Research On Data-driven Gaseous Phenomenon Reconstruction And Detail Enhancement

The modeling and simulation of realistic smoke animation has been a research hotspot in computer graphics,which plays an important role in the fields of film special effects,video game design,virtual reality,etc.With the development of fluid capture devices and technologies,how to reconstruct three-dimensional physical fields,infer and optimize physical parameters,and re-simulate smoke animation via capturing smoke phenomena in the real world,has gradually become an important research direction in smoke animation.However,the real-world captured smoke data usually has serious noise and environment interference,and traditional physics-inspired reconstruction methods suffer from a number of problems,such as high complexity,time-consuming computation,and difficulty during physical parameter estimation,which poses great challenges to the reconstruction and inverse modeling of smoke phenomenon.To tackle the above difficulties,this thesis deeply analyzes the mechanism of physicsbased model and data-driven method,and combine their advantages to propose a series of data-driven and physically correct rapid reconstruction and realistic detail enhancement methods for smoke phenomenon.The main contributions of this paper are:· A rapid reconstruction method for smoke phenomena from limited view based on generative adversarial network.Aiming at the problem of lacking global information and time-consuming computation for the reconstruction of smoke phenomena from sparse view,we combine the data-driven model and physics-based simulation to propose a density generation network considering the similarity of visual characteristics and a velocity estimation network considering the divergence constraint.Through the consistency penalty between multiple frames,the reconstructed density field,velocity field and the differentiable advection layer are tightly coupled to correct the cumulative error of long-sequence videos,achieving rapid and end-to-end reconstruction of smoke phenomena.· A smoke reconstruction method for captured video.In view of the serious environment interference in the captured video and the lack of reliable physical information in the single-view smoke video,a dataset for training the segmentation model is constructed using differentiable rendering technology combined with real-world background videos and simulation results of Eulerian models.Next,the smoke segmentation model analyze the different characteristics of high and low frequencies for input video by extracting multi-scale and spatiotemporal features,to remove the background interference in the captured video and obtain two-dimensional estimated density maps,so as to reconstruct the three-dimensional physical fields of smoke phenomena.Then we combine differentiable physics framework and various prior constraints to rapidly optimize the smoke source parameters,achieving rapid reconstruction of captured smoke phenomena.· A method for optimizing physical parameters of smoke simulation based on differentiable physics and deep learning.This method takes the reconstruction results of the captured video and corresponding physical parameters as the input of the differentiable physical solver to optimize the preset physical parameters that can be used for re-simulation,and proposes a physical similarity measurement network to measure the distance of two smoke sequences in the latent space,which is used as the objective function to optimize the physical parameters,so as to reduce the visual difference between the re-simulation and the captured scenes,and realize the rapid inference of smoke simulation parameters.· A detail enhancement method for smoke simulation based on convolution neural network.To ease the time-consuming generation of high-resolution smoke scenes,this method simulates high-resolution smoke scenes through Eulerian model,and constructs training sets with different resolutions but consistent motion trends through down-sampling.Then we deeply analyze the advantages and disadvantages of different physical fields in detail enhancement,and train convolutional neural networkswith physical constraints to generate high-resolution velocity fields,combined with the long-term mechanism and the differentiable physical advection algorithm to ensure the effectiveness and robustness of the long sequence.In this way,the rapid detail enhancement of the smoke scene is realized.A large number of experiments have been conducted in this paper,and the smoke scene reconstruction,re-simulation and detail enhancement results obtained from the experiments are rendered by the physics-based rendering engine.The realistic smoke simulation results verify the effectiveness and robustness of the methods proposed in this thesis.