Infrared sea surface background simulation method and system with texture intelligent generation

Abstract

The application provides an infrared sea surface background simulation method and system for texture intelligent generation, comprising the following steps: S1, extracting intrinsic radiation features from measured sea surface background data; S2, training a model by using the intrinsic radiation feature data and expanding sea surface texture; and S3, performing infrared sea surface texture random mapping and sea surface dynamic scene rendering. The application can provide efficient, fast, economical, large sample and high-fidelity simulation data for learning, training and testing of intelligent algorithms, improve the target characteristic modeling and simulation level and capability, and has great significance for the development of intelligent detection and identification technology of an infrared sensor system.

Description

Technical Field

[0001] This invention relates to the field of target characteristic modeling and simulation, specifically to a method and system for simulating infrared sea surface backgrounds with intelligent texture generation, and more specifically to a method and system for simulating high-fidelity infrared sea surface background images. Background Technology

[0002] The increasing complexity of battlefield environments and the uncertainty of target characteristics pose growing challenges to the development of infrared sensor systems and technologies. Intelligent technology is a key issue in the research and development of next-generation infrared sensor systems. The intelligence of an infrared sensor system refers to automatically making decisions using sufficient and effective data. Therefore, it is essential to establish a target feature database that includes an appropriate number of representative samples for effective training and testing.

[0003] The advantage of simulating infrared image data is its ability to generate data covering a wide range of operating conditions. Furthermore, it allows for the generation of annotations and precise metadata about the target, environment, and detection sensitivities during the data production process. However, unfortunately, infrared data is also very difficult to model accurately, making the generation of high-fidelity infrared images extremely complex and computationally challenging.

[0004] Currently, target characteristic modeling methods researched both domestically and internationally can be divided into two types. Method one belongs to the forward modeling process. By analyzing the generation mechanism of sample data from simulated entities, a 3D model of the simulated entity is constructed, completing the process of generating 2D image data from the 3D model. Its advantage is that it can generate sample data from any viewpoint and at any distance. The disadvantages of this method are that the modeling process is complex, consuming significant manpower and time costs. Furthermore, because it only targets a single scene or a single target, the simulation scene is singular and exhibits obvious periodicity, resulting in poor model generalization performance.

[0005] Method two involves a reverse modeling process. It uses measured data of the simulated entity and employs intelligent methods to extract model features from the measured data, establishing a corresponding intelligent generative model. By utilizing the correlation between the measured data and the generated sample data, more sample data is generated. This method completes the simulation process from small sample image data to large sample image data. The advantages of this method are richer detailed information, a simpler modeling process, lower labor and time costs, and higher conversion efficiency. Its disadvantages are unstable training networks, lower controllability of sample data, and insufficient sample coverage.

[0006] Based on the advantages and disadvantages of the two methods mentioned above, the intelligent algorithm of the infrared sensor system must construct a simulation of target environment scene sample data with a large sample size and high fidelity.

[0007] Literature CN103123670A (application number: CN201310066848.X) discloses a texture-based method for dynamic simulation of infrared rough sea surfaces, mainly addressing the problems of low realism and real-time performance in existing infrared rough sea surface simulation technologies. The implementation process is as follows: A formula for the infrared radiance of the sea surface is established using the Torrance-Sparrow illumination model; the sea surface emissivity, micro-facet distribution probability, solar radiance, sky radiance, atmospheric path radiance, and atmospheric transmittance are calculated using the atmospheric calculation software Atmosphere; the calculation results are stored as a DDS texture map, and the DDS texture map is written into a material script using the Cg language; the material script is parsed and compiled using a GPU, loaded into video memory, and executed; the executed code is used to complete the real-time simulation of the infrared rough sea surface. However, this invention does not extract texture features from small-sample measured data and expand the sample data using 3D modeling and simulation methods. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for simulating infrared sea surface backgrounds using intelligent texture generation.

[0009] An infrared sea surface background simulation method for intelligent texture generation according to the present invention includes:

[0010] Step S1: Extract intrinsic radiation features from measured sea surface background data;

[0011] Step S2: Train the model using intrinsic radiation feature data to expand the sea surface texture;

[0012] Step S3: Perform random mapping of infrared sea surface texture and dynamic rendering of sea surface scene.

[0013] Preferably, in step S1:

[0014] A sea surface infrared characteristic inversion model was established to extract intrinsic radiation features from measured sea surface background data; this included data calibration and effect removal, which included atmospheric effect removal and imaging system effect removal.

[0015] Specifically, this includes: using measured infrared image feature data to obtain the true temperature field distribution of the target, and using inversion to obtain the brightness radiation information and temperature distribution information of the sea scene; based on the infrared radiation transmission theory and the photoelectric signal conversion principle, by establishing an infrared illumination model, an atmospheric transmission model, a sensor imaging system effect model and an infrared inversion framework, the measured infrared radiation information of the target is used to back-calculate the target's intrinsic spontaneous infrared radiation distribution information.

[0016] Among them, data calibration is performed by converting and calculating the radiance value based on blackbody radiation calibration and Planck's formula; atmospheric effect removal is performed by calculating the atmospheric transmittance and path radiance value in the shooting band using MODTRAN software in combination with environmental and shooting distance information.

[0017] Based on the radiance value L before reaching the thermal imager ap The intrinsic radiance L of the target is calculated by combining the average atmospheric transmittance and atmospheric path radiance. o To remove imaging system effects, blind pixel compensation and noise suppression are performed on the image. Based on the modulation transfer function theory, the input infrared image is subjected to Fourier transform and inverse filtering transformation, followed by inverse Fourier transform to remove spatial transfer effects. The intrinsic temperature value or radiance value is obtained based on the data calibration results.

