Infrared simulation method and system based on neural network and GPU parallel computing

Abstract

The present invention discloses an infrared simulation method and system based on neural network and GPU parallel computing, belonging to the field of infrared simulation. The atmospheric transmittance and the radiation value received by the ground surface in all weather conditions are fitted by using an atmospheric model; the infrared simulation data is preprocessed, and the patch occlusion probability value and the patch temperature value in the simulation scene are calculated by using a distributed GPU parallel framework; wherein, the number of sampling points is dynamically adjusted by combining the patch geometric area and complexity through an adaptive patch sampling method to calculate the patch occlusion probability value, and the patch occlusion probability value is introduced when calculating the patch temperature value; the patch temperature value is optimized to obtain the patch steady-state temperature value, and the radiation brightness of the specified band is calculated according to the patch steady-state temperature value, and finally, the infrared simulation image of the specified viewing angle under the specified band is rendered. The present invention overcomes the problems of complex calculation of radiation value and temperature field, low efficiency of occlusion calculation, insufficient heat conduction simulation capability, and insufficient precision of adaptive sampling in the prior art.

Description

Technical Field

[0001] The invention belongs to the field of infrared simulation, and in particular relates to an infrared simulation method and system based on neural network and GPU parallel computing. Background Art

[0002] As an important simulation technology, infrared simulation can predict the infrared characteristics of objects without directly relying on physical testing, and has been widely used in security monitoring, virtual reality and other fields. However, the existing technology still has the following problems in large-scale infrared simulation:

[0003] 1. The calculation complexity of radiation value and temperature field is high:

[0004] In existing methods, direct solar radiation, scattered solar radiation and atmospheric long-wave radiation are usually obtained through offline calculation or static table. These methods have low calculation efficiency and cannot adapt to the rapid changes of complex lighting conditions in dynamic scenes.

[0005] 2. Low efficiency of occlusion calculation:

[0006] Large scenes involve the calculation of a large number of occlusion relationships. Existing methods of occlusion calculation based on ray tracing rely on single-threaded or limited parallel technology, which makes it too time-consuming to solve the occlusion relationship and it is difficult to achieve occlusion calculation with a high sampling rate.

[0007] 3. Insufficient efficiency of heat conduction simulation:

[0008] Solving the two-dimensional heat conduction equation is the core of temperature field simulation. When performing calculations on large-scale discrete grids, existing methods have low computational efficiency because they fail to fully utilize the parallel computing capabilities of GPUs.

[0009] 4. Adaptive sampling is not fine enough:

[0010] For large areas in complex scenes, existing methods lack efficient adaptive sampling strategies, resulting in suboptimal distribution of sampling points and allocation of computing resources, thus affecting simulation accuracy and efficiency.

[0011] Therefore, current technology urgently needs a technology that can efficiently complete infrared simulation in complex and large scenes to solve the problems of low computational efficiency and poor dynamic adaptability. Summary of the invention

[0012] The purpose of the present invention is to overcome the problems in the prior art such as complex calculation of radiation values ​​and temperature fields, low efficiency of occlusion calculation, insufficient heat conduction simulation capability, and insufficient precision of adaptive sampling, and to propose an infrared simulation method and system based on neural network and GPU parallel computing, which has the effects of high efficiency and strong dynamic adaptability.

[0013] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0014] In a first aspect, the present invention proposes an infrared simulation method based on neural network and GPU parallel computing, comprising:

[0015] The atmospheric model is used to perform nonlinear mapping of spatiotemporal parameters and environmental parameters, and the all-weather atmospheric transmittance and the radiation value received by the surface are fitted;

[0016] Preprocess infrared simulation data and use the distributed GPU parallel framework to calculate the patch occlusion probability value and patch temperature value in the simulation scene;

[0017] When calculating the patch occlusion probability value, the number of sampling points is dynamically adjusted according to the patch geometric area and complexity and the positions of the sampling points on the patch are determined by an adaptive patch sampling method, and the patch occlusion probability value is obtained by solving the occlusion relationship of each sampling point on the patch;

[0018] Introducing a facet occlusion probability value when calculating the facet temperature value;

[0019] The patch temperature value is optimized to obtain the patch steady-state temperature value, the radiant brightness of a specified band is calculated according to the patch steady-state temperature value, and finally an infrared simulation image of a specified viewing angle in a specified band is rendered.

