A physical perception fusion driven adaptive texture generation optimization method and system

Abstract

The application discloses a physical perception fusion driven adaptive texture generation optimization method and system. The method comprises the following steps: collecting material spectral reflectance data and material surface micro texture image data, and performing pretreatment; constructing an initial physical and perception fusion driven texture generation model; adopting a cross-domain joint training strategy, training and optimizing the model through an alternating optimization scheme of physical priority-perception priority-joint fine-tuning, to obtain a physical and perception fusion driven texture generation model; and performing actual texture generation based on the obtained physical and perception fusion driven texture generation model. The method fuses physical laws and human visual perception mechanisms in depth, constructs a new generation of adaptive texture optimization framework, solves the long-standing "physical-perception" split problem in the field of high-fidelity rendering, and realizes a breakthrough in spectral modeling accuracy, rendering efficiency and visual quality.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer graphics, artificial intelligence and high-fidelity rendering, and specifically relates to an adaptive texture generation optimization method and system driven by physical perception fusion. Background Technology

[0002] Existing texture generation technologies are mainly divided into two categories: Physically Based Rendering (PBR) and perception-driven deep learning models. Generally, PBR technology generates high-fidelity textures by accurately modeling the physical interaction between light and materials (such as microsurface theory, energy conservation, and Fresnel reflection). Representative models include: Disney BRDF: a general-purpose material model that parametrically controls properties such as roughness and metallicity, widely used in the film and game industries; and GGX microsurface distribution: simulating complex reflection behavior through more precise microsurface normal distribution. However, PBR technology suffers from insufficient spectral modeling and high computational complexity.

[0003] Perception-driven deep learning models utilize deep learning-based texture generation techniques to extract image features through pre-trained networks (such as VGG and ResNet) and combine them with generative adversarial networks (GANs) or style transfer algorithms to optimize visual quality. Typical methods include: VGG perceptual loss, which improves visual consistency of textures by matching the distance between generated and real images in the high-level feature space of the VGG network; and CycleGAN texture transfer, which achieves cross-domain texture transformation, but the generated textures often violate physical laws (e.g., incorrect specular highlight placement). However, perception-driven deep learning models suffer from physical inconsistencies and a lack of dynamic adaptability.

[0004] Therefore, existing texture generation techniques face a conflict between physical accuracy and technical efficiency. Traditional PBR technology pursues physical realism but suffers from high computational complexity and low efficiency, while deep learning models are more efficient but have poor physical consistency. Furthermore, existing methods struggle to simultaneously satisfy both the global spectral characteristics of the material and the human visual sensitivity to local details. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a physically-aware fusion-driven adaptive texture generation optimization method to achieve high-precision modeling of the spectral characteristics of complex materials while meeting real-time rendering requirements.

[0006] The second objective of this invention is to provide a system for implementing the aforementioned physical perception fusion-driven adaptive texture generation optimization method.

[0007] This invention provides a physically-aware fusion-driven adaptive texture generation optimization method, comprising the following steps:

[0008] S1. Collect material spectral reflectance data and material surface micro-texture image data, and preprocess them to obtain the training dataset;

[0009] S2. Construct an initial physical and perceptual fusion-driven texture generation model; the initial physical and perceptual fusion-driven texture generation model includes a spectral physics module, a visual perception module, and an adaptive fusion module;

[0010] S3. The loss function that combines the mean square error of reflectance and the physical constraint term is used as the loss function of the spectral physics module; the perceptual loss is used as the loss function of the visual perception module; a cross-domain joint training strategy is adopted, and the model is trained and optimized through an alternating optimization scheme of physics priority-perception priority-joint fine-tuning to obtain a texture generation model driven by the fusion of physics and perception.

[0011] S4. Based on the obtained texture generation model driven by the fusion of physics and perception, perform actual texture generation.

[0012] In step S1, a spectrophotometer is used to obtain the reflectance of several materials at a preset wavelength point as material reflectance data; a scanning electron microscope is used to obtain a texture image of the material surface at a preset resolution as material surface micro-texture image data.