[0018] Preferably, in step S2:

[0019] An intelligent texture generation model is established. The model is trained using a sample training set formed by intrinsic radiation feature data. After training, the sea surface texture is expanded based on StyleGAN, and the sea surface texture features are obtained based on the expansion results.

[0020] Specifically, the process includes: using a generative adversarial network (GAN) architecture, generating an infrared sea surface background image dataset with a larger number of measured sea surface image samples than a preset standard, thus expanding the data; extracting the infrared sample image dataset as input for infrared texture feature images; using a generative network to encode and decode the infrared texture feature images to generate infrared texture feature images to be judged; using a discriminative network to judge the generated results, thus completing the construction of the GAN structure; defining and selecting a loss function, and using an optimizer to optimize the loss function of the GAN during training, while gradually adjusting the hyperparameters to obtain the trained GAN; and using the trained GAN to generate sea surface infrared texture feature images that meet the conditions.

[0021] Preferably, step S2.1: Constructing the training sample dataset

[0022] Sea surface temperature data is obtained by inversion, and the pixels of the training set images are cropped to a preset size to generate the actual sample dataset required for training.

[0023] Step S2.2: Intelligent Model Construction and Training

[0024] StyleGAN employs a progressive training method from low to high resolution. StyleGAN consists of a generator and a discriminator.

[0025] First, a generative adversarial network (GAN) model is constructed, and the weights and data of the generator network and the discriminator network within the GAN are initialized.

[0026] Next, define the model loss function and use the Adam optimizer to optimize the loss function;

[0027] Step S2.3: Train the network model using the backpropagation algorithm. Backpropagation includes:

[0028] Step S2.3.1: Forward propagation, the input is fed into the network, and the network calculates the predicted value; the error between the predicted value and the true label of the input is calculated;

[0029] Step S2.3.2: Weight update. For each parameter in the network that needs to be updated, calculate the partial derivative of the error with respect to that parameter, and update the parameter accordingly.

[0030] Step S2.4: Adjust the hyperparameters to make the network converge, iterate the backpropagation process until the number of iterations exceeds the set number, and complete the training process;

[0031] Step S2.5: Fix the generator, save the parameters and network model after training, and generate images that meet the preset conditions.

[0032] Preferably, in step S3:

[0033] A sea surface infrared radiation model was established. A three-dimensional sea surface texture mapping method was adopted. The brightness distribution calculation results of the sea surface BRDF were modulated by the gray-scale feature modulation factor. The brightness characteristics of the sea surface height field were modulated by the UV unrolling mapping sampling method. The dynamic driving of the three-dimensional scene was realized by the rendering engine. By inputting different sensor parameters and adding atmospheric transmission effects, the sea surface scene simulation under different viewpoints and distances was realized.

[0034] Specifically, this includes: establishing a three-dimensional infrared radiation transmission model of the sea surface considering the shimmering of the sea surface and the radiation characteristics of the sea surface antenna; mapping the extended infrared texture features onto the sea surface radiation model; modulating the brightness distribution calculation results of the sea surface BRDF using grayscale feature modulation factors; modulating the brightness features of the sea surface height field using UV unrolling mapping sampling method; and generating a three-dimensional sea scene infrared image sequence data based on Unreal Engine rendering.

[0035] An infrared sea surface background simulation system for intelligent texture generation according to the present invention includes:

[0036] Module M1: Extracts intrinsic radiation features from measured sea surface background data;

[0037] Module M2: Trains the model using intrinsic radiation feature data and expands the sea surface texture;

[0038] Module M3: Performs random mapping of infrared sea surface textures and dynamic rendering of sea surface scenes.

[0039] Preferably, in module M1:

[0040] A sea surface infrared characteristic inversion model was established to extract intrinsic radiation features from measured sea surface background data; this included data calibration and effect removal, which included atmospheric effect removal and imaging system effect removal.

[0041] Specifically, this includes: using measured infrared image feature data to obtain the true temperature field distribution of the target, and using inversion to obtain the brightness radiation information and temperature distribution information of the sea scene; based on the infrared radiation transmission theory and the photoelectric signal conversion principle, by establishing an infrared illumination model, an atmospheric transmission model, a sensor imaging system effect model and an infrared inversion framework, the measured infrared radiation information of the target is used to back-calculate the target's intrinsic spontaneous infrared radiation distribution information.

[0042] Among them, data calibration is performed by converting and calculating the radiance value based on blackbody radiation calibration and Planck's formula; atmospheric effect removal is performed by calculating the atmospheric transmittance and path radiance value in the shooting band using MODTRAN software in combination with environmental and shooting distance information.

[0043] Based on the radiance value L before reaching the thermal imager ap The intrinsic radiance L of the target is calculated by combining the average atmospheric transmittance and atmospheric path radiance. o To remove imaging system effects, blind pixel compensation and noise suppression are performed on the image. Based on the modulation transfer function theory, the input infrared image is subjected to Fourier transform and inverse filtering transformation, followed by inverse Fourier transform to remove spatial transfer effects. The intrinsic temperature value or radiance value is obtained based on the data calibration results.

[0044] Preferably, in module M2:

[0045] An intelligent texture generation model is established. The model is trained using a sample training set formed by intrinsic radiation feature data. After training, the sea surface texture is expanded based on StyleGAN, and the sea surface texture features are obtained based on the expansion results.

