The method for image compression and editing based on a neural network bidirectional texture function according to claim 4, characterized in that, The decoupling stage described in step 2 specifically includes: Step 2-1, utilizing the embedding layer of the neural network and embedding layer , respectively index and index Transformed into an 8-dimensional feature tensor; Using the embedding layer of a neural network ,Will index Transformed into a 16-dimensional feature tensor; Step 2-2, using the linear layer Linear_xy1, ... index The 16-dimensional features are transformed into a 24-dimensional feature tensor; Steps 2-3 involve splitting the 24-dimensional feature tensor into three equal parts, where 2 / 3 of the parts are... index and index By concatenating the characteristic tensors of the transformation, the light intensity distribution tensor is obtained. ; The remaining 1 / 3 constitutes the albedo distribution tensor. ; Steps 2-4, for the albedo distribution tensor and light intensity distribution tensor Using an independent MLP decoder and decoder Decoding yields an albedo distribution map and a light intensity distribution map represented in RGB.
6. The method for image compression and editing based on a neural network bidirectional texture function according to claim 5, characterized in that, The compression stage described in step 3 specifically includes: Step 3-1, Light Intensity Distribution Tensor Passing through a linear layer; Step 3-2, Albedo Distribution Tensor After two linear layers; Step 3-3: Combine the outputs of Step 3-1 and Step 3-2 to obtain a 40-dimensional material feature tensor; Steps 3-4: Apply an independent MLP decoder to the material feature tensor. Decoding yields a bidirectional texture function result image represented in RGB, which is the compressed bidirectional texture function.
7. The method for image compression and editing based on a neural network bidirectional texture function according to claim 6, characterized in that, The loss function mentioned in step 4 is expressed as follows: ； in, For the total loss function, and This represents the actual albedo and light intensity distribution corresponding to the bidirectional texture function. and The albedo and light intensity distribution calculated in step 2 are used. This corresponds to the pixel value in the original bidirectional texture function. The pixel values in the compressed bidirectional texture function calculated in step 3, For MAE loss, This is the MSE loss.
8. The method for image compression and editing based on a neural network bidirectional texture function according to claim 7, characterized in that, Step 5, which involves modifying and replacing the albedo distribution map from step 2 and then retraining, specifically includes: Step 5-1: Using the albedo distribution map from Step 2 as the base map, modify and replace local details of the base map using digital image processing or drawing tools to obtain the guide map. ; Step 5-2, guide the image As an MLP decoder The result is that the remaining learning gradients are cleared to zero, retaining only those that are consistent with the previous learning gradients. index The gradients of the relevant parts are used for reverse training, and after a preset number of rounds, the modified and replaced bidirectional texture function is obtained.
9. The method for image compression and editing based on a neural network bidirectional texture function according to claim 8, characterized in that, The linear layer Linear_xy1 described in step 2-2 uses LEAKY_RELU as the activation function.
10. The method for image compression and editing based on a neural network bidirectional texture function according to claim 9, characterized in that, The linear layers described in steps 3-1 and 3-2 use LEAKY_RELU as the activation function.