[0013] The collected material spectral reflectance data is input into a spectral coding network to generate a continuous spectral response function. The spectral coding network is a fully connected network that maps reflectance data at discrete wavelengths to a continuous spectral response function. The spectral coding network includes an input layer, a hidden layer, and an output layer. The input layer has an N-dimensional network structure, where N is the number of preset wavelengths in the material reflectance data. The hidden layer consists of three fully connected layers with 256 neurons, using the ReLU activation function. The output layer uses 64-dimensional optical disc basis function coefficients, generating a continuous spectral response through basis function combination.

[0014] The continuous spectral response function and the acquired microscopic texture image data of the material surface are used as the training dataset.

[0015] In step S2, the input to the initial physical and perceptual fusion-driven texture generation model is a continuous spectral reflectance function and a set of micro-texture images;

[0016] The spectral reflectance function is input into the spectral physics module. The differential spectral BRDF model is used to model the interaction between illumination and material. Combined with the frequency division radiative transfer technology, the spectrum is divided into several sub-bands for parallel processing to obtain the reflectance characteristic parameters that conform to physical laws, and then input into the adaptive fusion module.

[0017] The micro-texture image is input into the visual perception module, where a multi-scale perception network is used to extract the global semantic features and local detail texture features of the image. A visual saliency attention module is also introduced, and the perceptual loss function is fused with physical constraint terms to finally obtain a texture perception feature map with human visual preferences, which is then input into the adaptive fusion module.

[0018] The adaptive fusion module predicts the metallicity of the material using a ResNet-18 network based on the input data, and then determines the weights of the two sets of input data in the fusion process based on the metallicity of the material, finally obtaining the fused texture map.

[0019] In the spectral physics module, the differential spectral BRDF models the interaction between illumination and materials, specifically as follows:

[0020] The RGB parameters of the traditional PBR are extended to wavelength correlation functions to obtain the spectral Disney BRDF, which enables reflection modeling of materials at different light wavelengths. This is expressed by the following formula: ;in, The wavelength of light; Let be the incident light direction vector; Here is the direction vector of the emitted light; The diffuse reflectance coefficient is wavelength-dependent. The wavelength-dependent specular reflectance coefficient; The normal distribution function of the micro-surface; The wavelength-dependent Fresnel reflection term; G is the geometric attenuation factor; is the spectral response function of the material; for and The included angle; for and The half-angle vector of the normalized average.

[0021] The aforementioned frequency-division radiative transfer technology divides the spectrum into several sub-bands for parallel processing, yielding reflection characteristic parameters that conform to physical laws, specifically:

[0022] The visible spectrum is divided into several sub-bands, and then several wavelengths are selected in each sub-band. These wavelengths are then input into the Disney BRDF spectral model to calculate the material's reflectance characteristics in different sub-bands. Where n is the number of sub-bands;

[0023] A lightweight multi-channel neural network is used to automatically generate the fusion weights for each sub-band based on the material's metallicity, roughness, and incident angle features. ,and ;

[0024] The final texture reflection function is calculated using the following formula: ;in, The final texture reflection characteristic parameters are used as the output of the spectral physics module.

[0025] In the visual perception module, the multi-scale perception network consists of two parallel branches, including a global semantic branch and a local detail branch;

[0026] The global semantic branch uses the convolutional pooling part of the pre-trained VGG-19 network to extract feature maps; the global semantic branch consists of a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, a third max pooling layer, a fourth convolutional layer, a fourth max pooling layer, a fifth convolutional layer, and a fifth max pooling layer connected in sequence; the feature map output by the fifth max pooling layer is selected to extract the semantic information and overall style distribution of the texture image;

[0027] The local detail branch adopts the U-Net encoding and decoding structure, and a visual saliency attention module is introduced between the encoder and decoder;

[0028] The formula for calculating the attention weights of the visual saliency attention module is as follows: ;in, For the input texture map Features extracted by the low-frequency feature extractor; For the input texture map Features extracted by a high-frequency feature extractor; For feature splicing; Use the Sigmoid activation function; The attention score map for each pixel, and .