[0020] Furthermore, the atmosphere model is a residual neural network; the training process of the atmosphere model includes:

[0021] The simulated atmospheric data is sampled to obtain data that conforms to the distribution of real data as the main training data, and the real atmospheric data is collected as the fine-tuning training data; the main training data and the fine-tuning training data both contain spatiotemporal parameters, environmental parameters, atmospheric transmittance, and the radiation value received by the surface; the spatiotemporal parameters include latitude, longitude, year, day, hour, and minute, and the environmental parameters include air temperature, humidity, air pressure, visibility, weather pattern, cloud height, and cloud thickness;

[0022] First, the main training data is used to pre-train the atmosphere model, and then the fine-tuning training data is used to fine-tune the pre-trained atmosphere model.

[0023] Furthermore, the radiation values ​​received by the surface include direct solar radiation values, scattered solar radiation values, and atmospheric long-wave radiation values.

[0024] Furthermore, three independent CPU nodes are used to complete the infrared simulation data preprocessing process in parallel, including:

[0025] CPU node 1: Calculates the depth map and patch data with material information based on the input material segmentation map, material database and model file;

[0026] CPU node 2: used to splice the all-weather atmospheric transmittance predicted by the atmospheric model and the radiation value received on the ground surface with the spatiotemporal parameters and environmental parameters at the corresponding time points to obtain multi-source fusion data;

[0027] CPU node 3: Calculate the all-weather solar azimuth according to the input time and space parameters to obtain the direct direction of the sun.

[0028] Furthermore, the adaptive patch sampling method is specifically as follows:

[0029] Determine the number of basic sampling points of the patch according to the density of the benchmark sampling points;

[0030] Multiply the number of basic sampling points by the complexity weight based on the depth map gradient to obtain the number of sampling points after complexity adjustment;

[0031] The number of sampling points after complexity adjustment is multiplied by the occlusion probability weight to obtain the number of sampling points after occlusion adjustment.

[0032] Determine whether the number of sampling points after occlusion adjustment is less than the preset minimum number of sampling points of the patch. If so, use the preset minimum number of sampling points of the patch as the final total number of sampling points; otherwise, use the number of sampling points after occlusion adjustment as the final total number of sampling points.

[0033] Furthermore, based on the final number of sampling points, the Poisson disk distribution method is used to determine the position of the sampling points on the patch. According to the direct direction of the sun and the GPU ray tracing algorithm, a bounding volume hierarchy is constructed to solve the occlusion relationship of each sampling point. The occlusion probability value corresponding to each patch at each moment is calculated based on the occlusion relationship of the sampling points. The occlusion probability value is the ratio of the occluded sampling points on the patch to the final total number of sampling points.

[0034] Furthermore, the calculation of the dough temperature value includes:

[0035] The energy value of the patch at each moment is calculated through the energy equation. The total solar radiation used to calculate the energy value is defined as the patch shading probability value multiplied by the direct solar radiation value received by the patch plus the solar diffuse radiation value received by the patch.

[0036] The longitudinal heat conduction is calculated based on the energy value to obtain the dough temperature value.

[0037] Furthermore, based on the surface temperature value, the multi-grid method is used to calculate the two-dimensional heat conduction to obtain the surface steady-state temperature value; based on the surface steady-state temperature value, the Planck formula is used to calculate the radiation brightness in the specified band.

[0038] Furthermore, the distributed GPU parallel framework refers to transmitting the results of infrared simulation data preprocessing to multiple GPU nodes after data segmentation to parallelly calculate the patch occlusion probability value, patch temperature value, patch steady-state temperature value and radiation brightness of the specified band.

[0039] In a second aspect, the present invention proposes an infrared simulation system based on neural network and GPU parallel computing, which is used to implement the above-mentioned infrared simulation method.

[0040] The beneficial effects of the present invention are:

[0041] (1) The present invention uses an atmospheric model to perform nonlinear mapping of spatiotemporal parameters and environmental parameters, and fits the all-weather atmospheric transmittance and the radiation value received by the surface, replacing the offline calculation and traditional table lookup method, thereby improving the calculation efficiency and accuracy of the radiation value;

[0042] (2) The present invention combines the adaptive patch sampling method and GPU ray tracing technology, and uses the distributed GPU parallel framework to calculate the patch occlusion probability value, patch temperature value, patch steady-state temperature value and specified band radiance in the simulation scene, thereby achieving efficient solution of occlusion relationship and greatly shortening the occlusion calculation time; by further optimizing the patch temperature value, the problem of insufficient heat conduction simulation is solved;

[0043] (3) The adaptive patch sampling method proposed in the present invention dynamically adjusts the number of sampling points according to the patch geometric area and complexity, optimizes the sampling point distribution and resource allocation, realizes efficient simulation calculation, and has more refined sampling;

[0044] (4) The present invention significantly improves the utilization rate of computing resources and can efficiently complete infrared simulation of complex dynamic scenes, and is suitable for security monitoring, virtual reality and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a schematic diagram of atmospheric model training.

