CN104702961A

CN104702961A - Code rate control method for distributed video coding

Info

Publication number: CN104702961A
Application number: CN201510086215.4A
Authority: CN
Inventors: 张登银; 蔡睿
Original assignee: Nanjing Post and Telecommunication University
Current assignee: Nanjing Post and Telecommunication University
Priority date: 2015-02-17
Filing date: 2015-02-17
Publication date: 2015-06-10
Anticipated expiration: 2035-02-17
Also published as: CN104702961B

Abstract

The invention discloses a method for establishing a correlation noise model and estimating model parameters in distributed video coding. The mixed model first uses K-Mediods to divide residual coefficients into small residuals and large residuals, and uses improved Lapla The Sri Lankan distribution is used to describe the distribution of small residual coefficients, and the Cauchy distribution is used to describe the large residual coefficients. The Hybrid Distribution Correlation Noise Model (HDCNM) proposed in this paper can more accurately describe the distribution of residual coefficients between WZ frames and side information, thus effectively improving the rate-distortion performance of transform-domain distributed video coding and reducing Computational complexity of the decoding side of the system.

Description

A Bit Rate Control Method in Distributed Video Coding

技术领域technical field

本发明涉及一种在分布式视频编码中码率控制方法，属于视频压缩领域。The invention relates to a code rate control method in distributed video coding, which belongs to the field of video compression.

背景技术Background technique

分布式视频编码(Distributed Video Coding，DVC)是基于20世纪70年代Slepian.Wolf以及Wyner.Ziv提出的信息理论而建立的，将编码器运算复杂度转移到解码器。其特点是编码简单、解码较复杂、压缩性能接近传统的编码方式、抗误码能力强，适用于无线网络中资源受限的视频编码设备等。Distributed Video Coding (DVC) is based on the information theory proposed by Slepian.Wolf and Wyner.Ziv in the 1970s, which transfers the computational complexity of the encoder to the decoder. It is characterized by simple encoding, complex decoding, compression performance close to traditional encoding methods, strong anti-error capability, and is suitable for video encoding equipment with limited resources in wireless networks.

分布式视频编码中多使用Turbo码或者LDPC码这样的信道编码对WZ帧进行编码。虽然Turbo码和LDPC码都是能够接近理论界的信道编码方案，但有研究结果表明LDPC码的整体性能要好于Turbo码，对于运动剧烈的视频，LDPC方案抗误码率性能更好。在分布式视频编码中，原始WZ帧与相应边信息间相关噪声模型的准确度对编码效率有很大影响，相关噪声模型越准确，LDPC码成功解码需要的校验位就越少，一方面降低了码率，提高了压缩效率；另一方面，减少了LDPC码解码的计算量。所以，解码端对相关噪声的准确建模是分布式视频编码的一个关键技术。In distributed video coding, channel coding such as Turbo code or LDPC code is often used to code WZ frames. Although both Turbo codes and LDPC codes are channel coding schemes close to the theoretical world, some research results show that the overall performance of LDPC codes is better than that of Turbo codes. For videos with intense motion, the LDPC scheme has better anti-bit error rate performance. In distributed video coding, the accuracy of the correlation noise model between the original WZ frame and the corresponding side information has a great influence on the coding efficiency. The more accurate the correlation noise model is, the less check digits are required for successful decoding of LDPC codes. On the one hand, The code rate is reduced, and the compression efficiency is improved; on the other hand, the calculation amount of LDPC code decoding is reduced. Therefore, accurate modeling of correlated noise at the decoder is a key technology in distributed video coding.