[0029] In the adaptive fusion module, a dynamic weight allocator is used to adjust the contribution weights of the spectral physics module and the visual perception module, calculated using the following formula: ;in, The contribution weights for the spectral physics module; The contribution weight to the visual perception module; is the metallicity parameter of the material; k is the parameter to be learned.

[0030] In step S3, the loss function of the spectral physics module is expressed using the following formula: ;in, This represents the mean square error of reflectivity. For physical constraints;

[0031] The physical constraint terms are expressed using the following formula: ;in, The total reflected energy per unit pixel for generating textures for the model; Theoretically, the actual reflected energy per pixel of texture; These are the physical constraint weighting coefficients;

[0032] The visual perception module employs self-supervised contrastive loss to train the local detail branch; and uses perceptual loss to fine-tune the global semantic branch and the local detail branch; the perceptual loss is expressed using the following formula: ;in, This is for global semantic branch processing; Texture image data generated for global semantic branches; It is real texture image data; To perceive the mean square error in the feature space; The weighting coefficients for the global semantic branch loss; The weighting coefficients are used to calculate the loss of local detail branches. For local detail branch loss;

[0033] The cross-domain joint training strategy includes three phases: a physics-first phase, a perception-first phase, and a joint fine-tuning phase.

[0034] In the physics-priority phase, the parameters of the perception network in the visual perception module are first fixed, and the loss function of the spectral physics module is minimized. Then, it enters the perception-first stage; in the perception-first stage, the model parameters in the spectral physics module are fixed, the loss function of the visual perception module is minimized, and then it enters the joint fine-tuning stage; in the joint fine-tuning stage, the following objective function is used to synchronously update and optimize the spectral physics module and the visual perception module: ;in, The regularization coefficient is used. It is a regular term; This refers to the complete set of parameters to be trained in a texture generation model driven by the fusion of physics and perception. The weighting factor is the physical loss factor. The weighting coefficient for perceived loss.

[0035] Step S4 includes the following steps:

[0036] The target material's spectral reflectance data and surface micro-texture image are acquired, and the target material's spectral reflectance data is input into a spectral coding network to obtain a continuous spectral reflectance function;

[0037] The spectral reflectance function and the micro-texture image of the material surface are input into the physical and perceptual fusion-driven texture generation model obtained in step S3, and finally the optimized texture map is obtained.

[0038] The present invention also provides a system for implementing the physical perception fusion-driven adaptive texture generation optimization method, including a data acquisition and processing module, a model building module, a model training module, and a texture optimization generation module;

[0039] The data acquisition and processing module collects material spectral reflectance data and material surface micro-texture image data, performs preprocessing to obtain a training dataset, and uploads the data to the model training module.

[0040] The model building module constructs an initial texture generation model driven by the fusion of physics and perception, and uploads the data to the model training module;

[0041] The model training module uses the loss function that fuses the mean square error of reflectance and the physical constraint term as the loss function of the spectral physics module based on the received data; it uses the perceptual loss as the loss function of the visual perception module; and it adopts a cross-domain joint training strategy to train and optimize the model through an alternating optimization scheme of physics priority-perception priority-joint fine-tuning to obtain a texture generation model driven by the fusion of physics and perception, and uploads the data to the texture optimization and generation module.

[0042] The texture optimization and generation module generates actual textures based on the received data and the obtained texture generation model driven by the fusion of physics and perception.

