【作者】 唐国维；

【导师】 顾国昌；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2010， 博士

【摘要】 随着Internet的迅速发展,图像压缩在多媒体信息的存储和传输中起着至关重要的作用。传统的基于DCT的编码方案在高压缩比条件下不可避免地会出现方块效应和飞蚊噪声,严重影响主观质量。小波变换因其特有的与人眼视觉相符的多分辨率分析能力以及方向选择能力而被广泛应用于图像压缩领域。JPEG2000标准的颁布,说明小波变换已成功取代DCT成为新一代编码算法的主要变换工具。嵌入式小波图像编码方法具有分辨率可分级、质量可分级以及较强的抗误码性能等优良特性,成为图像编码算法研究的主要方向。但目前的研究方法尚未能完全利用小波系数的所有统计特性,如何更好地利用小波变换系数的各种相关性以有效组织小波系数,并将小波变换技术与其它技术相结合以进一步提高压缩性能,是小波图像编码算法研究的主要问题。本文对现有的小波图像编码算法进行了比较详细的分析,主要从以下四方面对嵌入式小波图像编码算法及应用进行研究。(1)对基于分形的小波图像编码算法进行了深入研究。针对分形编码匹配搜索时间开销大的问题,提出了一种基于分块的分形搜索树结构,有效地减小了定义域池的搜索范围,并提高了匹配精度。对值域块则依据不同子带小波系数的重要性采取自适应的划分方式,然后寻找一组压缩仿射变换,使其构成的IFS逼近给定的吸引子,最后对获得的分形参数进行熵编码。该算法重构图像质量有所提高,特别在中低码率下PSNR提高明显。(2)由于分形编码具有编解码不平衡的问题,进一步从SVM的角度对小波图像编码算法进行了研究。在小波变换基础上,构建了一种适合SVM回归的树结构,并对树中显著性小波系数的分布规律进行了分析,提出了一种线性阈值选取方法,使得参与回归的小波系数分布趋于均匀,以利于回归。在此基础上提出依据阈值进行误差参数动态选取的方法,进一步提高了回归拟合精度。最后对支持向量及其权重进行熵编码。该算法实现了嵌入式特性,与现有算法相比压缩比有了较大提高。(3)虽然小波变换与分形和SVM结合取得了较好的编码效果,但小波变换在表示二维图像时并不是最优或最稀疏的函数表示方法。为此对基于Contourlet变换的图像编码算法进行了研究。将小波变换与Contourlet结合获取非冗余特性的Contourlet变换,然后对分解后的子带进行基于HVS的视觉加权处理。针对Contourlet方向分解没有考虑子带内系数分布特性的问题,研究了熵随子带方向分解数的变化情况,进而提出了基于熵的方向分解优化算法,增强了子带内部的局部相关性。在此基础上采用SPECK算法对经过方向优化分解的子带进行编码。该算法不仅峰值信噪比得到提高,而且对局部纹理失真较小。(4)对基于小波包的EBCOT岩心图像压缩进行了研究。基于岩心图像纹理特征突出的特点,采用小波包分解方式以增强对高频细节的表示能力。针对率失真方法计算量较大以及熵不能完全表征图像纹理信息的问题,提出了利用分形维数确定最优小波包基的方法,分解方式与子带方向特性比较吻合。在此基础上采用EBCOT算法对岩心图像进行压缩,并提出了通道并行扫描方法以提高EBCOT算法的执行速度。实验结果表明,算法的压缩比和速度均高于JPEG2000。更多还原

【Abstract】 With the rapid development of the Internet, image compression plays a vital role in the storage and transmission of multimedia information. Conventional DCT-based coding scheme will inevitably produce block artifacts and mosquito noise in the case of high compression ratio, which results in a seriously harming to the subjective quality of the reconstructed image. For the capability of multi-resolution analysis consistent with human visual system and the direction selectivity, wavelet transform has been widely used in the field of image compression. The promulgating of JPEG2000 indicates that wavelet transform has successfully replaced DCT and become a major transform tool of the new generation of image coding algorithm. Embedded wavelet image coding, taking the advantage of resolution scalable, quality scalable and strong error-resilient performance, has been an important research direction of image coding. But the current methods are still not able to take full use of all the statistical characteristics of wavelet coefficients, how to make better use of the variety of relevance of wavelet transform coefficients in order to effectively organize the wavelet coefficients, especially to combine wavelet transform with other techniques to further enhance the compression performance, are the main topics in the research of wavelet image coding.In this dissertation, the popular wavelet image coding algorithms are analyzed in detail, then embedded wavelet image coding algorithms are studied mainly from four aspect as follows.(1)The fractal based wavelet image coding algorithm was studied deeply. Aimed at the problem of too much time overhead for matching search of fractal image coding, a block based fractal searching tree structure was proposed, which reduced the searching range of domain pool effectively and improved the match precision. A self-adaptive way was adopted to partition range blocks according to the importance of wavelet coefficients in different sub-bands. Then seek for a group of compression affine transforms to make the IFS approximate to the given attractors, and at last carry out entropy coding to the fractal parameters obtained. The quality of reconstructed image is better than that of the existing algorithms, especially the PSNR increases significantly in medium and low bit-rate. (2)Considering the unbalance problem of fractal in coding and decoding, the wavelet image coding algorithm was studied further from the perspective of SVM. Basing on the wavelet transform, a tree structure suitable for SVM regression was constructed, and the distribution of significant coefficients in the tree was analyzed. Then a linear threshold selection method was proposed to make the distribution of coefficients involved in the regression tend to be uniform in order to facilitate the regression process. On that basis, a dynamic error parameter selection method was put forward according to the thresholds and the precision of the regression fitting was improved further. At last the support vectors and their weights are entropy coded. The algorithm is provided with embedded feature, and the compression ratio improved significantly.(3)Although better coding effect was achieved by combining with fractal and SVM, the wavelet transform is not the best or the most sparse in two dimensional image representation. So Contourlet transform based image coding algorithm was studied. Combine the wavelet transform with Contourlet to obtain non-redundant Contourlet transform, then perform visual weighting based on HVS to each sub-band. For that the direction decomposition of Contourlet do not consider the distribution characteristics of coefficients inside each sub-band, the changing situation of entropy with the direction number of sub-band decomposition was studied. Then an entropy based optimization algorithm of direction decomposition was presented and the local relevance inside each sub-band was enhanced. The SPECK algorithm was adopted to encode each sub-band decomposed through the optimized direction decomposition algorithm. Not only the PSNR of the reconstructed image is improved, but also the distortion for local texture is much less.(4)Core image compression using wavelet packet based EBCOT algorithm was studied. Considering that the texture feature is predominant in core images, wavelet packet decomposition was adopted to enhance the representation ability to high-frequency details. Due to the large amount of calculation of the rate-distortion method and the incapacity in representing texture information effectively of entropy method, a fractal dimension based optimal wavelet packet base selection method was proposed, in which the decomposition pattern is consistent with the direction feature of each sub-band. On this basis, use EBCOT algorithm to execute compression to core image, and a parallel pass scanning method was proposed to enhance the speed of EBCOT. The experimental results demonstrate that the compression ratio and the speed are both higher than JPEG2000.更多还原