CN114926593B - SVBRDF material modeling method and system based on Shan Zhanggao light images - Google Patents

Abstract

The present disclosure provides a SVBRDF material modeling method and system based on a single highlight image, the scheme belongs to the field of three-dimensional rendering material technology, and the scheme includes: obtaining a highlight image of the surface of an object material; removing highlights from the highlight image by dense feature fusion connection and highlight multi-level recognition to obtain a highlight-free image; inputting the highlight image and the highlight-free image into a pre-trained generator network to obtain a spatially varying bidirectional reflectance distribution function of the object surface, and then obtaining a corresponding material map; wherein the generator network includes a shared encoder and a plurality of decoders respectively connected to the shared encoder, and the decoders respectively correspond to the processing of a diffuse map, a normal map, a roughness map and a reflection map.

Description

SVBRDF material modeling method and system based on single highlight image

Technical Field

The present disclosure belongs to the technical field of three-dimensional rendering materials, and in particular, relates to a SVBRDF material modeling method and system based on a single highlight image.

Background technique

The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

In the field of computer graphics and vision, estimating the reflective properties of materials from images has been a widely studied topic. In the real world, the appearance of materials depends on the viewing and lighting directions, which makes appearance estimation a challenging task. The surface reflectance property is usually represented by a bidirectional reflectance distribution function (BRDF). Since most surfaces are not completely homogeneous, the BRDF is usually given as a function of the surface position, which is called a spatially varying bidirectional reflectance distribution function (SVBRDF), including diffuse albedo, specular albedo, surface normal, and roughness, corresponding to four material maps: diffuse map, reflection map, normal map, and roughness map.

Different reflectivities can produce the same observed image, so it is very difficult to reconstruct a high-quality spatially varying bidirectional reflectance distribution function from input images taken under different lighting and observation directions. Usually, the apparent material modeling of an object requires the use of professional instruments to measure the reflection characteristics of the light on the surface of a specific object material in the scene, or the use of professional tools to mark the material map through a large amount of manual intervention. The inventors have found that recently, a deep learning-based method has been used in material modeling to reconstruct various material maps from photographed material images, greatly simplifying the process of material modeling. However, due to the limited dynamic range of many consumer-grade cameras, some image areas will be contaminated by specular reflection highlights, which will produce obvious spot artifacts in the generated feature map, resulting in the generated material map effect failing to meet actual needs.

Summary of the invention

In order to solve the above problems, the present disclosure provides a SVBRDF material modeling method and system based on a single highlight image. The scheme automatically generates a material map that is close to the appearance of the real material based on a single image with highlights taken in the real world.

According to a first aspect of an embodiment of the present disclosure, a SVBRDF material modeling method based on a single highlight image is provided, comprising:

Get the highlight image of the object material surface;

Eliminate the highlights of the highlight image by dense feature fusion connection and highlight multi-level recognition to obtain a highlight-free image;

The highlight image and the non-highlight image are input into a pre-trained generator network to obtain a spatially varying bidirectional reflectance distribution function of the object surface, and then obtain a corresponding material map; wherein the generator network includes a shared encoder and a plurality of decoders respectively connected to the shared encoder, and the decoders respectively correspond to the processing of diffuse map, normal map, roughness map and reflection map.

Furthermore, the highlight removal is performed based on dense feature fusion connection and multi-level highlight recognition, specifically: an encoder-decoder group with dense feature fusion connection is used, a multi-level highlight recognition strategy is adopted, and a highlight removal operation is performed on the captured image to obtain a highlight-free image of the captured image.

Furthermore, the encoder includes a plurality of sequentially connected convolutional layers, instance normalization layers, linear rectification units and maximum pooling layers; the decoder includes a plurality of sequentially connected deconvolutional layers, instance normalization layers and linear rectification units.

Furthermore, the encoder performs a cascade operation on the features extracted by the convolution block and the features extracted by the convolution block in the previous layer based on dense feature fusion connection, then performs an upsampling operation, and finally performs a convolution operation.