[0046] Specifically, the process includes: using a generative adversarial network (GAN) architecture, generating an infrared sea surface background image dataset with a larger number of measured sea surface image samples than a preset standard, thus expanding the data; extracting the infrared sample image dataset as input for infrared texture feature images; using a generative network to encode and decode the infrared texture feature images to generate infrared texture feature images to be judged; using a discriminative network to judge the generated results, thus completing the construction of the GAN structure; defining and selecting a loss function, and using an optimizer to optimize the loss function of the GAN during training, while gradually adjusting the hyperparameters to obtain the trained GAN; and using the trained GAN to generate sea surface infrared texture feature images that meet the conditions.

[0047] Preferably, module M2.1: constructing the training sample dataset

[0048] Sea surface temperature data is obtained by inversion, and the pixels of the training set images are cropped to a preset size to generate the actual sample dataset required for training.

[0049] Module M2.2: Intelligent Model Construction and Training

[0050] StyleGAN employs a progressive training method from low to high resolution. StyleGAN consists of a generator and a discriminator.

[0051] First, a generative adversarial network (GAN) model is constructed, and the weights and data of the generator network and the discriminator network within the GAN are initialized.

[0052] Next, define the model loss function and use the Adam optimizer to optimize the loss function;

[0053] Module M2.3: The network model is trained using the backpropagation algorithm. Backpropagation includes:

[0054] Module M2.3.1: Forward propagation, feeds the input into the network, and the network calculates the predicted value; calculates the error between the predicted value and the true label of the input;

[0055] Module M2.3.2: Weight Update. For each parameter in the network that needs to be updated, calculate the partial derivative of the error with respect to that parameter, and update the parameter accordingly.

[0056] Module M2.4: Adjust hyperparameters to bring the network to converge, iterate the backpropagation process until the number of iterations exceeds the set number, and complete the training process;

[0057] Module M2.5: Fixed generator, saves the parameters and network model after training, and generates images that meet preset conditions.

[0058] Preferably, in module M3:

[0059] A sea surface infrared radiation model was established. A three-dimensional sea surface texture mapping method was adopted. The brightness distribution calculation results of the sea surface BRDF were modulated by the gray-scale feature modulation factor. The brightness characteristics of the sea surface height field were modulated by the UV unrolling mapping sampling method. The dynamic driving of the three-dimensional scene was realized by the rendering engine. By inputting different sensor parameters and adding atmospheric transmission effects, the sea surface scene simulation under different viewpoints and distances was realized.

[0060] Specifically, this includes: establishing a three-dimensional infrared radiation transmission model of the sea surface considering the shimmering of the sea surface and the radiation characteristics of the sea surface antenna; mapping the extended infrared texture features onto the sea surface radiation model; modulating the brightness distribution calculation results of the sea surface BRDF using grayscale feature modulation factors; modulating the brightness features of the sea surface height field using UV unrolling mapping sampling method; and generating a three-dimensional sea scene infrared image sequence data based on Unreal Engine rendering.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. This invention introduces intelligent methods on the basis of traditional 3D modeling methods to improve the realism of models in 3D modeling methods, and proposes an infrared sea surface background simulation method with intelligent texture generation;

[0063] 2. This invention extracts texture features from small sample data and expands the sample data by combining it with three-dimensional modeling and simulation methods to meet the requirements of rich sample size, full coverage, and high realism.

[0064] 3. This invention will provide efficient, fast, economical, large-sample, and highly realistic simulation data for the learning, training, and testing of intelligent algorithms, improve the level and capability of target characteristic modeling and simulation, and has great significance for the development of intelligent detection and recognition technology of infrared sensor systems. Attached Figure Description

[0065] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0066] Figure 1 A schematic diagram of an infrared sea surface background simulation method for intelligent texture generation;

[0067] Figure 2 Results for both bright band and non-bright band image sets;

[0068] Figure 3 A schematic diagram of the structure of the StyleGAN-based sea surface texture augmentation model;

[0069] Figure 4 This is a simulated infrared sea surface scene image that includes both bright and non-bright bands of sunlight on the sea surface. Detailed Implementation

[0070] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0071] Example 1:

[0072] This invention is primarily applied to the performance testing and verification of infrared sensor systems, relating to the field of target characteristic modeling and simulation. It is a method that supports high-fidelity infrared sea surface scene image simulation. This method is mainly used for the performance testing and verification of infrared sensor systems. It extracts intrinsic characteristics from measured image data, uses intelligent networks to learn and generate a large sample of sea surface infrared texture information, and adjusts the texture information in the sea surface infrared radiation characteristic model to achieve high-fidelity sea surface scene simulation. It can test the detection and recognition performance of infrared sensor systems in the laboratory and reproduce the entire process of field sea surface tests, allowing the tested products to identify problems and conduct performance verification.

[0073] The present invention provides a method for simulating infrared sea surface backgrounds using intelligent texture generation, such as... Figures 1-4 As shown, it includes:

[0074] Step S1: Extract intrinsic radiation features from measured sea surface background data;

[0075] Specifically, in step S1:

[0076] A sea surface infrared characteristic inversion model was established to extract intrinsic radiation features from measured sea surface background data; this included data calibration and effect removal, which included atmospheric effect removal and imaging system effect removal.