[0046] Figure 2 is the simulation data preprocessing flow chart, where Figure 2 (a) is a schematic diagram of calculating the depth map and the patch data with material information; Figure 2 (b) is a schematic diagram of calculating multi-source fusion data; Figure 2 (c) is a schematic diagram for calculating the direct direction of the sun.

[0047] Figure 3 This is the flow chart of GPU parallel calculation of scene temperature field and radiation field, where Figure 3 (a) is a schematic diagram of calculating the probability value of patch occlusion; Figure 3 (b) is a schematic diagram of calculating the temperature value of the dough sheet; Figure 3 (c) is a schematic diagram of calculating the steady-state temperature value of the surface; Figure 3 (d) in the figure is a schematic diagram for calculating the radiation brightness in a specified band.

[0048] Figure 4 It is the infrared image rendering flow chart.

[0049] Figure 5 It is an overall flow chart of the infrared simulation method proposed by the present invention.

[0050] Figure 6 It is the temperature field distribution map of the scene.

[0051] Figure 7 This is the mid-wave infrared (3-5 micron) simulation result of the scene.

[0052] Figure 8 This is the long-wave infrared (8-14 micron) simulation result of the scene. DETAILED DESCRIPTION

[0053] The present invention is further described and illustrated below in conjunction with specific embodiments. The embodiments are merely exemplary of the present disclosure and do not define the scope of limitation. The technical features of each embodiment of the present invention may be combined accordingly without conflicting with each other.

[0054] The accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0055] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the steps. For example, some steps may be decomposed, while some steps may be combined or partially combined, so the actual execution order may change according to the actual situation.

[0056] The large-scene infrared simulation method based on neural network and GPU parallelism proposed in the present invention is as follows: Figure 5 As shown in the figure, the main implementation process is as follows: for the input spatiotemporal parameters and environmental parameters, an atmospheric model based on a residual neural network is used to generate all-weather atmospheric transmittance and the radiation value received by the surface. After further preprocessing the simulation data, the temperature field is calculated in parallel using a distributed GPU parallel framework. Based on the temperature field distribution, infrared feature modeling and simulation are performed, and finally, the sensor effect is used to simulate the infrared simulation image.

[0057] More specifically, firstly, the atmospheric model based on residual neural network is used to perform nonlinear mapping of spatiotemporal parameters and environmental parameters, and the all-weather atmospheric transmittance and the radiation value received by the surface are quickly fitted, and the radiation value received by the surface includes direct solar radiation, solar scattered radiation, and atmospheric long-wave radiation value; after pre-processing the simulation data, the distributed GPU parallel framework is used to accelerate the calculation of the patch occlusion probability value, patch temperature value, patch steady-state temperature value, and radiation brightness of the specified band in the scene through CUDA technology; wherein, the adaptive patch sampling method is used to dynamically adjust the patch according to its geometric area and complexity. The number of sampling points is adjusted and their positions on the patch are determined. The bounding volume hierarchy (BVH) structure is constructed in combination with the GPU ray tracing algorithm to solve the occlusion relationship of each sampling point and obtain the patch occlusion probability value. The patch occlusion probability value is introduced into the temperature calculation method to calculate the patch temperature value. On the basis of the patch temperature value, the two-dimensional heat conduction simulation based on the multi-grid algorithm is used to calculate the patch steady-state temperature value. The Planck formula is then used to calculate the radiation brightness of the specified band. Finally, the rendering equation with the self-emission term and the sensor simulation module are combined to render the infrared simulation image of the specified viewing angle under the specified band.

[0058] The above technology is described in detail below.

[0059] (1) Atmospheric model training

[0060] The present invention uses a neural network model based on a residual structure as an atmospheric model, and can quickly fit all-weather direct solar radiation, solar scattered radiation, atmospheric long-wave radiation value, and atmospheric transmittance through nonlinear mapping of spatiotemporal parameters and environmental parameters. Figure 1 As shown in the figure, the process of predicting radiation eigenvalues ​​through feature engineering and neural networks from both simulated data and real data is demonstrated.

[0061] In a specific implementation of the present invention, a large amount of atmospheric data simulation is first performed by using Modtran software, and data sampling is used to obtain data distribution that conforms to reality as main training data. The main training data is then used to pre-train the atmospheric model, and the pre-trained atmospheric model is further fine-tuned using real data.