由于子带级的拉普拉斯分布模型在计算复杂度和精确度上具有良好的折中得到了广泛认可，现今的DVC系统大多采用拉普拉斯分布来描述原始WZ帧和边信息之间的噪声关系，近年来对相关噪声模型的研究主要在于如何获得更精确的拉普拉斯参数。但DVC系统中的DCT残差系数统计分布并不完全符合拉普拉斯分布这个假设，而是具有更尖的峰值特性和更长的尾部，为适应残差系数的这两个特性，本文提出基于K-Medoids聚类的拉普拉斯-柯西混合相关噪声模型的建模方法。该混合模型利用K-Medoids将子带残差分为大残差系数和小残差系数，对小残差系数的分布使用拉普拉斯模型来描述，对大残差系数的分布使用柯西分布来描述，增加了模型的精度，从而提高了系统的率失真性能。Since the sub-band-level Laplacian distribution model has a good trade-off in computational complexity and accuracy, it has been widely recognized. Most of today's DVC systems use Laplacian distribution to describe the relationship between the original WZ frame and side information. In recent years, the research on the correlation noise model mainly focuses on how to obtain more accurate Laplacian parameters. However, the statistical distribution of DCT residual coefficients in the DVC system does not fully conform to the assumption of Laplace distribution, but has sharper peak characteristics and longer tails. In order to adapt to these two characteristics of residual coefficients, this paper proposes A Laplacian-Cauchy Mixed Correlated Noise Modeling Method Based on K-Medoids Clustering. The hybrid model uses K-Medoids to divide the sub-band residuals into large residual coefficients and small residual coefficients. The distribution of small residual coefficients is described by the Laplace model, and the distribution of large residual coefficients is described by Cauchy distribution. to describe, increasing the accuracy of the model and thus improving the rate-distortion performance of the system.

发明内容Contents of the invention

技术问题：针对变换域分布式视频编码中原始Wyner-Ziv(WZ)帧与相应边信息的残差系数存在大残差和小残差系数统计分布与传统的拉普拉斯分布存在一定偏差的问题。为了减少这种差异，提出一种基于K-Mediods的混合分布相关噪声模型及其参数估计算法。该混合模型利用改进的拉普拉斯分布描述小残差系数的分布，采用柯西分布描述大残差系数。本文提出的混合模型建模方法能较精确地描述WZ帧和边信息间的残差系数分布，从而有效地改善了变换域分布式视频编码的率失真性能，并减少系统解码端计算复杂度。Technical problem: For the residual coefficients of the original Wyner-Ziv (WZ) frame and the corresponding side information in the transform domain distributed video coding, there are large residuals and small residual coefficients. The statistical distribution of the residual coefficients has a certain deviation from the traditional Laplace distribution. question. In order to reduce this difference, a K-Mediods-based mixed distribution correlated noise model and its parameter estimation algorithm are proposed. The mixed model uses the improved Laplace distribution to describe the distribution of the small residual coefficients, and uses the Cauchy distribution to describe the large residual coefficients. The hybrid model modeling method proposed in this paper can more accurately describe the distribution of residual coefficients between WZ frames and side information, thereby effectively improving the rate-distortion performance of distributed video coding in the transform domain and reducing the computational complexity of the system decoding end.

技术方案：Technical solutions:

1.一种分布式视频编码中相关噪声模型建模方法，其特征在于，该方法包含以下步骤：1. a correlation noise model modeling method in distributed video coding, it is characterized in that, the method comprises the following steps:

1)对于当前解码子带b_k，每个系数和它与该子带残差均方的距离的绝对值组成一个2维特征向量，即子带b_k中第n个残差特征向量表示成其中 $D_{b_{k}} (μ, v) = C_{R_{XY}}^{b_{x}} (μ, v) - E (C_{R_{XY}}^{b_{k}});$ 1) For the current decoding sub-band b _k , each coefficient and the absolute value of its distance from the sub-band residual mean square form a 2-dimensional feature vector, that is, the nth residual feature vector in sub-band b _k is expressed as in ${D.}_{b_{k}} (μ, v) = C_{R_{X Y}}^{b_{x}} (μ, v) - E. (C_{R_{X Y}}^{b_{k}});$

2)利用K-Medoids聚类算法将残差特征向量分成大残差类与小残差类；2) Use the K-Medoids clustering algorithm to divide the residual feature vector into a large residual class and a small residual class;

3)使用改进的拉普拉斯分布描述小残差类，使用柯西分布描述大残差类。分别计算相应的分布参数，最终得到混合分布式相关噪声模型；3) Use the improved Laplace distribution to describe the small residual class, and use the Cauchy distribution to describe the large residual class. Calculate the corresponding distribution parameters respectively, and finally get the mixed distributed correlated noise model;

在步骤2)中，按照如下步骤进行K-Medoids聚类：In step 2), perform K-Medoids clustering according to the following steps:

1)初始化聚类中心：选择开始的三个残差特征矢量作为初始聚类中心对应的类为S₁ ^(k)，S₂ ^(k)，S₃ ^(k)，令k＝0；1) Initialize the clustering center: select the first three residual feature vectors as the initial clustering center The corresponding classes are S ₁ ^(k) , S ₂ ^(k) , S ₃ ^(k) , let k=0;