[0043] This invention discloses an adaptive texture generation optimization method and system driven by physical perception fusion. By deeply integrating physical laws and human visual perception mechanisms, a new generation of adaptive texture optimization framework is constructed, which solves the long-standing problem of "physical-perception" separation in the field of high-fidelity rendering and achieves breakthroughs in spectral modeling accuracy, rendering efficiency and visual quality. Attached Figure Description

[0044] Figure 1 This is a schematic flowchart of the method of the present invention;

[0045] Figure 2 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0046] This invention provides an adaptive texture generation optimization method driven by physical perception fusion, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps:

[0047] S1. Collect material spectral reflectance data and material surface micro-texture image data, and preprocess them to obtain the training dataset;

[0048] In step S1, a spectrophotometer is used to obtain the reflectance of several materials at a preset wavelength point as material reflectance data; a scanning electron microscope is used to obtain a texture image of the material surface at a preset resolution as material surface micro-texture image data.

[0049] The collected material spectral reflectance data is input into a spectral coding network to generate a continuous spectral response function. The spectral coding network is a fully connected network that maps reflectance data at discrete wavelengths to a continuous spectral response function. The spectral coding network includes an input layer, a hidden layer, and an output layer. The input layer has an N-dimensional network structure, where N is the number of preset wavelengths in the material reflectance data. The hidden layer consists of three fully connected layers with 256 neurons, using the ReLU activation function. The output layer uses 64-dimensional optical disc basis function coefficients, generating a continuous spectral response through basis function combination.

[0050] The continuous spectral response function and the acquired microscopic texture image data of the material surface are used as the training dataset.

[0051] S2. Construct an initial physical and perceptual fusion-driven texture generation model; the initial physical and perceptual fusion-driven texture generation model includes a spectral physics module, a visual perception module, and an adaptive fusion module;

[0052] In step S2, the input to the initial physical and perceptual fusion-driven texture generation model is a continuous spectral reflectance function and a set of micro-texture images;

[0053] The spectral reflectance function is input into the spectral physics module. The differential spectral BRDF model is used to model the interaction between illumination and material. Combined with the frequency division radiative transfer technology, the spectrum is divided into several sub-bands for parallel processing to obtain the reflectance characteristic parameters that conform to physical laws, and then input into the adaptive fusion module.

[0054] The micro-texture image is input into the visual perception module, where a multi-scale perception network is used to extract the global semantic features and local detail texture features of the image. A visual saliency attention module is also introduced, and the perceptual loss function is fused with physical constraint terms to finally obtain a texture perception feature map with human visual preferences, which is then input into the adaptive fusion module.

[0055] The adaptive fusion module predicts the metallicity of the material using a ResNet-18 network based on the input data, and then determines the weights of the two sets of input data in the fusion process based on the metallicity of the material, finally obtaining the fused texture map.

[0056] In the spectral physics module, the differential spectral BRDF models the interaction between illumination and materials, specifically as follows:

[0057] The RGB parameters of the traditional PBR are extended to wavelength correlation functions to obtain the spectral Disney BRDF, which enables reflection modeling of materials at different light wavelengths. This is expressed by the following formula: ;in, The wavelength of light; Let be the incident light direction vector; Here is the direction vector of the emitted light; The diffuse reflectance coefficient is wavelength-dependent. The wavelength-dependent specular reflectance coefficient; The normal distribution function of the micro-surface; The wavelength-dependent Fresnel reflection term; G is the geometric attenuation factor; is the spectral response function of the material; for and The included angle; for and The half-angle vector of the normalized average.

[0058] The aforementioned frequency-division radiative transfer technology divides the spectrum into several sub-bands for parallel processing, yielding reflection characteristic parameters that conform to physical laws, specifically:

[0059] The visible spectrum is divided into several sub-bands, and then several wavelengths are selected in each sub-band. These wavelengths are then input into the Disney BRDF spectral model to calculate the material's reflectance characteristics in different sub-bands. Where n is the number of sub-bands;

[0060] A lightweight multi-channel neural network is used to automatically generate the fusion weights for each sub-band based on the material's metallicity, roughness, and incident angle features. ,and ;

[0061] The final texture reflection function is calculated using the following formula: ;in, The final texture reflection characteristic parameters are used as the output of the spectral physics module.