[0077] Specifically, this includes: using measured infrared image feature data to obtain the true temperature field distribution of the target, and using inversion to obtain the brightness radiation information and temperature distribution information of the sea scene; based on the infrared radiation transmission theory and the photoelectric signal conversion principle, by establishing an infrared illumination model, an atmospheric transmission model, a sensor imaging system effect model and an infrared inversion framework, the measured infrared radiation information of the target is used to back-calculate the target's intrinsic spontaneous infrared radiation distribution information.

[0078] Among them, data calibration is performed by converting and calculating the radiance value based on blackbody radiation calibration and Planck's formula; atmospheric effect removal is performed by calculating the atmospheric transmittance and path radiance value in the shooting band using MODTRAN software in combination with environmental and shooting distance information.

[0079] Based on the radiance value L before reaching the thermal imager ap The intrinsic radiance L of the target is calculated by combining the average atmospheric transmittance and atmospheric path radiance. o To remove imaging system effects, blind pixel compensation and noise suppression are performed on the image. Based on the modulation transfer function theory, the input infrared image is subjected to Fourier transform and inverse filtering transformation, followed by inverse Fourier transform to remove spatial transfer effects. The intrinsic temperature value or radiance value is obtained based on the data calibration results.

[0080] Step S2: Train the model using intrinsic radiation feature data to expand the sea surface texture;

[0081] Specifically, in step S2:

[0082] An intelligent texture generation model is established. The model is trained using a sample training set formed by intrinsic radiation feature data. After training, the sea surface texture is expanded based on StyleGAN, and the sea surface texture features are obtained based on the expansion results.

[0083] Specifically, the process includes: using a generative adversarial network (GAN) architecture, generating an infrared sea surface background image dataset with a larger number of measured sea surface image samples than a preset standard, thus expanding the data; extracting the infrared sample image dataset as input for infrared texture feature images; using a generative network to encode and decode the infrared texture feature images to generate infrared texture feature images to be judged; using a discriminative network to judge the generated results, thus completing the construction of the GAN structure; defining and selecting a loss function, and using an optimizer to optimize the loss function of the GAN during training, while gradually adjusting the hyperparameters to obtain the trained GAN; and using the trained GAN to generate sea surface infrared texture feature images that meet the conditions.

[0084] Specifically, step S2.1: Constructing the training sample dataset

[0085] Sea surface temperature data is obtained by inversion, and the pixels of the training set images are cropped to a preset size to generate the actual sample dataset required for training.

[0086] Step S2.2: Intelligent Model Construction and Training

[0087] StyleGAN employs a progressive training method from low to high resolution. StyleGAN consists of a generator and a discriminator.

[0088] First, a generative adversarial network (GAN) model is constructed, and the weights and data of the generator network and the discriminator network within the GAN are initialized.

[0089] Next, define the model loss function and use the Adam optimizer to optimize the loss function;

[0090] Step S2.3: Train the network model using the backpropagation algorithm. Backpropagation includes:

[0091] Step S2.3.1: Forward propagation, the input is fed into the network, and the network calculates the predicted value; the error between the predicted value and the true label of the input is calculated;

[0092] Step S2.3.2: Weight update. For each parameter in the network that needs to be updated, calculate the partial derivative of the error with respect to that parameter, and update the parameter accordingly.

[0093] Step S2.4: Adjust the hyperparameters to make the network converge, iterate the backpropagation process until the number of iterations exceeds the set number, and complete the training process;

[0094] Step S2.5: Fix the generator, save the parameters and network model after training, and generate images that meet the preset conditions.

[0095] Step S3: Perform random mapping of infrared sea surface texture and dynamic rendering of sea surface scene.

[0096] Specifically, in step S3:

[0097] A sea surface infrared radiation model was established. A three-dimensional sea surface texture mapping method was adopted. The brightness distribution calculation results of the sea surface BRDF were modulated by the gray-scale feature modulation factor. The brightness characteristics of the sea surface height field were modulated by the UV unrolling mapping sampling method. The dynamic driving of the three-dimensional scene was realized by the rendering engine. By inputting different sensor parameters and adding atmospheric transmission effects, the sea surface scene simulation under different viewpoints and distances was realized.

[0098] Specifically, this includes: establishing a three-dimensional infrared radiation transmission model of the sea surface considering the shimmering of the sea surface and the radiation characteristics of the sea surface antenna; mapping the extended infrared texture features onto the sea surface radiation model; modulating the brightness distribution calculation results of the sea surface BRDF using grayscale feature modulation factors; modulating the brightness features of the sea surface height field using UV unrolling mapping sampling method; and generating a three-dimensional sea scene infrared image sequence data based on Unreal Engine rendering.

[0099] Example 2:

[0100] Example 2 is a preferred embodiment of Example 1, and is used to illustrate the present invention in more detail.

[0101] The present invention also provides an infrared sea surface background simulation system with intelligent texture generation. The infrared sea surface background simulation system with intelligent texture generation can be implemented by executing the process steps of the infrared sea surface background simulation method with intelligent texture generation. That is, those skilled in the art can understand the infrared sea surface background simulation method with intelligent texture generation as a preferred embodiment of the infrared sea surface background simulation system with intelligent texture generation.

[0102] An infrared sea surface background simulation system for intelligent texture generation according to the present invention includes:

[0103] Module M1: Extracts intrinsic radiation features from measured sea surface background data;

[0104] Specifically, in module M1:

[0105] A sea surface infrared characteristic inversion model was established to extract intrinsic radiation features from measured sea surface background data; this included data calibration and effect removal, which included atmospheric effect removal and imaging system effect removal.

