基于图像分割的视频纹理合成技术研究

发布时间：2018-05-21 11:49

本文选题：纹理合成 + 动画　；参考：《上海交通大学》2011年硕士论文

【摘要】：随着计算机图形学以及多媒体技术的不断发展和成熟,简单的图像纹理已经不能满足所有应用的需求,比如在视频制作、虚拟仿真、动画生成和游戏引擎等领域中,都需要使用到视频纹理。视频纹理是指描述某种动态景观的具有时间相关重复特征的帧序列,视频纹理在自然界中广泛存在,比如火焰、瀑布、飘扬的旗帜、飞翔的鸟群等。不过受到拍摄条件和存储设备的限制,所拍摄视频只能持续有限长时间,虽然在一定程度上反映了拍摄事物随时间变化的行为,但是在很多应用中,还是需要使用“无限长时间”的视频纹理。比如,网页上的一个广告,棕榈树在沙滩上迎风摆动,应该使用视频纹理而不是静态图片。类似的,视频纹理还可以在计算机游戏中作为动态背景使用。视频纹理的合成一直是计算机图形学的主要目标和难点之一,本文研究了基于图像分割的视频纹理合成算法。传统的视频纹理合成方法是从一段输入视频中生成连续的、无限长时间的视频的帧序列,这是一种产生视频纹理的很好的方法,缺点是输入视频的体积一般比较大。与此相比,我们提出一种从一些静态图片中生成视频纹理的方法,输入的静态图片在一定程度上反映了场景的动态信息。我们首先利用图像相似函数恢复输入图片的时间先后顺序,然后用二阶马尔科夫链模型对输入图片进行采样,产生后续视频的帧序列,最后我们采用变形技术对相邻帧进行插值,消除相邻帧序列之间的不连续性和不一致性,最后生成连续的、无限长时间的,具有良好视觉效果的视频纹理。本文关注的重点是利用输入的静态图片的信息,生成无限长时间的,具有良好视觉效果的视频纹理。本文的创新主要有以下两个方面:1)图像相似度函数的设计,我们同时考虑图片的颜色和运动物体的形状特征。当运动物体的形状发生较大变化时,我们的图像相似度函数比仅仅使用欧拉距离描述图像相似度效果更好。2)使用基于shape context和thin-plate spline的变形技术做视频帧之间的插值,并且使用反距离权重方法来消除生成的视频帧中存在的空洞。由此可见,研究视频纹理的合成具有重要的应用价值。设计好的视频纹理合成算法,能够提高视频纹理的生成质量,减少生成视频纹理所需要的时间空间耗费,为其他相关应用奠定良好的基础。
[Abstract]:With the development and maturity of computer graphics and multimedia technology, simple image texture can not meet the needs of all applications, such as video production, virtual simulation, animation generation and game engine, etc. Both need to use video textures. Video texture is a time-dependent frame sequence that describes a dynamic landscape. Video texture widely exists in nature, such as flame, waterfall, flying flag, flying birds and so on. However, due to the limitation of shooting conditions and storage devices, the video can only last for a limited period of time. Although to some extent, it reflects the behavior of shooting things over time, but in many applications, Still need to use "infinite time" video texture. For example, an ad on a web page with palm trees swinging against the sand should use video textures rather than static images. Similarly, video textures can be used as dynamic backgrounds in computer games. The synthesis of video texture has been one of the main objectives and difficulties of computer graphics. In this paper, a video texture synthesis algorithm based on image segmentation is studied. The traditional method of video texture synthesis is to generate frame sequences of continuous and infinite video from an input video. This is a good method to generate video texture. The disadvantage is that the volume of input video is generally large. In contrast, we propose a method to generate video texture from some static images, and the input static images reflect the dynamic information of the scene to a certain extent. We first use the image similarity function to restore the time sequence of the input pictures, and then we sample the input pictures by using the second-order Markov chain model to generate the frame sequence of the subsequent video. Finally, we use the deformation technique to interpolate the adjacent frames to eliminate the discontinuity and disconsistency between the adjacent frame sequences, and finally generate the continuous, infinite time video texture with good visual effect. The focus of this paper is to use the input information of static images to generate video textures with infinite time and good visual effect. The innovation of this paper mainly includes the following two aspects: 1) the design of image similarity function. We also consider the color of images and the shape features of moving objects. When the shape of moving object changes greatly, our image similarity function is better than only using Euler distance to describe image similarity. And the inverse distance weight method is used to eliminate the holes in the generated video frames. Therefore, the research of video texture synthesis has important application value. The designed video texture synthesis algorithm can improve the quality of video texture generation, reduce the time and space cost of video texture generation, and lay a good foundation for other related applications.
【学位授予单位】：上海交通大学
【学位级别】：硕士
【学位授予年份】：2011
【分类号】：TP391.41

【相似文献】