[0062] In the visual perception module, the multi-scale perception network consists of two parallel branches, including a global semantic branch and a local detail branch;

[0063] The global semantic branch uses the convolutional pooling part of the pre-trained VGG-19 network to extract feature maps; the global semantic branch consists of a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, a third max pooling layer, a fourth convolutional layer, a fourth max pooling layer, a fifth convolutional layer, and a fifth max pooling layer connected in sequence; the feature map output by the fifth max pooling layer is selected to extract the semantic information and overall style distribution of the texture image;

[0064] The local detail branch adopts the U-Net encoding and decoding structure, and a visual saliency attention module is introduced between the encoder and decoder;

[0065] The formula for calculating the attention weights of the visual saliency attention module is as follows: ;in, For the input texture map Features extracted by the low-frequency feature extractor; For the input texture map Features extracted by a high-frequency feature extractor; For feature splicing; Use the Sigmoid activation function; The attention score map for each pixel, and .

[0066] In the adaptive fusion module, a dynamic weight allocator is used to adjust the contribution weights of the spectral physics module and the visual perception module, calculated using the following formula: ;in, The contribution weights for the spectral physics module; The contribution weight to the visual perception module; is the metallicity parameter of the material; k is the parameter to be learned.

[0067] S3. The loss function that combines the mean square error of reflectance and the physical constraint term is used as the loss function of the spectral physics module; the perceptual loss is used as the loss function of the visual perception module; a cross-domain joint training strategy is adopted, and the model is trained and optimized through an alternating optimization scheme of physics priority-perception priority-joint fine-tuning to obtain a texture generation model driven by the fusion of physics and perception.

[0068] In step S3, the loss function of the spectral physics module is expressed using the following formula: ;in, This represents the mean square error of reflectivity. For physical constraints;

[0069] The physical constraint terms are expressed using the following formula: ;in, The total reflected energy per unit pixel for generating textures for the model; Theoretically, the actual reflected energy per pixel of texture; These are the physical constraint weighting coefficients;

[0070] The visual perception module employs self-supervised contrastive loss to train the local detail branch; and uses perceptual loss to fine-tune the global semantic branch and the local detail branch; the perceptual loss is expressed using the following formula: ;in, This is for global semantic branch processing; Texture image data generated for global semantic branches; It is real texture image data; To perceive the mean square error in the feature space; The weighting coefficients for the global semantic branch loss; The weighting coefficients are used to calculate the loss of local detail branches. For local detail branch loss;

[0071] The cross-domain joint training strategy includes three phases: a physics-first phase, a perception-first phase, and a joint fine-tuning phase.

[0072] In the physics-priority phase, the parameters of the perception network in the visual perception module are first fixed, and the loss function of the spectral physics module is minimized. Then, it enters the perception-first stage; in the perception-first stage, the model parameters in the spectral physics module are fixed, the loss function of the visual perception module is minimized, and then it enters the joint fine-tuning stage; in the joint fine-tuning stage, the following objective function is used to synchronously update and optimize the spectral physics module and the visual perception module: ;in, The regularization coefficient is used. It is a regular term; This refers to the complete set of parameters to be trained in a texture generation model driven by the fusion of physics and perception. The weighting factor is the physical loss factor. The weighting coefficient for perceived loss.

[0073] S4. Based on the obtained texture generation model driven by the fusion of physics and perception, perform actual texture generation.

[0074] Step S4 includes the following steps:

[0075] The target material's spectral reflectance data and surface micro-texture image are acquired, and the target material's spectral reflectance data is input into a spectral coding network to obtain a continuous spectral reflectance function;

[0076] The spectral reflectance function and the micro-texture image of the material surface are input into the physical and perceptual fusion-driven texture generation model obtained in step S3, and finally the optimized texture map is obtained.

[0077] In practical applications, the texture generation model driven by the fusion of physics and perception is lightweighted through channel pruning and precision quantization. The model is exported in ONNX format and packaged into a Unity plugin. Combined with the hardware acceleration of CUDA core and NVIDIA RTX graphics card, the 4K resolution real-time rendering function in Unity HDRP rendering pipeline is realized, and finally a high-precision texture map and plugin call are obtained for real-time graphics system.