[0062] During the training process, spatiotemporal parameters and environmental parameters are used as input, where the spatiotemporal parameters include latitude, longitude, year, day, hour, and minute, and the environmental parameters include air temperature, humidity, air pressure, visibility, weather pattern, cloud height, and cloud thickness; the input parameters are preprocessed, that is, feature engineering is applied to first convert the time parameters (year, day, hour, minute) in the input spatiotemporal parameters using sin function or cos function for periodic characteristic conversion; then the correlation information between the spatiotemporal parameters and environmental parameters is extracted, and strong correlation information is screened out from the correlation information to improve the network training efficiency and prediction accuracy; the spatiotemporal and environmental parameters and strong correlation information after the periodic characteristic conversion are used as the input of the atmospheric model to enhance the representation ability of the input data; the atmospheric model predicts the direct solar radiation, solar scattered radiation, atmospheric long-wave radiation values, and atmospheric transmittance at all time points of the day.

[0063] In this embodiment, polynomial expansion and interactive feature extraction methods are used to obtain correlation information between parameters, and a recursive feature elimination (RFECV) method is used to select the strongest correlation from the correlation information as strong correlation information. Other means known in the art may also be used to achieve this.

[0064] During the pre-training and fine-tuning stage of the atmospheric model, the simulated or real direct solar radiation, scattered solar radiation, atmospheric long-wave radiation value, and atmospheric transmittance of the training data are used as labels to optimize the model parameters until convergence. The neural network model based on the residual structure is composed of a fully connected layer and a residual block. Batch normalization and Dropout techniques are introduced to improve the training stability and generalization ability of the model. The above can be achieved according to known technologies and will not be repeated here. The present invention significantly improves the efficiency and accuracy of the calculation of direct solar radiation, scattered radiation, and atmospheric radiation values ​​by replacing offline calculations and traditional table lookup methods with optimized deep neural networks. (2) Simulation data preprocessing

[0065] Obtain the data that needs to be simulated, including spatiotemporal parameters, environmental parameters, material segmentation maps, material databases and model files, and preprocess the simulation data to obtain depth maps, patch data with material information, multivariate fusion data and direct sunlight direction.

[0066] like Figure 2 As shown, this embodiment uses three independent CPU nodes to complete the above tasks in parallel:

[0067] CPU node 1 calculates the depth map and the patch data with material information based on the input material segmentation map, material database and model file; Figure 2As shown in (a), by loading the model file, extracting the depth information in the model file, and generating a depth map; the loaded model is composed of patches, and according to the loaded model, material segmentation map and material database, the patch data with material information is obtained through patch material mapping. Here, patch material mapping can be implemented according to the traditional material matching algorithm to obtain the material sequence corresponding to each patch.

[0068] CPU node 2 Figure 2 As shown in (b), the direct solar radiation, scattered solar radiation, atmospheric long-wave radiation values ​​and atmospheric transmittance predicted by the atmospheric model at all time points of the day are spliced ​​with the spatiotemporal parameters and environmental parameters at the corresponding time points to obtain multi-source fusion data; here, running the atmospheric model can be completed in the GPU to speed up the processing process.

[0069] CPU node 3 Figure 2 As shown in (c) in the figure, the all-weather solar azimuth is calculated based on the input time and space parameters to obtain the direct direction of the sun.

[0070] The above three CPU nodes are used for parallel processing to accelerate the calculation time. The depth map, patch data with material information, multi-element fusion data and direct sunlight direction data obtained after preprocessing are evenly distributed to each computing node through the MPI interface using a data segmentation algorithm, and transmitted from the CPU to the GPU to ensure load balancing. Subsequently, the distributed GPU parallel framework will be used to accelerate the calculation of patch occlusion probability values, patch temperature values, patch steady-state temperature values, and radiation brightness in the specified band in the scene through CUDA technology.

[0071] (3) Calculation of patch occlusion probability value

[0072] like Figure 3 As shown in (a) in the figure, the present invention uses an adaptive patch sampling method to dynamically adjust the number of sampling points and determine the positions of the sampling points on the patch according to the geometric area and complexity of the patch, and constructs a bounding volume hierarchy (BVH) structure based on the direct direction of the sun in combination with the GPU ray tracing algorithm to solve the occlusion relationship of each sampling point, thereby obtaining the occlusion probability value of each patch.

