基于SLWE的自适应熵编码器概率估计模型应用研究

发布时间：2018-08-07 12:36

【摘要】：熵编码,特别是自适应熵编码,是大多数图像和视频压缩编码标准和许多非标准编码器的核心部分,例如JPEG2000中的MQ编码器和H.264/AVC中的CABAC编码器均用到了自适应二值算数编码。熵编码的编码性能主要与两个因素有关,一是概率模型与信源的实际特性相符的程度,二是编码器依据概率模型为待编码符号分配码字的方式。对于第二点,目前,当给定概率模型时,很多熵编码方法的平均码长已经非常接近给定概率模型下的香农熵,关于这方面的性能提升已经非常有限。而对于第一点,当待编码数据统计特性较平稳时,概率模型的建立问题相对容易,但当待编码数据统计特性经常发生变化时,概率模型往往会与信源的实际特性有所偏离,影响编码性能,若概率模型能实时的反映待编码数据这种特性变化,理论上可以取得更好的编码效果。基于以上背景,本文对数据特性经常变化时熵编码器概率模型的建立问题展开了研究。首先介绍了与熵编码密切相关的信源熵、概率模型和估计熵的相关定义,分析了熵编码的码长与待编码数据的信源熵和在概率模型约束下的估计熵之间的关系,阐述了自适应熵编码器中概率估计模型的重要作用。基于以上基本理论,接着给出了本文中平稳数据和非平稳数据的定义以及相关数学表示,并分析了静态模型以及传统的基于贝叶斯参数估计理论的概率估计算法不适合对非平稳数据进行概率估计的原因并进行了相关实验验证,最后介绍了加窗法和遗忘因子法两种经典的非平稳环境下的概率估计算法,并分析了各概率估计算法对于不同特性数据的概率估计效果,最后探讨了非平稳数据特性变化的剧烈程度对整个概率估计过程的影响,提出了根据数据的变化特性调节概率估计算法的自适应能力的基本思想。其次,研究了随机学习弱估计理论(stochastic learning weak estimators,SLWE)在非平稳数据概率估计问题中的应用。首先介绍了基于SLWE的二项分布和多项分布参数估计过程,并从定性和定量两方面分析了其适用性,深入剖析了其概率更新的内在原理,并将其与加窗法比较,分析了其与加窗法的内在联系。最后利用区间编码的架构设计了利用SLWE算法进行概率估计的区间编码方法,详细叙述了将SLWE算法移植到区间编码中时可能会引起的区间退化问题和由于浮点累加的舍入误差引起的解码问题及相应的解决办法。最后通过实验分析了新的区间编码方法对不同特性数据的编码效果。最后,提出了根据数据局部特性变化情况自适应调节算法自适应能力的参数自适应SLWE算法。该算法针对实际数据特性变化情况比较复杂的特点,首先利用数据的局部统计特性的变化情况分析数据的非平稳度,然后提取出特性变化比较大的位置作为变化点,最后根据变化点的位置和特性变化程度自适应的调节SLWE算法的学习因子,相应的改变算法的自适应能力和收敛能力以适应数据的局部特性变化。最后给出了实验分析。
[Abstract]:Entropy coding, especially adaptive entropy coding, is the core of most image and video compression coding standards and many non-standard encoders. For example, MQ encoders in JPEG2000 and CABAC encoders in H.264/AVC are used for adaptive two value arithmetic coding. The coding performance of entropy coding is mainly related to two factors, one is probability model. The degree to which the actual characteristic of the source is consistent with the source, two is the way the encoder assigns the code word to the coded symbol based on the probability model. For the second point, the average code length of many entropy coding methods is very close to the Shannon entropy of the given probability model, and the performance improvement in this respect is very limited. For the first point, when the statistical characteristics of the coded data are relatively stable, the establishment of the probability model is relatively easy, but when the statistical characteristics of the coded data often change, the probability model often deviates from the actual characteristics of the source and affects the coding performance. If the probability model can reflect the characteristics of the data to be encoded in real time, the probability model can change the characteristics in real time. Based on the above background, this paper studies the establishment of the entropy encoder probability model when the data characteristics are changing frequently. Firstly, it introduces the correlation definition of the source entropy, the probability model and the estimation entropy, which is closely related to the entropy coding, and analyzes the code length of entropy coding and the data to be encoded. The relationship between the source entropy and the estimation entropy under the constraint of the probability model, the important role of the probability estimation model in the adaptive entropy encoder is expounded. Based on the above basic theory, the definition of the flat and non-stationary data and the relative mathematical representation are given, and the static model and the traditional Bayesian based on the Bayes are analyzed. The probability estimation algorithm of the parameter estimation theory is not suitable for the reason of the probability estimation of non stationary data and has been verified by the related experiments. At last, the probability estimation algorithm of two classical non stationary environments is introduced, and the probability estimation effect of each probability estimation method for different characteristic data is analyzed. Finally, the influence of the intensity of the nonstationary data characteristics on the whole probability estimation process is discussed. The basic idea of the adaptive ability of the probability estimation algorithm based on the change characteristics of the data is proposed. Secondly, the probability of the stochastic learning weak estimation (stochastic learning weak estimators, SLWE) is studied in the nonstationary data probability. The application of the estimation problem is introduced. First, the two distribution and multi distribution parameter estimation process based on SLWE is introduced, and its applicability is analyzed from two aspects of qualitative and quantitative. The inherent principle of the probability updating is analyzed deeply, and compared with the window method, the inner connection between the method and the window method is analyzed. Finally, the framework of the interval coding is used. The interval coding method of probability estimation using SLWE algorithm is designed. The problem of interval degradation which may be caused by the transplantation of SLWE algorithm into the interval coding and the decoding problem caused by the rounding error of floating point accumulation are described in detail. Finally, the new interval coding method is analyzed by the actual analysis. In the end, a parameter adaptive SLWE algorithm based on adaptive ability of adaptive adjustment algorithm based on local characteristic changes of data is proposed. The algorithm aims at the complex characteristics of the change of the actual data characteristics. First, the non stationary degree of the data is analyzed by the change of the local statistical characteristics of the data. In the end, the position of the characteristic change is taken as the change point, and the learning factor of the SLWE algorithm is adjusted adaptively according to the position of the change point and the degree of change of the characteristic. The adaptive ability and convergence ability of the algorithm are changed to adapt to the change of the local characteristic of the data. Finally, the experimental analysis is given.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN762

【相似文献】