[0078] The present invention also provides a system for implementing the adaptive texture generation optimization method driven by physical perception fusion, the structural schematic diagram of which is shown below. Figure 2 As shown, it includes a data acquisition and processing module, a model building module, a model training module, and a texture optimization and generation module;

[0079] The data acquisition and processing module collects material spectral reflectance data and material surface micro-texture image data, performs preprocessing to obtain a training dataset, and uploads the data to the model training module.

[0080] The model building module constructs an initial texture generation model driven by the fusion of physics and perception, and uploads the data to the model training module;

[0081] The model training module uses the loss function that fuses the mean square error of reflectance and the physical constraint term as the loss function of the spectral physics module based on the received data; it uses the perceptual loss as the loss function of the visual perception module; and it adopts a cross-domain joint training strategy to train and optimize the model through an alternating optimization scheme of physics priority-perception priority-joint fine-tuning to obtain a texture generation model driven by the fusion of physics and perception, and uploads the data to the texture optimization and generation module.

[0082] The texture optimization and generation module generates actual textures based on the received data and the obtained texture generation model driven by the fusion of physics and perception.

[0083] The method of the present invention will be further described below with reference to an embodiment:

[0084] The reflectance of 100 materials at 32 wavelength points (400-700nm, spaced 10nm apart) was obtained using a spectrophotometer (such as Ocean Optics USB4000);

[0085] High-resolution images of the material surface with texture at a resolution of 0.1 μm / pixel were obtained using a scanning electron microscope (SEM, such as Hitachi SU5000). In this embodiment, 100,000 high-resolution material photos containing textures were used, sourced from MIT OpenMaterials, Adobe Stock, and a self-built database.

[0086] A pre-trained spectral coding network is used to map reflectance data into a continuous spectral function;

[0087] The model was trained using the data above.

[0088] The number of network parameters was reduced by 40% through channel pruning and quantization (FP32→INT8);

[0089] A custom kernel was written using CUDA to support ray tracing hardware acceleration for NVIDIA RTX series graphics cards;

[0090] The trained PhysPerceptNet was converted to ONNX format, and a Unity plugin was developed to enable one-click material generation and real-time preview.

[0091] Finally, the obtained model was compared with existing models such as Disney BRDF, NeRF-Texture, and PurePercept in a quantitative comparison experiment. The results are shown in the table below:

[0092] Table 1 Quantitative Comparison Experiment Results

[0093]

[0094] As shown in Table 1, this invention constructs a new generation of adaptive texture optimization framework by deeply integrating physical laws and human visual perception mechanisms, solving the long-standing problem of "physics-perception" separation in the field of high-fidelity rendering. Experiments demonstrate that this method achieves industry-leading levels in spectral accuracy, rendering efficiency, and visual quality, providing core technical support for emerging fields such as digital twins.