Furthermore, the encoder of the generator network includes eight convolutional layers for downsampling, the first seven convolutional layers are connected to an instantiated normalization layer and a linear rectifier unit, and the last convolutional layer is connected to an instantiated normalization layer.

Furthermore, the generator network and the multi-discriminator network are trained collaboratively, specifically:

Obtain a public data set, and construct a training data set by removing duplicates, randomly cropping, and randomly screening the public data set;

Based on a pre-constructed minimization loss function, the generator network and the multi-discriminator network are collaboratively trained through a constructed training data set, wherein the quality of the material map generated by the generator network is judged through the multi-discriminator network and the pre-constructed loss function.

According to a second aspect of an embodiment of the present disclosure, a SVBRDF material modeling system based on a single highlight image is provided, comprising:

A data acquisition unit, which is used to acquire a highlight image of the surface of an object material;

A highlight removal unit, which is used to remove the highlights of the highlight image by means of dense feature fusion connection and highlight multi-level recognition to obtain a highlight-free image;

An object surface material estimation unit is used to input the highlight image and the non-highlight image into a pre-trained generator network to obtain a spatially varying bidirectional reflectance distribution function of the object surface, and then obtain a corresponding material map; wherein the generator network includes a shared encoder and a plurality of decoders respectively connected to the shared encoder, and the decoders respectively correspond to the processing of diffuse maps, normal maps, roughness maps and reflection maps.

According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising a memory, a processor, and a computer program stored and running on the memory, wherein when the processor executes the program, the SVBRDF material modeling method based on a single highlight image is implemented.

According to a fourth aspect of an embodiment of the present disclosure, a non-transitory computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the SVBRDF material modeling method based on a single highlight image is implemented.

Compared with the prior art, the beneficial effects of the present invention are:

(1) The scheme disclosed in the present invention provides an SVBRDF material modeling method based on a single highlight image. The generator proposed is an encoder-decoder network with a shared encoder and four decoders. The four decoders process the diffuse map, normal map, roughness map and reflection map respectively. The generator network is used to generate a material map. The generated map is rendered into a rendered image using a rendering module. A multi-discriminator network is used to discriminate between the generated map and the real map, and between the rendered image and the real image, so as to generate a high-quality material appearance map.

(2) The scheme uses a highlight removal module based on dense feature fusion connection and highlight multi-level recognition to reduce the impact of overexposed areas on material estimation during the SVBRDF estimation process, and can remove highlight artifacts existing in existing SVBRDF estimation technologies.

(3) The scheme proposes a loss function based on mapping loss, rendering loss and adversarial loss, and introduces feature matching loss to stabilize training.

Advantages of additional aspects of the present disclosure will be given in part in the following description, and in part will become apparent from the following description, or will be learned through practice of the present disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

FIG1 is a network structure diagram of the SVBRDF material modeling method based on a single highlight image described in an embodiment of the present disclosure;

FIG2 is a network structure diagram of a highlight removal module described in an embodiment of the present disclosure;

FIG3( a ) is a structural diagram of an encoder of a highlight removal module described in an embodiment of the present disclosure;

FIG3( b ) is a structural diagram of a decoder of a highlight removal module described in an embodiment of the present disclosure;

FIG4 is a schematic diagram of the highlight removal module results described in an embodiment of the present disclosure;

FIG5( a ) is a diagram of a generator network encoder structure according to an embodiment of the present disclosure;

FIG5( b ) is a diagram of a generator network decoder structure in an embodiment of the present disclosure;

FIG6 is a diagram of a discriminator network structure described in an embodiment of the present disclosure;

FIG. 7( a ) and FIG. 7( b ) are comparison diagrams of the results of the SVBRDF material modeling method based on a single highlight image described in an embodiment of the present disclosure.

Detailed ways

The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and/or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and/or combinations thereof.

In the absence of conflict, the embodiments in the present disclosure and the features in the embodiments may be combined with each other.

Embodiment 1:

The purpose of this embodiment is to provide a SVBRDF material modeling method based on a single highlight image.

