【作者】 宫铭举；

【导师】 王汝霖；

【作者基本信息】 中国海洋大学 ， 通信与信息系统， 2005， 硕士

【摘要】 视觉是人类获取外部信息最为重要的途径。“百闻不如一见”,视觉信息给人们以直观、生动的形象。随着计算机与数字通信技术的迅速发展,特别是网络和多媒体技术的兴起,图像压缩已受到人们越来越多的关注。 小波分析是从付里叶分析发展而来的一种新的时频分析方法,具有良好的空间——频率局部化特性,作为小波分析的再发展,多小波不仅保持了小波的优点,而且克服了小波的一些缺陷,同时具备了实际应用中十分重要的光滑性、紧支性、对称性、正交性。这使得多小波尤比单小波具备更优良的性能(其在图像压缩方面)这决定了其在这方面将被越来越广泛的研究和应用。 作为图像有损压缩的方法之一,矢量量化的压缩比极大。一般地讲,压缩比越大,失真越大:压缩比越小,失真越小;同时,矢量量化也是一个大量计算的过程,随着一些改进算法的提出,矢量量化技术越来越成熟。 本文重要的工作就是探讨了多小波系数特性,针对其特点采用合适的矢量量化方法。实验中,采用了适合图像压缩的CL多小波与自组织特征映射神经网络SOFM,提出了一套切实可行的两者相结合的图像矢量量化方法,其创新点有二: 其一:创新性地将多小波与SOFM结合在一起进行矢量量化,填补了这方面的研究空白; 其二:根据多小波系数的特点,将其一分为三各自进行并行矢量量化,解决了实际中内存不足的问题,提高了编解码的速度,并同时提高了编码质量。 当然,对自组织特征映射网络来说,选择初始的矢量训练集是非常重要的。实验中,我们选用了各频量丰富的Lena,图像作为初始训练集。这样,在一定程度上避免了码本的局部最优化,并可以比较均匀的划分频率空间,以提高该码本的通用性能。实验结果表明,我们获得的码本具有较好的通用性,能够针对不同统计特性的图像进行良好的矢量量化。 最后,作为一个应用方法的探索,我们总结和展望了多小波图像矢量量化的发展前景,并提出一些可能的技术。更多还原

【Abstract】 Vision is the most important way for one to get information from the outer world ."It is better to see for oneself rather than to hear for many times" , and one can get intuitionistic, visualized impression by vision .With the fast development of computer and digital communication technology , especially with the rise of Internet and multimedia, the image compression has attracted more and more attention.Wavelet analysis is a new kind of time-frequency technology .which developed on the basis of Fourier Transform, and it has good space-frequency localization characteristic.As the further development of wavelet multiwavelet not only keeps the excellent characteristics of time and frequency field owned by wavelet but also overcomes the defects of wavelet. Multiwavelet combines smoothness, compactness, symmetry and orthogonality which are very important in practice ,and especially in image compression it has better functions than wavlet, which results in more and more research and applications on it.As one method of image lossy compression vector quantization has great compression ratio. Generally speaking, compression ratio is bigger and hence distortion is bigger, on the other hand, compression ratio is smaller and hence distortion is smaller. At the same time vector quantization is also a process with a lot of computing, but with some ameliorative algorithms , vector quantization has become more and more maturational .The central work of this paper is to probe into the coefficients which are made from a image by multiwavlet, and according to their distribution characteristic a vector quantization algorithm is put forward. In experiment based on CL multiwavelet and SOFM a image vectorquantization design is advanced. There are two creations, as follows, First, combine multiwavlet and SOFM to make a vector quantization. Second, divide coefficients into three parts according to the characteristic of them. Then resolve the problem of shortage of memory, and at the same time improve the speed of coding and en-coding.Of course, it is very important for a self organization feature map network to select a good initial training collection, and image Lena is adopted to do the vital job, image Lena has a rich frequency spectrum, which can avoid the local optimization of code table, and partition the frequency space equably to make the code table fit for as many image as possible .The experiment result show that the code table which is made by image Lena can fit for many different statistic feature images to make vector quantizations.In the end, we make a conclusion and look forward to the future of multiwavlet with vector quantization , and put forward some possible technology.更多还原