基于分层稠密对应关系的视频风格化技术和系统

发布时间：2018-05-24 22:13

本文选题：视频风格化 + 分层稠密对应关系　；参考：《中山大学》2013年硕士论文

【摘要】：图像，视频的各种艺术风格渲染(风格化)是计算机图形学非真实感渲染(Non-Photorealistic Rendering, NPR)的重要分支。图像风格化技术通过对输入图像进行分析和渲染，得到不同艺术风格的图像。视频风格化则在图像风格化的基础上，通过视频分析，效果渲染和效果传播，得到具有艺术效果的视频。图像和视频的非真实渲染在动漫，娱乐等领域有广泛的应用，因此得到了众多研究者的关注。近年来，得益于计算机图形学社区的快速发展，视频风格化技术也有了长足的发展。目前，主流的视频风格化技术是基于笔触的视频风格化技术(Stroke-BasedVideo Stylization Rendering)。该视频风格化技术由三方面组成：图像渲染，笔触传播和视频内对应关系提取。图像渲染决定了最终渲染风格；笔触传播通过应用对应关系到笔触，将渲染效果传播到下一帧；视频内对应关系确定了笔触的时空位置，将会影响最终结果是否出现抖动，闪烁，脱落等影响观赏性的问题。为了减轻抖动，闪烁等问题，准确健壮的对应关系是必不可少的。本文结合光流法和运动分层，提出基于运动分层的自适应光流法(Layered AdaptiveOptical Flow, LAOF)用于提取视频内稠密对应关系。首先，使用自适应的光流算法获得视频内的运动信息，从而获得视频内的稠密对应关系。然后在运动信息的基础上，进行运动分层。最后，，将运动分层作为输入，重新求解视频内的稠密对应关系，形成一个迭代优化的模型。此外，本文还通过预渲染(BeforehandRendering)技术来进一步减少效果传播过程导致的抖动。针对视频风格化的组成，本文总结出通用的视频风格化4层架构，并利用该架构实现视频风格化系统，包括本文提出的LAOF算法和预渲染技术。本系统可以与任何基于笔触渲染(Stroke-based Rendering, SBR)的图像风格化算法结合使用，生成具有艺术效果的视频。在实验部分，将根据LAOF的模型特点进行验证实验。并且通过对不同运动类型的视频进行风格化处理，表明本文提出的视频风格化算法在遮挡和非刚性运动区域能很好地减少渲染上的抖动。最后将给出根据本文提出的视频风格化算法实现的视频风格化系统。
[Abstract]:All kinds of artistic style rendering (stylization) of images and videos is an important branch of Non-Photorealistic rendering (NPRR) in computer graphics. Image styling technology through the input image analysis and rendering, the image of different artistic styles. Video stylization is based on image stylization. Through video analysis, effect rendering and effect propagation, the artistic video is obtained. Non-realistic rendering of images and video has been widely used in animation, entertainment and other fields, so it has attracted many researchers' attention. In recent years, thanks to the rapid development of computer graphics community, video styling technology has made great progress. At present, the mainstream video styling technology is Stroke-Based Video Stylization rendering. The video styling technology consists of three parts: image rendering, stroke propagation and video correspondence extraction. Image rendering determines the final rendering style; stroke propagation spreads the rendering effect to the next frame by applying the correspondence relation; the corresponding relationship in the video determines the temporal and spatial position of the stroke, which will affect whether the final result is jitter or not. Flicker, shedding, etc., affecting the ornamental. In order to reduce the jitter, flicker and other problems, accurate and robust correspondence is essential. In this paper, an adaptive layered AdaptiveOptical Flow, LAOF) based on motion stratification is proposed to extract dense correspondence in video by combining optical flow method and motion stratification. Firstly, the adaptive optical flow algorithm is used to obtain the motion information in the video, and the dense correspondence in the video is obtained. Then on the basis of the motion information, the motion stratification is carried out. Finally, the motion stratification is used as input to solve the dense correspondence in the video again, and an iterative optimization model is formed. In addition, the jitter caused by the propagation process is further reduced by prerendering BeforehandRendering. Aiming at the composition of video stylization, this paper summarizes the general four-layer architecture of video stylization, and uses this architecture to implement video stylization system, including the LAOF algorithm and pre-rendering technology proposed in this paper. This system can be combined with any Stroke-based rendering (SBR-based) based image styling algorithm to generate artistic video. In the experiment part, the verification experiment will be carried out according to the characteristics of LAOF model. Through the stylization of different motion types of video, it is shown that the video styling algorithm proposed in this paper can reduce the jitter in shading and non-rigid motion regions. Finally, a video styling system based on the video styling algorithm proposed in this paper is presented.
【学位授予单位】：中山大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP391.41

【共引文献】