A SVBRDF material modeling method based on a single highlight image, comprising:

Get the highlight image of the object material surface;

Eliminate the highlights of the highlight image by dense feature fusion connection and highlight multi-level recognition to obtain a highlight-free image;

The highlight image and the non-highlight image are input into a pre-trained generator network to obtain a spatially varying bidirectional reflectance distribution function of the object surface, and then obtain a corresponding material map; wherein the generator network includes a shared encoder and a plurality of decoders respectively connected to the shared encoder, and the decoders respectively correspond to the processing of diffuse maps, normal maps, roughness maps and reflection maps.

Furthermore, the highlight removal is performed based on dense feature fusion connection and multi-level highlight recognition, specifically: an encoder-decoder group with dense feature fusion connection is used, a multi-level highlight recognition strategy is adopted, and a highlight removal operation is performed on the captured image to obtain a highlight-free image of the captured image.

Furthermore, the encoder includes a plurality of sequentially connected convolutional layers, instance normalization layers, linear rectification units and maximum pooling layers; the decoder includes a plurality of sequentially connected deconvolutional layers, instance normalization layers and linear rectification units.

Furthermore, the encoder performs a cascade operation on the features extracted by the convolution block and the features extracted by the convolution block in the previous layer based on dense feature fusion connection, then performs an upsampling operation, and finally performs a convolution operation.

Furthermore, the encoder of the generator network includes eight convolutional layers for downsampling, the first seven convolutional layers are connected to an instantiated normalization layer and a linear rectifier unit, and the last convolutional layer is connected to an instantiated normalization layer.

Furthermore, the generator network and the multi-discriminator network are trained collaboratively, specifically:

Obtain a public data set, and construct a training data set by removing duplicates, randomly cropping, and randomly screening the public data set;

Based on a pre-constructed minimization loss function, the generator network and the multi-discriminator network are collaboratively trained through a constructed training data set, wherein the quality of the material map generated by the generator network is judged through the multi-discriminator network and the pre-constructed loss function.

Furthermore, the multi-discriminator network specifically includes a discriminator for processing diffuse maps, a discriminator for processing normal maps, a discriminator for processing roughness maps, a discriminator for processing reflection maps, and a discriminator for processing rendering results.

Specifically, for ease of understanding, the solution described in this embodiment is described in detail below with reference to the accompanying drawings:

This embodiment provides an SVBRDF material modeling method based on a single highlight image. The scheme can be divided into four main stages: a training data set construction stage, a generator network and a multi-discriminator network collaborative training stage, a highlight elimination stage, and an object surface material estimation stage. The flowchart is shown in FIG1 , which specifically includes the following steps:

Step 1: Remove duplicates, randomly crop, and randomly select large-scale data sets to construct training data sets. By minimizing the loss function, use the training data set to co-train the generator network and the multi-discriminator network, use the multi-discriminator to discriminate the material map generated by the generator, and calculate the loss function according to the formula. The loss function includes map loss, rendering loss, and adversarial loss, where the adversarial loss includes feature matching loss and standard adversarial loss.

S101: Collect public datasets, using the large-scale dataset proposed by Deschaintre et al. The public datasets were deduplicated, and only one set of training data was retained for images generated from the same SVBRDF but with slightly different viewing or lighting directions, and finally 195,284 instances were collected.

S102: Crop the images in the dataset, and randomly crop all the images in the training dataset to 256x256.

S103: Randomly select 18K instances in the data set as the training data set, and the remaining instances as the test data set.

S104: Calculate adversarial loss.

The feature matching loss is calculated using formula (1) The feature matching loss stabilizes training by minimizing the distance between the estimated and true values of high-dimensional features.

Where G represents the generator, I represents the input image, and M represents the material map generated by the generator. _Di refers to any discriminator in _Dmaps = ( _Dd , _Dn , _Dr , _Ds ), and each discriminator has the same structure, where _Dd , _Dn , _Dr , and _Ds are used to process diffuse maps, normal maps, roughness maps, and reflection maps, respectively.

The standard adversarial loss is calculated using formula (2) By minimizing the standard adversarial loss, the generator tries to generate maps that are indistinguishable from reference maps, while the discriminator tries to effectively distinguish the generated maps from the reference maps.