Claims

1. A physically-perceived fusion-driven adaptive texture generation optimization method, characterized in that, Includes the following steps: S1. Collect material spectral reflectance data and material surface micro-texture image data, and preprocess them to obtain the training dataset; S2. Construct an initial physical and perceptual fusion-driven texture generation model; the initial physical and perceptual fusion-driven texture generation model includes a spectral physics module, a visual perception module, and an adaptive fusion module; S3. The loss function that combines the mean square error of reflectance and the physical constraint term is used as the loss function of the spectral physics module; the perceptual loss is used as the loss function of the visual perception module; a cross-domain joint training strategy is adopted, and the model is trained and optimized through an alternating optimization scheme of physics priority-perception priority-joint fine-tuning to obtain a texture generation model driven by the fusion of physics and perception. S4. Based on the obtained texture generation model driven by the fusion of physics and perception, perform actual texture generation; In step S3, the loss function of the spectral physics module is expressed using the following formula: ;in, This represents the mean square error of reflectivity. For physical constraints; The physical constraint terms are expressed using the following formula: ;in, The total reflected energy per unit pixel for generating textures for the model; Theoretically, the actual reflected energy per pixel of texture. These are the physical constraint weighting coefficients; The visual perception module employs self-supervised contrastive loss to train the local detail branch; and uses perceptual loss to fine-tune the global semantic branch and the local detail branch; the perceptual loss is expressed using the following formula: ;in, This is for global semantic branch processing; Texture image data generated for global semantic branches; It is real texture image data; To perceive the mean square error in the feature space; The weighting coefficients for the global semantic branch loss; The weighting coefficients are used to calculate the loss of local detail branches. For local detail branch loss; The cross-domain joint training strategy includes three phases: a physics-first phase, a perception-first phase, and a joint fine-tuning phase. In the physics-priority phase, the parameters of the perception network in the visual perception module are first fixed, and the loss function of the spectral physics module is minimized. Then, it enters the perception-first stage; in the perception-first stage, the model parameters in the spectral physics module are fixed, the loss function of the visual perception module is minimized, and then it enters the joint fine-tuning stage; in the joint fine-tuning stage, the following objective function is used to synchronously update and optimize the spectral physics module and the visual perception module: ;in, The regularization coefficient is used. It is a regularization term; This refers to the complete set of parameters to be trained in a texture generation model driven by the fusion of physics and perception. The weighting factor is the physical loss factor. The weighting coefficient for perceived loss.

2. The adaptive texture generation optimization method driven by physical perception fusion according to claim 1, characterized in that, In step S1, a spectrophotometer is used to obtain the reflectance of several materials at a preset wavelength point as material reflectance data; a scanning electron microscope is used to obtain a texture image of the material surface at a preset resolution as material surface micro-texture image data. The collected material spectral reflectance data is input into a spectral coding network to generate a continuous spectral response function. The spectral coding network is a fully connected network that maps reflectance data at discrete wavelengths to a continuous spectral response function. The spectral coding network includes an input layer, a hidden layer, and an output layer. The input layer has an N-dimensional network structure, where N is the number of preset wavelengths in the material reflectance data. The hidden layer consists of three fully connected layers with 256 neurons, using the ReLU activation function. The output layer uses 64-dimensional optical disc basis function coefficients, generating a continuous spectral response through basis function combination. The continuous spectral response function and the acquired microscopic texture image data of the material surface are used as the training dataset.

3. The adaptive texture generation optimization method driven by physical perception fusion according to claim 1, characterized in that, In step S2, the input to the initial physical and perceptual fusion-driven texture generation model is a continuous spectral reflectance function and a set of micro-texture images; The spectral reflectance function is input into the spectral physics module. The differential spectral BRDF model is used to model the interaction between illumination and material. Combined with the frequency division radiative transfer technology, the spectrum is divided into several sub-bands for parallel processing to obtain the reflectance characteristic parameters that conform to physical laws, and then input into the adaptive fusion module. The micro-texture image is input into the visual perception module, where a multi-scale perception network is used to extract the global semantic features and local detail texture features of the image. A visual saliency attention module is also introduced, and the perceptual loss function is fused with physical constraint terms to finally obtain a texture perception feature map with human visual preferences, which is then input into the adaptive fusion module. The adaptive fusion module predicts the metallicity of the material using a ResNet-18 network based on the input data, and then determines the weights of the two sets of input data in the fusion process based on the metallicity of the material, finally obtaining the fused texture map.

4. The adaptive texture generation optimization method driven by physical perception fusion according to claim 3, characterized in that, In the spectral physics module, the differential spectral BRDF models the interaction between illumination and materials, specifically as follows: The RGB parameters of the traditional PBR are extended to wavelength correlation functions to obtain the spectral Disney BRDF, which enables reflection modeling of materials at different light wavelengths. This is expressed by the following formula: ;in, The wavelength of light; Let be the incident light direction vector; Here is the direction vector of the emitted light; The diffuse reflectance coefficient is wavelength-dependent. The wavelength-dependent specular reflectance coefficient; The normal distribution function of the micro-surface; The wavelength-dependent Fresnel reflection term; G is the geometric attenuation factor; is the spectral response function of the material; for and The included angle; for and The half-angle vector of the normalized average.

