【作者】 甘涛；

【导师】 朱维乐；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2008， 博士

【摘要】 目前，数字图像已在诸多领域有着广泛的应用。日益增长的海量数据，给存储器的容量，通信信道的带宽以及计算机的处理速度带来了前所未有的极大压力。如何有效地组织、存储、传输和恢复图像数据，或探索和研究压缩比更高，质量更好，复杂度更低的图像编码技术是现代信息处理的关键任务之一。随着因特网在全球范围的日益普及，各种新兴多媒体业务的相继出现，传统的编码方法面临巨大的挑战。除了应具有好的压缩效果外，图像编码还需要适应多变的应用环境，如用户的不同需求，异构网络支路的不同传输条件和终端的不同处理能力等等。本论文以探索满足上述要求的高性能图像编码方法为目标，围绕图像压缩的变换、量化和编码等几个主要环节而展开研究，重点开发了基于小波变换和基于稀疏分解的两类高性能的可伸缩图像编码器。论文的创新性研究的主要内容为：1．系统深入地研究嵌入式小波图像编码技术，对当前先进的小波系数组织和编码方法进行分类总结和探讨。2．深入研究小波系数集合划分的原理和方法，提出一种自适应的分层树集合分裂方法，扩展了传统的空间方向树结构，将多个方向树组合起来共同表示一个非重要系数集合，节约了集合表示的比特数。3．探索研究矢量量化技术在嵌入式图像编码中的应用，提出一个格型矢量量化与分层树集合分裂算法相结合的编码方案。该方案选用规则的格点作为码书，不需要训练和存储，克服了传统矢量量化计算和存储复杂度高的缺点。改进多级增益一形状矢量量化方法并使之与集合分裂算法有机结合，实现了对细节较丰富图像更好的编码效果。4．在分析和借鉴多种先进编码方法的基础上，提出一个高性能的嵌入式小波编码算法。该算法联合采用了基于聚类的重要系数表示法和基于零块结构的非重要系数表示法，以及基于上下文的算术编码，共同挖掘小波子带内聚集和子带间相似的特性。在嵌入准则的指导下，提出基于聚类表征的分类和排序方案，借助链表结构有效地实现了精细的分数位平面编码，在较低复杂度下实现了比当前先进算法更好的有失真和无失真编码性能。在此基础之上提出了一种高伸缩性的图像编码方案。该方案利用位平面编码和小波分解固有的多分辨率特性实现了质量和分辨率的可伸缩。它解除了编解码之间的约束，编码方无需了解解码方的具体状况，生成的具有层次结构的码流可以很方便高效地被解析和解码，以满足用户对质量和分辨率的不同需求。5．研究信号稀疏逼近问题，设计适合图像表示的冗余几何原子库，探索快速图像稀疏分解方法。对流行的匹配追踪算法进行深入地分析和研究。总结了匹配追踪分解的特点，并提出用库原子的空间相对距离来估计其互相关信息的方法。在此基础上提出一系列快速图像稀疏分解算法，在保持主客观质量的前提下，大大提高了图像分解的速度。比如对512×512测试图像，相对于最新的全局匹配追踪算法，本文算法的速度增益达到115．39倍，而PSNR平均下降0．25 dB。6．探索研究稀疏分解在图像编码中的应用。对图像进行快速稀疏分解并分析和总结分解原子的分布规律，提出与之相适应的块划分编码方法，对原子位置参数和系数进行联合编码。形成的编码器在计算复杂度、编码效率和伸缩性能等方面都优于最新的匹配追踪编码器，特别是在前两方面，其优势十分明显。比如对512×512测试图像，编码率为0．5 bpp时本文编码器的平均PSNR增益高达1．75 dB。特别地，该编码器提供了较传统方法更灵活的伸缩性，允许通过编码域的参数变换来获得任意分辨率大小的重建图像，可在网络或移动终端的图像业务中得到很好的应用。更多还原

【Abstract】 Nowadays,digital image has found wide applications in many areas.The vastamount of data imposes unprecedented pressure on storage capacity,transmissionbandwidth and computer processing speed.To find efficient ways to organize,store,transmit and restore image data or develop coding methods which enjoy the propertiesof high compression ratio,good quality and low complexity,is one of the key tasks ofmodern signal processing.With the popularity of Internet and the emergence of diversemultimedia applications,great challenges have been presented to traditional methods.Besides the need of high compression efficiency,image coding is also required to adaptto heterogeneous environments,including different demands from users,varioustransmission conditions as well as different receiver capacities.Aiming at developing high performance image coding methods which meet theabove requirements,this dissertation investigates major issues of image coding,including transform,quantization and coding.The main results are as follows:1.The embedded wavelet image coding techniques are studied in detail.Theadvanced organization and coding methods of wavelet coefficients are classified anddiscussed.2.The principle and methods of set partitioning are investigated.An adaptive setpartitioning algorithm is proposed.The structure of spatial orientation tree is extendedand several trees are joined together to represent one insignificant set,resulting in bitsavings in set representation.3.The application of vector quantization to embedded image coding is investigated.A coding scheme which combines the algorithm of set partitioning in hierarchical tree(SPIHT) and lattice vector quantization is proposed.In this scheme,regular lattices arechosen as the codebook,no training or storage of the codebook or significant encodingcomputations is required and thus the high complexity problem of traditional vectorquantization is overcome.The multistage gain-shape vector quantization is modifiedand incorporated into SPIHT system,which provides better coding performances forimages with more details. 4.By combining several advanced techniques,an efficient embedded waveletcoding algorithm is proposed.It employs morphological representation,quadtreepartitioning as well as efficient context-based adaptive coding to jointly exploit bothwithin-subband clustering and cross-subband similarity of wavelet coefficients.Guidedby the embedding principle,a cluster-based classification and sorting strategy isproposed.Based on the list structure,a fine fractional bit-plane coding is achieved withrelatively low complexity.Experimental results show that the proposed algorithmoutperforms the state-of-the-art coders for both lossy and lossless compression.Inaddition,a highly flexible image codec is proposed based on above algorithm.Itexploits inherent multi-resolution and multi-precision nature of subband bitplane codingto achieve both resolution and rate scalability.Importantly,the encoding can beindependently conducted without knowledge of the final decoding situation.Ahierarchical codestream structure is established and it can be easily parsed and reorderedto meet the different requirements of the users.5.The problem of sparse approximation is studied and fast methods of sparseexpansion over the redundant geometric dictionary are investigated.The popularmatching pursuit (MP) algorithm and its properties are analyzed thoroughly.A novelidea of estimating the cross-correlation between dictionary atoms is introduced and aseries of fast MP algorithms are proposed.Experimental results show that the proposedalgorithms offer important complexity reductions while maintaining the highapproximation performances.For instance,compared with the latest full searchalgorithm,a speed-up gain up to 115.39 times is achieved for 512×512 test images withan average PSNR loss of 0.25 dB.6.A geometric image coding scheme based on fast MP algorithms is designed.Thedistributions of selected atoms and coefficients resulting from image decomposition arestudied.A block partitioning coding method is proposed to jointly code the atompositions and coefficient magnitudes.The proposed method has striking advantagesover the latest MP coder in computational complexity,coding efficiency as well asscalability.For instance,for 512×512 test images an average PSNR gain of 1.75 dB isachieved.Notably,thanks to the geometrical structure of the dictionary,the new coderprovides interesting adaptability features which allow the codestream to be easily andefficiently decoded at any spatial resolution.This makes it very attractive for various imaging applications over heterogeneous networks.更多还原