Among them, D _render represents the rendering discriminator;

Using formula (3), the adversarial loss is calculated by combining feature matching loss and standard adversarial loss:

S105: Calculate the mapping loss using formula (4) The map loss reflects the pixel-by-pixel difference between the generated map and the reference map.

in, Represents the reference material map corresponding to the input image I.

S106: Draw a rendered image using the diffuse map, normal map, roughness map and reflection map generated by the generator, and calculate the rendering loss using formula (5) The rendering loss reflects the pixel-by-pixel difference between the real image and the rendered image.

Where R(·) represents the rendering module. The rendering module uses the Cook-Torrance BRDF model and the GGX microsurface normal distribution function.

S107: Combine adversarial loss, mapping loss, and rendering loss, and calculate the total loss function by weighted summation:

Among them, λ _map , λ _render , and λ _adv are weights used to balance each loss item. In this embodiment, λ _map =10, λ _render =5, and λ _adv =1.

S108: Using the training data set, the generator network and the multi-discriminator network are collaboratively trained according to formula (7).

Where N is the number of samples in the training set, θ is the generator parameter set, and the network structure is shown in Figure (1).

In this embodiment, the PyTorch framework and the Adam optimizer are used, and the initial learning rate is 0.00002. The training is performed 1,000,000 times. The graphics processor is NVIDIA TITAN RTX, and the video memory is 12 GB.

Step 2: Use the captured image of the surface of the object material and input it into the highlight removal module for multi-level highlight recognition and highlight removal to obtain a highlight-free image. Input the captured image and the highlight-free image into the trained generator network obtained in step 1 to estimate the surface material of the object, and estimate the diffuse map, normal map, roughness map and reflection map of the material.

S201: Acquire a captured image of the surface of an object material.

S202: Based on a highlight removal module, perform multi-level highlight recognition and removal processing on the captured image to obtain a highlight-free image.

The network structure of the highlight removal module is shown in Figure 2. An encoder-decoder group with dense feature fusion connections is used to adopt a multi-level highlight recognition strategy to perform highlight removal operations on the captured image to obtain a highlight-free image of the captured image. The highlight recognition and highlight removal results are shown in Figure 4.

Using the captured image I as input, the highlight feature F is extracted using the highlight removal module encoder and the highlight removal module decoder.

Based on the highlight feature F, the highlight mask M is predicted, where M represents the visible highlight position.

The highlight density S is predicted based on the highlight feature F and the highlight mask M.

According to formula (8), the highlight-free image D is calculated based on M, S and I:

in, Represents the multiplication of elements at corresponding positions.

Figure 3(a) shows the encoder structure of the highlight removal module. The encoder of the highlight removal module consists of five convolutional blocks consisting of a convolutional layer, an instance normalization layer, a linear rectifier unit, and a maximum pooling layer. The convolutional layer is used to extract image features, the instance normalization layer is used to receive the output of the convolutional layer and perform normalization, the linear rectifier unit is used to map the output of the instance normalization layer, and the maximum pooling layer is used to reduce the error of the estimated mean caused by the convolutional layer parameter error and retain more texture information. The maximum pooling window size is 2.

Figure 3(b) shows the decoder structure of the highlight removal module. The decoder of the highlight removal module consists of five deconvolution blocks consisting of a deconvolution layer, an instance normalization layer, and a linear rectifier unit. The deconvolution layer is used to expand the dimension of the feature vector, the instance normalization layer is used to receive the output of the deconvolution layer and perform normalization, and the linear rectifier unit is used to map the output of the instance normalization layer.

In the encoder of the highlight removal module, a dense feature fusion connection is used. The dense feature fusion strategy is expressed by formula (9) and formula (10). The features extracted by the convolution block are cascaded with the features extracted by the convolution block of the previous layer, and then upsampled and finally convolved. Generally, the number of channels increases with the number of layers. In order to avoid the dominance of high-level information in the fused features, a convolution operation is introduced to reduce the number of channels to a fixed value so that each feature will contribute the same number of channels in the fused features.

in, Represents the features extracted by convolution block i, up represents upsampling operation, cat represents cascade operation, and conv represents convolution operation.