[0106] Specifically, this includes: using measured infrared image feature data to obtain the true temperature field distribution of the target, and using inversion to obtain the brightness radiation information and temperature distribution information of the sea scene; based on the infrared radiation transmission theory and the photoelectric signal conversion principle, by establishing an infrared illumination model, an atmospheric transmission model, a sensor imaging system effect model and an infrared inversion framework, the measured infrared radiation information of the target is used to back-calculate the target's intrinsic spontaneous infrared radiation distribution information.

[0107] Among them, data calibration is performed by converting and calculating the radiance value based on blackbody radiation calibration and Planck's formula; atmospheric effect removal is performed by calculating the atmospheric transmittance and path radiance value in the shooting band using MODTRAN software in combination with environmental and shooting distance information.

[0108] Based on the radiance value L before reaching the thermal imager ap The intrinsic radiance L of the target is calculated by combining the average atmospheric transmittance and atmospheric path radiance. o To remove imaging system effects, blind pixel compensation and noise suppression are performed on the image. Based on the modulation transfer function theory, the input infrared image is subjected to Fourier transform and inverse filtering transformation, followed by inverse Fourier transform to remove spatial transfer effects. The intrinsic temperature value or radiance value is obtained based on the data calibration results.

[0109] Module M2: Trains the model using intrinsic radiation feature data and expands the sea surface texture;

[0110] Specifically, in module M2:

[0111] An intelligent texture generation model is established. The model is trained using a sample training set formed by intrinsic radiation feature data. After training, the sea surface texture is expanded based on StyleGAN, and the sea surface texture features are obtained based on the expansion results.

[0112] Specifically, the process includes: using a generative adversarial network (GAN) architecture, generating an infrared sea surface background image dataset with a larger number of measured sea surface image samples than a preset standard, thus expanding the data; extracting the infrared sample image dataset as input for infrared texture feature images; using a generative network to encode and decode the infrared texture feature images to generate infrared texture feature images to be judged; using a discriminative network to judge the generated results, thus completing the construction of the GAN structure; defining and selecting a loss function, and using an optimizer to optimize the loss function of the GAN during training, while gradually adjusting the hyperparameters to obtain the trained GAN; and using the trained GAN to generate sea surface infrared texture feature images that meet the conditions.

[0113] Specifically, module M2.1: Constructing the training sample dataset.

[0114] Sea surface temperature data is obtained by inversion, and the pixels of the training set images are cropped to a preset size to generate the actual sample dataset required for training.

[0115] Module M2.2: Intelligent Model Construction and Training

[0116] StyleGAN employs a progressive training method from low to high resolution. StyleGAN consists of a generator and a discriminator.

[0117] First, a generative adversarial network (GAN) model is constructed, and the weights and data of the generator network and the discriminator network within the GAN are initialized.

[0118] Next, define the model loss function and use the Adam optimizer to optimize the loss function;

[0119] Module M2.3: The network model is trained using the backpropagation algorithm. Backpropagation includes:

[0120] Module M2.3.1: Forward propagation, feeds the input into the network, and the network calculates the predicted value; calculates the error between the predicted value and the true label of the input;

[0121] Module M2.3.2: Weight Update. For each parameter in the network that needs to be updated, calculate the partial derivative of the error with respect to that parameter, and update the parameter accordingly.

[0122] Module M2.4: Adjust hyperparameters to bring the network to converge, iterate the backpropagation process until the number of iterations exceeds the set number, and complete the training process;

[0123] Module M2.5: Fixed generator, saves the parameters and network model after training, and generates images that meet preset conditions.

[0124] Module M3: Performs random mapping of infrared sea surface textures and dynamic rendering of sea surface scenes.

[0125] Specifically, in module M3:

[0126] A sea surface infrared radiation model was established. A three-dimensional sea surface texture mapping method was adopted. The brightness distribution calculation results of the sea surface BRDF were modulated by the gray-scale feature modulation factor. The brightness characteristics of the sea surface height field were modulated by the UV unrolling mapping sampling method. The dynamic driving of the three-dimensional scene was realized by the rendering engine. By inputting different sensor parameters and adding atmospheric transmission effects, the sea surface scene simulation under different viewpoints and distances was realized.

[0127] Specifically, this includes: establishing a three-dimensional infrared radiation transmission model of the sea surface considering the shimmering of the sea surface and the radiation characteristics of the sea surface antenna; mapping the extended infrared texture features onto the sea surface radiation model; modulating the brightness distribution calculation results of the sea surface BRDF using grayscale feature modulation factors; modulating the brightness features of the sea surface height field using UV unrolling mapping sampling method; and generating a three-dimensional sea scene infrared image sequence data based on Unreal Engine rendering.

[0128] Example 3:

[0129] Example 3 is a preferred example of Example 1, and is used to illustrate the present invention in more detail.

[0130] This invention introduces intelligent methods to improve the realism of 3D modeling, based on traditional 3D modeling methods. It proposes an infrared sea surface background simulation method with intelligent texture generation. The aim is to extract texture features from measured data and combine this with 3D modeling simulation methods to expand the sample data, meeting the requirements of rich sample size, comprehensive coverage, and high realism. The infrared sea surface background simulation method with intelligent texture generation is characterized by the following steps:

[0131] Step 1: Extract intrinsic radiation features from measured sea surface background data;

[0132] Step 2: Enhance the ocean surface texture based on StyleGAN;

[0133] Step 3: Random and rapid mapping of infrared sea surface texture and dynamic rendering of sea surface scene.

