Research On Three-dimensional Scene Rendering Synthesis Method Based On Cycle-Consistent Adversarial Networks

With the rapid development of science and technology,computer graphics has achieved a breakthrough in terms of rendering as well as a large number of excellent rendering model appears.However,there is still a gap between the result of those models and photorealistic image.Because the real world scene is complicated and changeable.Meanwhile,there are many objective conditions and the demand of photorealistic rendering is more and more urgent.It's not easy to achieve progress in photorealistic rendering.The emergence of the generative adversarial networks makes it possible to render photorealistic image.In this paper,I combines generative adversarial networks and OpenGL and trains network to get an appropriate model.Then,by modifying the rendering pipeline,the ability of rendering photorealistic image can be improved.The main work has been done in the following aspects:First of all,OpenGL has been learned and mastered in depth.It,s used to build 3D scene and render illumination and shadow.At the same time,I have studied the rendering pipeline of OpenGL.Second,I studied and analyzed generative adversarial networks and its various ways of improvement.Generative adversarial networks can constantly try to generate realistic-looking data in the course of training.One of this improvements,cycle-consisitent adversarial networks can not only learn a lot of real 3D scene data sets,but also transform data in different fields.It is hoped to learn the rendering style from the real data set and transform the non-rendering model to photorealistic image.However,because the training process is not stable enough and in order to obtain better results,I did two groups of experiments.The one of them is using WGAN object function replace the target function of GAN in the CycleGAN network,the other is the original training mode of the CycleGAN network.In the course of training,it is found that the method of WGAN substituting GAN can speed up the convergence rate of the cyclic consistent loss function.In order to explore its reason,two sets of contrast experiments have been done in this paper,and the results show that the weight-curtting method used in Wgan training plays a major role.In order to bring the final training results to the target,this paper proposes an improved scheme:A norm constraint is added to the cyclic consistency loss function,and the weight clipping constraint is done in the original training process.Finally,the experiment proves that the scheme not only accelerates the training speed but also obtain better results.Then referring to OpenGL's rendering pipeline and TensorFlow open source software library features,This paper presents a OpenGL-GAN rendering pipeline,which is mainly embedded in the graph model frolm the training of the CycleGAN network,which makes OpenGL and GAN combine perfectly so that the method of using GAN in OpenGL can render the more realistic image.It lays a good foundation for the future high realistic rendering model,and opens up a new idea for later research.Finally,I use C++/CLR Window Forms to write a system of OpenGL interactive rendering based on the new rendering pipeline,then verify the feasibility of the new rendering pipeline and meet the application needs.