[0073] In a specific implementation of the present invention, the process of dynamically adjusting the number of sampling points according to the geometric area and complexity of the patch is as follows:

[0074] Assume the geometric area of ​​the patch is , the number of sampling points for the patch Positively correlated with the area of ​​the patch. Set the density of the benchmark sampling points according to the size of the area. (number of sampling points per unit area), then the number of basic sampling points of the patch for:

[0075]

[0076] The complexity of a patch is measured by its depth map gradient, which reflects the surface undulation and occlusion sensitivity of the patch. The complexity weight based on the depth map gradient is defined as follows:

[0077]

[0078] in, The coordinates of the patch in the depth map The depth value at and The depth values ​​are The rate of change in direction; the denominator of the entire formula is the maximum depth gradient integral value of the entire scene patch, which is used for normalization.

[0079] According to the complexity weight, calculate the number of sampling points after complexity adjustment for:

[0080]

[0081] Calculate the occlusion probability weight of the patch for:

[0082]

[0083] in, is the occlusion probability weight adjustment factor, represents the occlusion probability of patch i based on the depth map, and represents the light occlusion sensitivity of the patch; The number of sampling points that are judged to be less than the center point depth (i.e., occluded) when performing hemispherical sampling at the center of the patch according to the depth map; Indicates the total number of sampling points for hemisphere sampling at the center of the patch according to the depth map.

[0084] According to the occlusion probability weight, calculate the number of sampling points after occlusion adjustment for:

[0085]

[0086] Above The geometric area, depth gradient complexity and occlusion probability are combined. In order to ensure the stability of the calculation, the minimum number of sampling points of the patch is set , determine the final number of sampling points based on the number of sampling points after occlusion adjustment and the minimum number of sampling points for:

[0087]

[0088] The Poisson Disk Sampling method is used to determine the position of the sampling points on the patch to ensure the uniformity of the sampling points on the patch. The main steps are: 1. Generate random points in the two-dimensional projection of the patch; 2. Filter the minimum distance between adjacent points based on the final number of sampling points calculated to ensure the uniform distribution between adjacent sampling points; 3. Project the filtered points back to the three-dimensional space and map them to the patch surface. This solution effectively improves the sampling efficiency based on occlusion calculation.

[0089] According to the direct sunlight direction, the bounding volume hierarchy (BVH) structure is constructed in combination with the GPU ray tracing algorithm to solve the occlusion relationship of each sampling point. According to the occlusion relationship of the sampling points, the occlusion probability value corresponding to each patch at each moment is calculated. The occlusion probability value is the ratio of the occluded sampling points on the patch to the total number of sampling points. The occlusion probability value calculated here includes the occlusion probability values ​​of each observation direction under each direct sunlight direction.

[0090] (4) Calculation of dough temperature

[0091] like Figure 3 As shown in (b), the patch temperature value is calculated based on the multi-source fusion data, the patch occlusion probability value and the patch data with material information.

[0092] First, the energy value of the patch at each moment is calculated by the energy equation:

[0093]

[0094]

[0095] in, is the surface layer of the dough (assuming the thickness is ) average absorbed power; the terms on the right side of the equation are solar radiation, atmospheric long-wave radiation, thermal radiation from the test object surface, convection term, thermal conduction term, surrounding scene energy term and latent heat term, and are the absorption coefficients of the patch surface to solar radiation and atmospheric long-wave radiation, respectively.

[0096] From the above formula, we can see that It is obtained by multiplying the patch occlusion probability value by the direct solar radiation value received by the surface and adding the received solar scattered radiation value. The calculation accuracy is improved by introducing the patch occlusion probability.

[0097] The longitudinal heat conduction is calculated according to the energy value. In this embodiment, the longitudinal heat conduction is calculated by using a one-dimensional difference method:

[0098]

[0099] In the formula, is the thermal diffusion coefficient of the patch material, for Always in vertical position The temperature at is the unit time interval, is the unit vertical position change.

[0100] The data at all time points of the day are continuously iterated and calculated until a steady state is reached, and the final dough temperature value at each moment is obtained.

[0101] (5) Calculation of steady-state temperature of dough

[0102] like Figure 3 As shown in (c), based on the calculated temperature value of each surface patch, the multi-grid method is used to calculate the two-dimensional heat conduction to obtain the steady-state temperature value of the surface patch.