[0134] In step one, a sea surface infrared characteristic inversion model is established to extract intrinsic radiation characteristics from measured sea surface background data; this mainly includes data calibration and effect removal, which includes atmospheric effect removal and imaging system effect removal.

[0135] In step two, an intelligent texture generation model is established, and a small sample training set is formed using intrinsic radiation feature data to train the intelligent model. After training, the sea surface texture can be expanded, and the sea surface texture features can be obtained based on the expansion results.

[0136] In step three, an infrared radiation model of the sea surface is established. A three-dimensional sea surface texture mapping method is adopted. The brightness distribution calculation results of the sea surface BRDF are modulated using the gray-scale feature modulation factor. The brightness characteristics of the sea surface height field are modulated using the UV unrolling mapping sampling method, so that the sea surface radiation feature distribution is closer to the real feature distribution. The rendering engine is used to realize the dynamic driving of the three-dimensional scene. By inputting different sensor parameters and adding atmospheric transmission effects, the simulation of the sea surface scene under different viewpoints and distances is realized.

[0137] This invention proposes an infrared sea surface background simulation method for intelligent texture generation. The principle of the method is as follows: Figure 1 As shown:

[0138] Establish a sea surface infrared characteristic inversion model (1) and extract intrinsic radiation features from measured sea surface background data; establish an intelligent texture generation model (2) and use intrinsic radiation feature data to form a small sample training set to train the intelligent model. After training, sea surface texture can be expanded and sea surface texture features can be obtained based on the expansion results; establish a sea surface infrared radiation model, adopt a three-dimensional sea surface texture mapping method, apply sea surface infrared texture features to the three-dimensional sea surface height field grid, and use a rendering engine to realize the dynamic driving of the three-dimensional scene (3). By inputting different sensor parameters and adding atmospheric transmission effects, sea surface scene simulation under different perspectives and distances can be realized.

[0139] This invention provides a method for simulating an infrared sea surface background using intelligent texture generation, the specific steps of which include:

[0140] Step 1: Extract intrinsic radiation features from measured sea surface background data;

[0141] This technique uses measured infrared image feature data to obtain the true temperature field distribution of a target, and then inverts this data to obtain the brightness radiation information and temperature distribution information of the sea scene. The principle is based on infrared radiative transfer theory and photoelectric signal conversion principles. By establishing an infrared illumination model, an atmospheric transmission model, a sensor imaging system effect model, and an infrared inversion framework, the intrinsic infrared spontaneous radiation distribution information of the target is derived from the measured infrared radiation information. The main steps include data calibration and effect removal. Effect removal includes atmospheric effect removal and imaging system effect removal, as shown in the diagram.

[0142] Data calibration can be performed using blackbody radiation calibration and Planck's formula to calculate radiance values; atmospheric effect removal can be achieved by using MODTRAN software to calculate atmospheric transmittance and path radiance in the shooting band, taking into account environmental factors and shooting distance. The radiance value L before reaching the thermal imager is then used. ap The intrinsic radiance L of the target can be calculated by combining the average atmospheric transmittance and the atmospheric path radiance. o Imaging system effects can be removed by first performing blind pixel compensation and noise suppression on the image, then performing an inverse Fourier transform on the input infrared image based on modulation transfer function theory, and finally performing an inverse Fourier transform to remove spatial transfer effects; finally, the intrinsic temperature value or radiance value can be obtained based on the data calibration results.

[0143] Step 2: Enhance the ocean surface texture based on StyleGAN;

[0144] By employing a generative adversarial network (GAN) architecture, a large dataset of infrared sea surface background images is generated using a small number of measured sea surface image samples, thus expanding the data. The process involves inverting and extracting the infrared sample image dataset as input for infrared texture feature images; using a generative network to encode and decode these infrared texture feature images to generate infrared texture feature images to be judged; using a discriminative network to judge the generated results, thus completing the construction of the GAN structure; defining and selecting an appropriate loss function, and using an optimizer during training to optimize the loss function of the GAN, while progressively adjusting hyperparameters to obtain the trained GAN; and finally, using the trained GAN to generate sea surface infrared texture feature images that meet the specified conditions.

[0145] I. Constructing the training sample dataset

[0146] Sea surface temperature data was obtained through inversion. Since this project requires sea state information, the training set images were uniformly cropped to a size of 128×128 pixels for easier subsequent data processing. The above steps were repeated to generate the measured sample dataset required for training, such as... Figure 2 As shown.

[0147] II. Intelligent Model Construction and Training

[0148] StyleGAN employs a progressive training method, starting from low resolution and gradually increasing it. This not only facilitates the generation of high-resolution images but also helps improve image details. In terms of network structure, StyleGAN follows the GAN concept, with its main components consisting of a generator and a discriminator.

[0149] The training process of the network model is as follows Figure 3 As shown:

[0150] First, a generative adversarial network (GAN) model is constructed, and the weights and data of the generator network and discriminator network within the GAN are initialized.

[0151] Next, we define the model loss function and optimize it using the widely used Adam optimizer to accelerate its descent.

[0152] Third, training the network model aims to optimize the objective function, which requires the use of the backpropagation algorithm. Backpropagation is a common training method for most neural networks and is often used in conjunction with Adam's momentum optimization method. Backpropagation mainly consists of two steps:

[0153] 1) Forward propagation. First, the input is fed into the network, which calculates the predicted value; then, the error between the predicted value and the true label of the input is calculated.

