快速分形图像编码及高压缩比纹理压缩方法研究

【作者】 高雪峰；

【导师】 叶正麟；

【作者基本信息】 西北工业大学 ， 计算数学， 2005， 硕士

【摘要】 分形图像压缩作为一种新的压缩方法,因其具有压缩比高、解码速度快、解码图像与分辨率无关等优点,十余年来引起了众多学者的关注和研究。但是,这种方法存在一个最大的问题:编码时间过长,主要是因为编码时定义域块的搜索量非常可观。这实际上已成为了该方法走向高效能实用化的最主要障碍,因此分形编码加速方法已成为了近些年来分形压缩的一个研究热点。目前许多分形编码加速方法或者以牺牲解码图像质量为代价,或者加速效果不明显。本文提出了一种混合的快速分形编码方法,将分形编码的父块库看作向量空间,通过正交特征量化将子块和父块的匹配转化为欧式空间的最小距离问题,在分类的基础上,将向量子空间建立成kd-tree,匹配问题进一步转化为在树结构上的最近邻居查找问题,再引入(1+ε)最近邻居查找进一步加速,可达到实时编码。试验结果表明,同以往的方法相比,本方法在基本保证解码质量的基础上,可以极大地提高分形编码速度。同不加速分形编码相比,混合方法带旋转反射变换编码速度可以提高66倍,不带旋转反射变换可以提高653倍。 纹理是图像处理,图像真实感造型领域的很重要的概念,其自身有很多独特的性质使其可获得高压缩比。将纹理看成是图像本身的局部性质,就可将纹理压缩与传统压缩方法结合起来弥补后者的不足。成熟的纹理压缩方法都是基于硬件的,纹理压缩的高压缩比潜力并没有充分被发挥出来。矢量量化技术的码书训练原理与纹理的纹元重复特性刚好一致,并且纹理合成方法的成熟发展为纹理的纯算法压缩提供了方向。本文提出了新的纹理压缩方法,将纹理进行采样,用矢量量化技术从中训练出码书,与传统矢量量化不同,不再进行键集编码,而是以码书作为编码结果,这就是高压缩比的关键,以此码书作为纹理合成的“样图”,利用纹理合成技术作为解码过程。如果是块采样,则用基于块的纹理合成,如果是象素采样,则用基于象素的纹理合成。试验结果表明本方法保证解码质量的基础上可以获得很高的压缩比,而且可以根据图像的质量控制压缩比,由于纹理合成的优势,还可以解码出任意尺寸的图像,在保证逼真度的基础上,可以获得32:1或更高的压缩比。更多还原

【Abstract】 Fractal image compression, as a new scheme of image compression, has received a great deal of attention and study from researchers all over the world in the field of image compression because of its desirable properties such as fast decoding, resolution independence of decoded image and high compression ratio. However, there is a especially unsatisfying problem in this method: too long encoding time, mainly because of the considerable number of domain blocks to compare with for each range block in encoding phase, which, in fact, prevent fractal image compression from becoming a practical method for image compression, hence fast encoding has become a hot issue in fractal image compression. Existed fast encoding methods are often at the cost of image quality, or can only obtain poor speed-up ratio. In this paper, we give a hybrid fractal image encoding technique, which takes on the domain blocks of fractal coding is a vector space, transforms the match of the range block and the domain block into the minimized distance problem of the Euclid space, establishes kd-tree for each sub-space at basis of classifying, further translates the match problem into the nearest neighbor search on tree structure and introduces 1+ε nearest neighbor search to more speed up. It can reach run-time coding. The experiments indicate that compared to the old method, our method can speed up the fractal encoding perfectly. Compared with no speeding-up coding, when ε=14, the hybrid fractal image coding can improve 66 times with rotation and reflection transformations, and can improve 653 times without rotation and reflection transformations.Texture is an important concept of the image processing and image third dimension sculpting whose many special properties make it obtain high compression ratio. Taking texture as the local property of images we can combine the texture compression and traditional compression methods to offset the disadvantage of the latter. The advanced texture compression methods are all based on hardware, so the potential of high compress ratio couldn’t showed completely. The theory of training codebook of vector quantization accords with the repetition property of the texton, and the development of the synthesis methods give the texture compression the newway. In this paper, we give a new texture compression method, which samples the texture, then trains the codebook by vector quantization and takes the codebook as a result of coding, as is different form conventional vector quantization. The codebook is taken on as the "sample texture" of texture synthesis, and the decoding just is texture synthesis. If using blocks sampling, the decoding process is texture synthesis based on blocks. And instead of pixels sampling, the one is based on pixels. The experiments indicate that when insuring the quality we can attain high compression ratio, control it, and get decoding image of random size. Guaranteeing the fidelity, we can get a ratio of 32:1 or more.更多还原