Research On Some Key Technologies Of Ambient Occlusion In Computer Graphics And Images

Ambient occlusion is a computer graphics method of simulating low-frequency global illumination and has been widely used in real-time rendering,data visualization and computer-aided design.Meanwhile,the estimation and decomposition of light is often an important step in image processing.Ambient occlusion can offer a simple but efficient illumination model to handle various tasks,such as intrinsic image decomposition,multiview stereo and photometric stereo.Therefore,in-depth research on ambient occlusion will help to advance computer graphics and image processing technologies.However,in the field of image processing,the estimation of ambient occlusion is still a major problem,because it requires imposing additional constraints on geometry and illumination priors.For computer graphics,it is difficult to ensure accurate estimation of real-time ambient occlusion.To address these limitations,this thesis focuses on several key problems of ambient occlusion by solving ambient occlusion in with more generalized inputs,improving the estimation accuracy of real-time ambient occlusion and applying key ideas of ambient occlusion to other computer graphics tasks.The contributions of this thesis are listed as following:1.An implicit method for estimating ambient occlusion of dynamic-lighting multiframe images is proposed.Previous methods are unable to estimate ambient occlusion from dynamic-lighting multi-frame images with complex object geography or intense ambient light.By using an implicit strategy,our method removes the albedo from the proposed maximum prior so as to estimate ambient occlusion and lighting ratio.Then,the albedo is updated.Experiments show that the proposed method is able to estimate ambient occlusion consistently and precisely.Moreover,it can handle objects with complex geometries.2.A new method is proposed to estimate the ambient occlusion of single-frame natural images.The existing approaches require multiple images and are inaccurate in estimating ambient occlusion.By using CNN,the proposed method is able to estimate ambient occlusion of single-frame natural images in an end-to-end manner.Three different network structures are designed and considered.A synthetic dataset is generated to train the network.Results show that the proposed method is able to handle more generalized inputs and achieves higher accuracy than the existing methods.Moreover,the proposed method is the first to estimate ambient occlusion of single-frame natural images.3.A platform for computing ambient occlusion in real time is proposed.To address the under/over-estimation problems of screen-space-based ambient occlusion methods suffer,a hybrid platform based on low-frequency ray tracing and new platform is proposed combing low frequent ray tracing based method and Mento Carlo denoising.An optimized CNN structure is introduced to handle Mento Carlo denoising tasks in real time.To improve network efficiency,trade-offs of different loss functions are compared and high-quality training samples are generated.Experiments show that the proposed method is able to produce hundreds of frames per second with greater accuracy.4.A view-volume network is proposed to supplement semantic information of 3D scenes.Inspired by screen space methods,a novel network structure is designed,which is consisted of a view network,a volume network and a reverse rasterization projection module.The view network extracts local features.The volume network learns 3D context information.All these information is organized by the reverse rasterization projection module.A comparison is made between previous methods and the proposed network.Results show that the proposed network achieves the best results in accuracy and improves training/test speed by 4-7 times.