5. The adaptive texture generation optimization method driven by physical perception fusion according to claim 3, characterized in that, The aforementioned frequency-division radiative transfer technology divides the spectrum into several sub-bands for parallel processing, yielding reflection characteristic parameters that conform to physical laws, specifically: The visible spectrum is divided into several sub-bands, and then several wavelengths are selected in each sub-band. These wavelengths are then input into the DisneyBRDF spectral model to calculate the reflectance characteristics of the material in different sub-bands. Where n is the number of sub-bands; A lightweight multi-channel neural network is used to automatically generate the fusion weights for each sub-band based on the material's metallicity, roughness, and incident angle features. ,and ; The final texture reflection function is calculated using the following formula: ;in, The final texture reflection characteristic parameters are used as the output of the spectral physics module.

6. The adaptive texture generation optimization method driven by physical perception fusion according to claim 3, characterized in that, In the visual perception module, the multi-scale perception network consists of two parallel branches, including a global semantic branch and a local detail branch; The global semantic branch uses the convolutional pooling part of the pre-trained VGG-19 network to extract feature maps; the global semantic branch consists of a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, a third max pooling layer, a fourth convolutional layer, a fourth max pooling layer, a fifth convolutional layer, and a fifth max pooling layer connected in sequence; the feature map output by the fifth max pooling layer is selected to extract the semantic information and overall style distribution of the texture image; The local detail branch adopts the U-Net encoding and decoding structure, and a visual saliency attention module is introduced between the encoder and decoder; The formula for calculating the attention weights of the visual saliency attention module is as follows: ;in, For the input texture map Features extracted by the low-frequency feature extractor; For the input texture map Features extracted by a high-frequency feature extractor; For feature splicing; Use the Sigmoid activation function; The attention score map for each pixel, and .

7. The adaptive texture generation optimization method driven by physical perception fusion according to claim 3, characterized in that, In the adaptive fusion module, a dynamic weight allocator is used to adjust the contribution weights of the spectral physics module and the visual perception module, calculated using the following formula: ;in, The contribution weights for the spectral physics module; The contribution weight to the visual perception module; is the metallicity parameter of the material; k is the parameter to be learned.

8. The adaptive texture generation optimization method driven by physical perception fusion according to claim 1, characterized in that, Step S4 includes the following steps: The target material's spectral reflectance data and surface micro-texture image are acquired, and the target material's spectral reflectance data is input into a spectral coding network to obtain a continuous spectral reflectance function; The spectral reflectance function and the micro-texture image of the material surface are input into the physical and perceptual fusion-driven texture generation model obtained in step S3, and finally the optimized texture map is obtained.

9. A system for implementing the physically-aware fusion-driven adaptive texture generation optimization method according to any one of claims 1 to 8, characterized in that, It includes a data acquisition and processing module, a model building module, a model training module, and a texture optimization and generation module; The data acquisition and processing module collects material spectral reflectance data and material surface micro-texture image data, performs preprocessing to obtain a training dataset, and uploads the data to the model training module. The model building module constructs an initial texture generation model driven by the fusion of physics and perception, and uploads the data to the model training module; The model training module uses the loss function that fuses the mean square error of reflectance and the physical constraint term as the loss function of the spectral physics module based on the received data; it uses the perceptual loss as the loss function of the visual perception module; and it adopts a cross-domain joint training strategy to train and optimize the model through an alternating optimization scheme of physics priority-perception priority-joint fine-tuning to obtain a texture generation model driven by the fusion of physics and perception, and uploads the data to the texture optimization and generation module. The texture optimization and generation module generates actual textures based on the received data and the obtained texture generation model driven by the fusion of physics and perception.

Images