[0154] 2) Weight update. For each parameter in the network that needs to be updated, calculate the partial derivative of the error with respect to that parameter, and use this derivative to update the parameter.

[0155] 4. Adjust hyperparameters to bring the network to convergence. Continuously iterate the backpropagation process until the number of iterations exceeds the set number, completing the training process.

[0156] 5. Fixed Generator. Save the trained parameters and network model, which can then be used to generate images that meet the specified conditions.

[0157] Step 3: Random and rapid mapping of infrared sea surface textures and dynamic rendering of sea surface scenes;

[0158] First, a three-dimensional infrared radiation transmission model of the sea surface is established, considering the shimmering of sea surface scales and the radiation characteristics of the sea-line antenna. Second, extended infrared texture features are mapped onto the sea surface radiation model. The brightness distribution calculation results of the sea surface BRDF are modulated using a grayscale feature modulation factor, and the brightness features of the sea surface height field are modulated using a UV unrolling mapping sampling method, making the sea surface radiation feature distribution closer to the real feature distribution. Finally, highly realistic three-dimensional sea scene infrared image sequence data is generated using Unreal Engine rendering. The infrared sea surface sunlight bright band image generated by this method is shown below. Figure 4 As shown, the non-bright bands on the sea surface are as follows: Figure 4 As shown.

[0159] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0160] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for simulating infrared sea surface backgrounds using intelligent texture generation, characterized in that, include: Step S1: Extract intrinsic radiation features from measured sea surface background data; Step S2: Train the model using intrinsic radiation feature data to expand the sea surface texture; Step S3: Perform random mapping of infrared sea surface texture and dynamic rendering of sea surface scene; In step S1: Establish a sea surface infrared characteristic inversion model to extract intrinsic radiation features from measured sea surface background data; This includes data calibration and effect removal, with effect removal encompassing atmospheric effect removal and imaging system effect removal. Specifically, this includes: using measured infrared image feature data to obtain the true temperature field distribution of the target, and using inversion to obtain the brightness radiation information and temperature distribution information of the sea scene; based on the infrared radiation transmission theory and the photoelectric signal conversion principle, by establishing an infrared illumination model, an atmospheric transmission model, a sensor imaging system effect model and an infrared inversion framework, the measured infrared radiation information of the target is used to back-calculate the target's intrinsic spontaneous infrared radiation distribution information. Among them, data calibration is performed by converting and calculating the radiance value based on blackbody radiation calibration and Planck's formula; atmospheric effect removal is performed by calculating the atmospheric transmittance and path radiance value in the shooting band using MODTRAN software in combination with environmental and shooting distance information. Based on the radiance value before reaching the thermal imager The intrinsic radiance of the target is calculated by combining the average atmospheric transmittance and atmospheric path radiance. To remove imaging system effects, blind pixel compensation and noise suppression are performed on the image. Then, based on the modulation transfer function theory, Fourier transform is performed on the input infrared image for inverse filtering and Fourier transform to remove spatial transfer effects. The intrinsic temperature or radiance value is obtained based on the data calibration results. In step S2: An intelligent texture generation model is established. The model is trained using a sample training set formed by intrinsic radiation feature data. After training, the sea surface texture is expanded based on StyleGAN, and the sea surface texture features are obtained based on the expansion results. Specifically, the process includes: using a generative adversarial network (GAN) architecture, generating an infrared sea surface background image dataset with a larger number of measured sea surface image samples than a preset standard, thus expanding the data; extracting the infrared sample image dataset as input for infrared texture feature images; using a generative network to encode and decode the infrared texture feature images to generate infrared texture feature images to be judged; using a discriminative network to judge the generated results, thus completing the construction of the GAN structure; defining and selecting a loss function, and using an optimizer to optimize the loss function of the GAN during training, while gradually adjusting the hyperparameters to obtain the trained GAN; and using the trained GAN to generate sea surface infrared texture feature images that meet the conditions. In step S3: A sea surface infrared radiation model was established. A three-dimensional sea surface texture mapping method was adopted. The brightness distribution calculation results of the sea surface BRDF were modulated by the gray-scale feature modulation factor. The brightness characteristics of the sea surface height field were modulated by the UV unrolling mapping sampling method. The dynamic driving of the three-dimensional scene was realized by the rendering engine. By inputting different sensor parameters and adding atmospheric transmission effects, the sea surface scene simulation under different viewpoints and distances was realized. Specifically, this includes: establishing a three-dimensional infrared radiation transmission model of the sea surface considering the shimmering of the sea surface and the radiation characteristics of the sea surface antenna; mapping the extended infrared texture features onto the sea surface radiation model; modulating the brightness distribution calculation results of the sea surface BRDF using grayscale feature modulation factors; modulating the brightness features of the sea surface height field using UV unrolling mapping sampling method; and generating a three-dimensional sea scene infrared image sequence data based on Unreal Engine rendering.