S203: Using the trained generator network, the illuminated image and the non-highlight image are processed to generate an SVBRDF map of the material corresponding to the input image.

Figure 5(a) shows the encoder structure of the generator network. The encoder of the generator network has 8 convolutional layers for downsampling. The first 7 convolutional layers are followed by an instantiation normalization layer and a linear rectifier unit, and the last convolutional layer is followed by an instantiation normalization layer. Figure 5(b) shows the decoder structure of the generator network. The four decoders of the generator network have the same structure. Each decoder network has 8 deconvolutional layers for upsampling. The first 7 deconvolutional layers are followed by an instantiation normalization layer and a linear rectifier unit, and the last deconvolutional layer is followed by an instantiation normalization layer. The highlight-free image generated by the highlight removal module is fed into the encoder to extract features, and the extracted diffuse features are fed into the decoder De _d to generate a diffuse map. The captured image is fed into the encoder to extract features, and the extracted features are input into the decoders _Den , De _r , and De _s to generate normal maps, roughness maps, and reflection maps, respectively.

The discriminators D _d , D _n , _Dr , D _s , and D _render have the same structure. Figure 6 shows the structure of the discriminator network. The discriminator network consists of a cascade operation and three convolution blocks consisting of a convolution layer, an instance normalization layer, a linear rectifier unit, and a maximum pooling layer. The cascade operation concatenates the generated map with the real map, the convolution layer is used to extract image features, the instance normalization layer is used to receive the output of the convolution layer and perform normalization, and the linear rectifier unit is used to map the output of the instance normalization layer.

Compared with the prior art, the quality of the generated material maps and the re-rendering results are significantly improved on both synthetic data and real data. This embodiment reduces the re-rendering root mean square error RMSE to 0.079, the normal map error to 0.042, the normal map error to 0.062, the roughness map error to 0.102, and the reflection map error to 0.062, and significantly removes the highlight artifact problem existing in the prior art.

The present disclosure proposes a method for estimating the SVBRDF of a single image. The method proposes a highlight removal module with multi-level highlight recognition and dense feature fusion connections. An encoder-decoder network with a shared encoder and four decoders is used as the generator network, and a set of discriminators is used to train the generator network to distinguish between generated maps and real maps, and rendered images and real images. Extensive experiments on synthetic and real image datasets show that compared with the existing technology, high-quality SVBRDF can be generated, and more realistic rendering results can be produced for input images with a large number of highlight areas.

Embodiment 2:

The purpose of this embodiment is to provide a SVBRDF material modeling system based on a single highlight image.

A SVBRDF material modeling system based on a single highlight image, comprising:

A data acquisition unit, which is used to acquire a highlight image of the surface of an object material;

A highlight removal unit, which is used to remove the highlights of the highlight image by means of dense feature fusion connection and highlight multi-level recognition to obtain a highlight-free image;

An object surface material estimation unit is used to input the highlight image and the non-highlight image into a pre-trained generator network to obtain a spatially varying bidirectional reflectance distribution function of the object surface, and then obtain a corresponding material map; wherein the generator network includes a shared encoder and a plurality of decoders respectively connected to the shared encoder, and the decoders respectively correspond to the processing of diffuse maps, normal maps, roughness maps and reflection maps.

It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one. The technical details of the system described in this embodiment have been described in detail in Example 1, so they will not be repeated here.

In further embodiments, there is also provided:

An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, it will not be described in detail here.

It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

A computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the method described in embodiment 1 is completed.

The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

Those skilled in the art will appreciate that the units, i.e., algorithm steps, of the various examples described in the present embodiment can be implemented in electronic hardware or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this disclosure.

The SVBRDF material modeling method and system based on a single highlight image provided in the above embodiment can be implemented and have broad application prospects.

The above description is only a preferred embodiment of the present disclosure and is not intended to limit the present disclosure. For those skilled in the art, the present disclosure may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.