[0103] The multigrid method described accelerates the solution process by constructing a series of grids with different resolutions (from fine to coarse). Each grid level corresponds to a different discretization accuracy, with coarse grids used to quickly eliminate low-frequency errors and fine grids used to fine-tune high-frequency errors. At each grid level, the Gauss-Seidel iteration method is used to "smooth" the errors. The smoothing process mainly eliminates high-frequency errors because these errors are more significant on the fine grid. The residual (error) on the fine grid is constrained to the coarser grid. This step converts the high-frequency errors into low-frequency errors on the coarse grid, facilitating effective error correction on the coarse grid. The constrained error equation is solved on the coarse grid, and a small number of iterations are used to obtain an approximation of the error on the coarse grid. Since the problem size on the coarse grid is smaller, the computational efficiency is higher. The error approximation on the coarse grid is extended back to the fine grid through interpolation methods and used to correct the solution on the fine grid. This step ensures that the low-frequency errors on the fine grid are also effectively corrected.

[0104] Figure 6 This is a temperature field distribution diagram of a certain scene calculated by this embodiment. The present invention uses a multi-grid algorithm and CUDA parallel computing to improve the efficiency of heat conduction simulation to several times that of existing methods, and dynamically adjusts the grid resolution to improve the simulation accuracy.

[0105] (6) Calculation of radiant brightness in a specified band

[0106] like Figure 3 As shown in (d) in the figure, based on the calculated steady-state temperature value of the patch, the radiation brightness of the specified band is obtained according to the Planck formula:

[0107]

[0108] in, Indicates that the black body has a frequency and absolute temperature The radiation intensity is expressed in power per unit area, per unit solid angle, and per unit frequency; is the frequency of electromagnetic waves, measured in Hertz (Hz); is the absolute temperature of a black body in Kelvin (K); is Planck's constant, which is approximately Joule·second (J·s); is the speed of light, about Meters per second (m / s); is the Boltzmann constant, which is approximately Joule per Kelvin (J / K); is the base of the natural exponential function, which is approximately 2.718. This formula is used to describe the radiation intensity distribution of a black body at different frequencies and temperatures, revealing the quantized nature of thermal radiation. For the radiation energy in a certain wavelength range, it can be calculated by integration:

[0109]

[0110] in, Expressed as a black body in the wavelength range The total radiation energy, and is the wavelength range of the integration, in meters (m).

[0111] The above calculations of the patch occlusion probability value, patch temperature value, patch steady-state temperature value, and radiant brightness of a specified band in the scene adopt a distributed GPU parallel framework, and the calculation process is accelerated by CUDA technology.

[0112] (7) Rendering

[0113] like Figure 4 As shown, according to the radiance of the specified band, the patch data with material information, the occlusion rate of the patch in the specified direction, the infrared material database and the atmospheric transmittance, combined with the rendering equation with the self-emission term and the sensor simulation module, the infrared simulation image of the specified viewing angle under the specified band is rendered.

[0114] In this embodiment, the self-emission term is added to the ray tracing rendering equation to obtain the following formula to calculate the irradiance of the unobstructed position of the surface in the incident direction in the specified direction:

[0115]

[0116] in, is the emissivity of the patch material, For the temperature Next specified band The radiance of Describes the infrared bidirectional reflectance distribution function, indicating the direction Incident to position After that, in the direction The outgoing energy distribution, Represents the position variable on the patch, represents the incident direction, Indicates the emission direction, Indicates the position of the unobstructed surface in the specified direction In the incident direction The irradiance, represents the patch normal, represents integration over the hemisphere.

[0117] Introduce the occlusion rate of the specified direction of the patch to obtain the radiation brightness at the final observation point:

[0118]

[0119] in, for The occlusion probability value of the direction.

[0120] The irradiance at the observation point is introduced into the sensor effect simulation for post-processing to obtain the final infrared simulation image. The present invention introduces a rendering equation with a self-emission term to render the infrared image of a specified viewing angle, which effectively improves the simulation accuracy.

[0121] like Figure 7 Shown is the mid-wave infrared (3-5 micron) simulation result for the scene. Figure 8 The long-wave infrared (8-14 micron) simulation results of the scene are shown. As can be seen from the image, both medium-wave and long-wave infrared images well present the thermal radiation distribution of buildings, roads and water bodies in the area, especially emphasizing different thermal features in different bands. The color mapping of the two images shows the radiation intensity gradient (in W / m²) in different areas. Especially in the long-wave infrared image, the brightness distribution of high-radiation areas is very significant, which can well distinguish high and low radiation areas. These two simulated images are excellent in accurately presenting thermal radiation distribution and regional characteristics, providing an intuitive basis for thermal analysis and infrared detection in complex scenes.