2. The infrared sea surface background simulation method for intelligent texture generation according to claim 1, characterized in that: Step S2.1: Construct the training sample dataset Sea surface temperature data is obtained by inversion, and the pixels of the training set images are cropped to a preset size to generate the actual sample dataset required for training. Step S2.2: Intelligent Model Construction and Training StyleGAN employs a progressive training method from low to high resolution. StyleGAN consists of a generator and a discriminator. First, a generative adversarial network (GAN) model is constructed, and the weights and data of the generator network and the discriminator network within the GAN are initialized. Next, define the model loss function and use the Adam optimizer to optimize the loss function; Step S2.3: Train the network model using the backpropagation algorithm. Backpropagation includes: Step S2.3.1: Forward propagation, the input is fed into the network, and the network calculates the predicted value; the error between the predicted value and the true label of the input is calculated; Step S2.3.2: Weight update. For each parameter in the network that needs to be updated, calculate the partial derivative of the error with respect to that parameter, and update the parameter accordingly. Step S2.4: Adjust the hyperparameters to make the network converge, iterate the backpropagation process until the number of iterations exceeds the set number, and complete the training process; Step S2.5: Fix the generator, save the parameters and network model after training, and generate images that meet the preset conditions.

3. An infrared sea surface background simulation system with intelligent texture generation, characterized in that, include: Module M1: Extracts intrinsic radiation features from measured sea surface background data; Module M2: Trains the model using intrinsic radiation feature data and expands the sea surface texture; Module M3: Performs random mapping of infrared sea surface textures and dynamic rendering of sea surface scenes; In module M1: Establish a sea surface infrared characteristic inversion model to extract intrinsic radiation features from measured sea surface background data; This includes data calibration and effect removal, with effect removal encompassing atmospheric effect removal and imaging system effect removal. Specifically, this includes: using measured infrared image feature data to obtain the true temperature field distribution of the target, and using inversion to obtain the brightness radiation information and temperature distribution information of the sea scene; based on the infrared radiation transmission theory and the photoelectric signal conversion principle, by establishing an infrared illumination model, an atmospheric transmission model, a sensor imaging system effect model and an infrared inversion framework, the measured infrared radiation information of the target is used to back-calculate the target's intrinsic spontaneous infrared radiation distribution information. Among them, data calibration is performed by converting and calculating the radiance value based on blackbody radiation calibration and Planck's formula; atmospheric effect removal is performed by calculating the atmospheric transmittance and path radiance value in the shooting band using MODTRAN software in combination with environmental and shooting distance information. Based on the radiance value before reaching the thermal imager The intrinsic radiance of the target is calculated by combining the average atmospheric transmittance and atmospheric path radiance. To remove imaging system effects, blind pixel compensation and noise suppression are performed on the image. Then, based on the modulation transfer function theory, Fourier transform is performed on the input infrared image for inverse filtering and Fourier transform to remove spatial transfer effects. The intrinsic temperature or radiance value is obtained based on the data calibration results. In module M2: An intelligent texture generation model is established. The model is trained using a sample training set formed by intrinsic radiation feature data. After training, the sea surface texture is expanded based on StyleGAN, and the sea surface texture features are obtained based on the expansion results. Specifically, the process includes: using a generative adversarial network (GAN) architecture, generating an infrared sea surface background image dataset with a larger number of measured sea surface image samples than a preset standard, thus expanding the data; extracting the infrared sample image dataset as input for infrared texture feature images; using a generative network to encode and decode the infrared texture feature images to generate infrared texture feature images to be judged; using a discriminative network to judge the generated results, thus completing the construction of the GAN structure; defining and selecting a loss function, and using an optimizer to optimize the loss function of the GAN during training, while gradually adjusting the hyperparameters to obtain the trained GAN; and using the trained GAN to generate sea surface infrared texture feature images that meet the conditions. In module M3: A sea surface infrared radiation model was established. A three-dimensional sea surface texture mapping method was adopted. The brightness distribution calculation results of the sea surface BRDF were modulated by the gray-scale feature modulation factor. The brightness characteristics of the sea surface height field were modulated by the UV unrolling mapping sampling method. The dynamic driving of the three-dimensional scene was realized by the rendering engine. By inputting different sensor parameters and adding atmospheric transmission effects, the sea surface scene simulation under different viewpoints and distances was realized. Specifically, this includes: establishing a three-dimensional infrared radiation transmission model of the sea surface considering the shimmering of the sea surface and the radiation characteristics of the sea surface antenna; mapping the extended infrared texture features onto the sea surface radiation model; modulating the brightness distribution calculation results of the sea surface BRDF using grayscale feature modulation factors; modulating the brightness features of the sea surface height field using UV unrolling mapping sampling method; and generating a three-dimensional sea scene infrared image sequence data based on Unreal Engine rendering.

4. The infrared sea surface background simulation system with intelligent texture generation according to claim 3, characterized in that: Module M2.1: Constructing the training sample dataset Sea surface temperature data is obtained by inversion, and the pixels of the training set images are cropped to a preset size to generate the actual sample dataset required for training. Module M2.2: Intelligent Model Construction and Training StyleGAN employs a progressive training method from low to high resolution. StyleGAN consists of a generator and a discriminator. First, a generative adversarial network (GAN) model is constructed, and the weights and data of the generator network and the discriminator network within the GAN are initialized. Next, define the model loss function and use the Adam optimizer to optimize the loss function; Module M2.3: The network model is trained using the backpropagation algorithm. Backpropagation includes: Module M2.3.1: Forward propagation, which feeds the input into the network and calculates the predicted value through the network; Calculate the error between the predicted value and the true label of the input; Module M2.3.2: Weight Update. For each parameter in the network that needs to be updated, calculate the partial derivative of the error with respect to that parameter, and update the parameter accordingly. Module M2.4: Adjust hyperparameters to bring the network to converge, iterate the backpropagation process until the number of iterations exceeds the set number, and complete the training process; Module M2.5: Fixed generator, saves the parameters and network model after training, and generates images that meet preset conditions.

Images