Claims (8)

1. A SVBRDF material modeling method based on a single highlight image, characterized by comprising: Get the highlight image of the object material surface; Eliminate the highlights of the highlight image by dense feature fusion connection and highlight multi-level recognition to obtain a highlight-free image; Inputting the highlight image and the non-highlight image into a pre-trained generator network to obtain a spatially varying bidirectional reflectance distribution function of the object surface, and then obtaining a corresponding material map; wherein the generator network includes a shared encoder and a plurality of decoders respectively connected to the shared encoder, and the decoders respectively correspond to the processing of a diffuse map, a normal map, a roughness map, and a reflection map; The generator network is trained collaboratively with the multi-discriminator network, specifically: Obtain a public data set, and construct a training data set by removing duplicates, randomly cropping, and randomly screening the public data set; Based on a pre-constructed minimization loss function, the generator network and the multi-discriminator network are collaboratively trained through a constructed training data set, wherein the quality of the material map generated by the generator network is judged through the multi-discriminator network and the pre-constructed loss function; The multi-discriminator network specifically includes a discriminator for processing a diffuse map, a discriminator for processing a normal map, a discriminator for processing a roughness map, a discriminator for processing a reflection map, and a discriminator for processing a rendering result. 2. An SVBRDF material modeling method based on a single highlight image as described in claim 1 is characterized in that the highlight elimination is performed based on dense feature fusion connection and multi-level highlight recognition, specifically: using an encoder-decoder group with dense feature fusion connection, adopting a multi-level highlight recognition strategy, performing a highlight removal operation on the captured image, and obtaining a highlight-free image of the captured image. 3. An SVBRDF material modeling method based on a single highlight image as described in claim 2, characterized in that the encoder includes a plurality of convolutional layers, instance normalization layers, linear rectification units and maximum pooling layers connected in sequence; the decoder includes a plurality of deconvolutional layers, instance normalization layers and linear rectification units connected in sequence. 4. The SVBRDF material modeling method based on a single highlight image as described in claim 2 is characterized in that the encoder performs a cascade operation on the features extracted by the convolution block and the features extracted by the convolution block in the previous layer based on dense feature fusion connection, then performs an upsampling operation, and finally performs a convolution operation. 5. A SVBRDF material modeling method based on a single highlight image as described in claim 1, characterized in that the encoder of the generator network includes eight convolutional layers for downsampling, the first seven convolutional layers are connected to an instantiated normalization layer and a linear rectification unit, and the last convolutional layer is connected to an instantiated normalization layer. 6. A SVBRDF material modeling system based on a single highlight image, characterized by comprising: A data acquisition unit, which is used to acquire a highlight image of the surface of an object material; A highlight removal unit, which is used to remove the highlights of the highlight image by means of dense feature fusion connection and highlight multi-level recognition to obtain a highlight-free image; An object surface material estimation unit, which is used to input the highlight image and the non-highlight image into a pre-trained generator network to obtain a spatially varying bidirectional reflectance distribution function of the object surface, and then obtain a corresponding material map; wherein the generator network includes a shared encoder and a plurality of decoders respectively connected to the shared encoder, and the decoders respectively correspond to the processing of diffuse reflection map, normal map, roughness map and reflection map; The generator network is trained collaboratively with the multi-discriminator network, specifically: Obtain a public data set, and construct a training data set by removing duplicates, randomly cropping, and randomly screening the public data set; Based on a pre-constructed minimization loss function, the generator network and the multi-discriminator network are collaboratively trained through a constructed training data set, wherein the quality of the material map generated by the generator network is judged through the multi-discriminator network and the pre-constructed loss function; The multi-discriminator network specifically includes a discriminator for processing a diffuse map, a discriminator for processing a normal map, a discriminator for processing a roughness map, a discriminator for processing a reflection map, and a discriminator for processing a rendering result. 7. An electronic device, comprising a memory, a processor and a computer program stored and running on the memory, characterized in that when the processor executes the program, the SVBRDF material modeling method based on a single highlight image as described in any one of claims 1-5 is implemented. 8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the SVBRDF material modeling method based on a single highlight image as described in any one of claims 1 to 5 is implemented.