[0122] Based on the same inventive concept, an infrared simulation system based on neural network and GPU parallel computing is also provided in this embodiment, including:

[0123] The atmospheric model module is used to perform nonlinear mapping of spatiotemporal parameters and environmental parameters, and fit the all-weather atmospheric transmittance and the radiation value received by the surface;

[0124] A preprocessing module, which is used to preprocess infrared simulation data;

[0125] A distributed GPU computing module is used to calculate the patch occlusion probability value and the patch temperature value in the simulation scene using a distributed GPU parallel framework, optimize the patch temperature value to obtain the patch steady-state temperature value, and calculate the radiant brightness of a specified band according to the patch steady-state temperature value;

[0126] Wherein, when calculating the patch occlusion probability value, the number of sampling points is dynamically adjusted according to the patch geometric area and complexity and the positions of the sampling points on the patch are determined by an adaptive patch sampling method, and the patch occlusion probability value is obtained by solving the occlusion relationship of each sampling point on the patch; the patch occlusion probability value is introduced when calculating the patch temperature value;

[0127] The rendering module is used to render an infrared simulation image of a specified viewing angle in a specified band according to the calculation results of the preprocessing module and the calculation results of the distributed GPU calculation module.

[0128] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment, and the implementation methods of the remaining modules will not be repeated here. The system embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0129] The embodiments of the system of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The system embodiments can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, the corresponding computer program instructions in the non-volatile memory are read into the memory by the processor of any device with data processing capabilities and run.

[0130] The above-mentioned embodiments only express several implementation modes of the present invention, and the description is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. For those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. An infrared simulation method based on neural network and GPU parallel computing, characterized in that: include: The atmospheric model is used to perform nonlinear mapping of spatiotemporal parameters and environmental parameters, and the all-weather atmospheric transmittance and the radiation value received by the surface are fitted; Preprocess infrared simulation data and use the distributed GPU parallel framework to calculate the patch occlusion probability value and patch temperature value in the simulation scene; When calculating the patch occlusion probability value, the number of sampling points is dynamically adjusted according to the patch geometric area and complexity and the positions of the sampling points on the patch are determined by an adaptive patch sampling method, and the patch occlusion probability value is obtained by solving the occlusion relationship of each sampling point on the patch; The adaptive patch sampling method is specifically as follows: Determine the number of basic sampling points of the patch according to the density of the benchmark sampling points; The number of basic sampling points is multiplied by the complexity weight based on the depth map gradient to obtain the number of sampling points after complexity adjustment. The complexity weight calculation formula is: Where D(x,y) represents the depth value of the patch at the coordinate (x,y) in the depth map. and Respectively represent the rate of change of depth value in x and y directions, represents the complexity weight; The number of sampling points after complexity adjustment is multiplied by the occlusion probability weight to obtain the number of sampling points after occlusion adjustment. The occlusion probability calculation formula is: Among them, β is the occlusion probability weight adjustment factor; n occluded Indicates the number of sampling points whose depth is judged to be less than the center point depth when hemisphere sampling is performed at the center of the patch according to the depth map; n total Indicates the total number of sampling points for hemispherical sampling at the center of the patch according to the depth map; represents the occlusion probability; Determine whether the number of sampling points after occlusion adjustment is less than the preset minimum number of sampling points of the patch. If so, use the preset minimum number of sampling points of the patch as the final total number of sampling points; otherwise, use the number of sampling points after occlusion adjustment as the final total number of sampling points; Introducing a facet occlusion probability value when calculating the facet temperature value; The patch temperature value is optimized to obtain the patch steady-state temperature value, the radiant brightness of a specified band is calculated according to the patch steady-state temperature value, and finally an infrared simulation image of a specified viewing angle in a specified band is rendered.

2. The infrared simulation method based on neural network and GPU parallel computing according to claim 1 is characterized in that: The atmospheric model is a residual neural network; The training process of the atmospheric model includes: Sampling simulated atmospheric data to obtain data that conforms to the distribution of real data as main training data, and collecting real atmospheric data as fine-tuning training data; The main training data and fine-tuning training data both contain spatiotemporal parameters, environmental parameters, atmospheric transmittance, and radiation values ​​received by the surface; the spatiotemporal parameters include latitude, longitude, year, day, hour, and minute, and the environmental parameters include air temperature, humidity, air pressure, visibility, weather patterns, cloud height, and cloud thickness; First, the main training data is used to pre-train the atmosphere model, and then the fine-tuning training data is used to fine-tune the pre-trained atmosphere model.

3. The infrared simulation method based on neural network and GPU parallel computing according to claim 1 is characterized in that: The radiation values ​​received by the surface include direct solar radiation values, scattered solar radiation values, and atmospheric long-wave radiation values.

4. The infrared simulation method based on neural network and GPU parallel computing according to claim 1, characterized in that: Three independent CPU nodes are used to complete the infrared simulation data preprocessing process in parallel, including: CPU node 1: Calculates the depth map and patch data with material information based on the input material segmentation map, material database and model file; CPU node 2: used to splice the all-weather atmospheric transmittance predicted by the atmospheric model and the radiation value received on the ground surface with the spatiotemporal parameters and environmental parameters at the corresponding time points to obtain multi-source fusion data; CPU node 3: Calculate the all-weather solar azimuth according to the input time and space parameters to obtain the direct direction of the sun.

5. The infrared simulation method based on neural network and GPU parallel computing according to claim 1, characterized in that: According to the final number of sampling points, the Poisson disk distribution method is used to determine the position of the sampling points on the patch. According to the direct direction of the sun and the GPU ray tracing algorithm, a bounding volume hierarchy is constructed to solve the occlusion relationship of each sampling point. According to the occlusion relationship of the sampling points, the occlusion probability value corresponding to each patch at each moment is calculated. The occlusion probability value is the ratio of the occluded sampling points on the patch to the final total number of sampling points.

6. The infrared simulation method based on neural network and GPU parallel computing according to claim 1, characterized in that: The calculation of the dough temperature value includes: The energy value of the patch at each moment is calculated through the energy equation. The total solar radiation used to calculate the energy value is defined as the patch shading probability value multiplied by the direct solar radiation value received by the patch plus the solar diffuse radiation value received by the patch. The longitudinal heat conduction is calculated based on the energy value to obtain the dough temperature value.

7. The infrared simulation method based on neural network and GPU parallel computing according to claim 1, characterized in that: Based on the surface temperature value, the multi-grid method is used to calculate the two-dimensional heat conduction and obtain the surface steady-state temperature value; The radiance in the specified band is calculated using Planck's formula based on the steady-state temperature value of the patch.

8. The infrared simulation method based on neural network and GPU parallel computing according to claim 1, characterized in that: The distributed GPU parallel framework refers to the data segmentation of the results of infrared simulation data preprocessing and then transmitting them to multiple GPU nodes for parallel calculation of patch occlusion probability values, patch temperature values, patch steady-state temperature values ​​and specified band radiation brightness.

9. An infrared simulation system based on neural network and GPU parallel computing, characterized in that: include: The atmospheric model module is used to perform nonlinear mapping of spatiotemporal parameters and environmental parameters, and fit the all-weather atmospheric transmittance and the radiation value received by the surface; A preprocessing module, which is used to preprocess infrared simulation data; A distributed GPU computing module is used to calculate the patch occlusion probability value and the patch temperature value in the simulation scene using a distributed GPU parallel framework, optimize the patch temperature value to obtain the patch steady-state temperature value, and calculate the radiant brightness of a specified band according to the patch steady-state temperature value; Wherein, when calculating the patch occlusion probability value, the number of sampling points is dynamically adjusted according to the patch geometric area and complexity and the positions of the sampling points on the patch are determined by an adaptive patch sampling method, and the patch occlusion probability value is obtained by solving the occlusion relationship of each sampling point on the patch; the patch occlusion probability value is introduced when calculating the patch temperature value; The adaptive patch sampling method is specifically as follows: Determine the number of basic sampling points of the patch according to the density of the benchmark sampling points; The number of basic sampling points is multiplied by the complexity weight based on the depth map gradient to obtain the number of sampling points after complexity adjustment. The complexity weight calculation formula is: Where D(x,y) represents the depth value of the patch at the coordinate (x,y) in the depth map. and Respectively represent the rate of change of depth value in x and y directions, represents the complexity weight; The number of sampling points after complexity adjustment is multiplied by the occlusion probability weight to obtain the number of sampling points after occlusion adjustment. The occlusion probability calculation formula is: Among them, β is the occlusion probability weight adjustment factor; n occluded Indicates the number of sampling points whose depth is judged to be less than the center point depth when hemisphere sampling is performed at the center of the patch according to the depth map; n total Indicates the total number of sampling points for hemispherical sampling at the center of the patch according to the depth map; represents the occlusion probability; Determine whether the number of sampling points after occlusion adjustment is less than the preset minimum number of sampling points of the patch. If so, use the preset minimum number of sampling points of the patch as the final total number of sampling points; otherwise, use the number of sampling points after occlusion adjustment as the final total number of sampling points; The rendering module is used to render an infrared simulation image of a specified viewing angle in a specified band according to the calculation results of the preprocessing module and the calculation results of the distributed GPU calculation module.

Images