2)样本聚类：将待分类的特征向量集逐个按最小距离原则划分给三类中的某一类，即：如果 $d_{ij}^{(k)} = \min_{j} [d_{ij}^{(k)}], (i = 1,2, . . ., N), j = (1,2,3),$ 则式中表示和类的中心的距离，上角标k表示迭代次数，这里的距离选择欧氏距离，于是产生新聚类 2) Sample clustering: the feature vector set to be classified One by one according to the principle of the minimum distance into one of the three categories, that is: if $d_{ij}^{(k)} = \min_{j} [d_{ij}^{(k)}], (i = 1,2, . . ., N), j = (1,2,3),$ but In the formula express and class center of The superscript k indicates the number of iterations, and the distance here is Euclidean distance, so a new cluster is generated

3)重新计算聚类中心：与K-Means不同，K-Means选择当前cluster中所有数据点的平均值为新的中心点，而在K-Medoids中，将从当前cluster中选取一个到其他所有(当前cluster中的)点的距离之和最小的点作为中心点；3) Recalculate the cluster center: Unlike K-Means, K-Means selects the average of all data points in the current cluster as the new center point, while in K-Medoids, one will be selected from the current cluster to all other (in the current cluster) the point with the smallest sum of distances as the center point;

4)判断终止条件：如果(j＝1,2,3)，则结束，得到3个聚类，DCT残差系数被分为三个集合S₁，S₂，S₃，否则，k＝k+1，转至2)。4) Determine the termination condition: if (j=1,2,3), then end, get 3 clusters, DCT residual coefficients are divided into three sets S ₁ , S ₂ , S ₃ , otherwise, k=k+1, go to 2) .

在所述步骤3)中，计算拉普拉斯分布的参数以及柯西分布的参数的具体流程为：In said step 3), the specific process for calculating the parameters of the Laplace distribution and the parameters of the Cauchy distribution is:

聚类之后，计算三个残差系数集合S_j(j＝1,2,3)各自的方差，然后按从小到大的顺序，将其对应的残差系数集合分别记为S₁，S₂，S₃，其中S₁记为小系数集，而S₂，S₃则记为大系数集，S₁基本关于0对称，为了计算柯西分布的参数的方便，对S₁进行修正，记S₁的上界和下界的绝对值的最小值为TL，之后将用柯西分布来描述大残差系数集，而用拉普拉斯分布描述小残差系数集；After clustering, calculate the respective variances of the three residual coefficient sets S _j (j=1,2,3), and then record the corresponding residual coefficient sets as S ₁ and S ₂ respectively in ascending order , S ₃ , where S ₁ is recorded as a small coefficient set, while S ₂ and S ₃ are recorded as a large coefficient set, and S ₁ is basically symmetrical about 0. In order to facilitate the calculation of the parameters of the Cauchy distribution, S ₁ is modified and recorded as The minimum value of the absolute value of the upper and lower bounds of S ₁ is TL, and then the Cauchy distribution will be used to describe the large residual coefficient set, and the Laplace distribution will be used to describe the small residual coefficient set;

柯西分布可以表示为：The Cauchy distribution can be expressed as:

$p p ((x x)) = = \frac{11}{π π} \frac{λ λ}{{λ λ}^{22} + + {((x x - - μ μ))}^{22}}$

其中λ是形状参数，μ是位置参数，因为视频压缩编码中残差系数基本关于0对称，所以令μ＝0。为了保持最后由拉普拉斯分布和柯西分布组合表示的概率密度还能满足那么对于柯西分布，得找到一个λ使得其概率密度在[-TL,TL]区间内的积分值等于拉普拉斯的概率密度在[-TL,TL]区间内的积分值，并且仍然能维持它的重尾特性。这样就能推导出λ。Wherein λ is a shape parameter, μ is a position parameter, because the residual coefficient in video compression coding is basically symmetrical about 0, so μ=0. In order to maintain the final probability density represented by the combination of Laplace distribution and Cauchy distribution can also satisfy Then for the Cauchy distribution, it is necessary to find a λ such that the integral value of its probability density in the [-TL, TL] interval is equal to the integral value of Laplace’s probability density in the [-TL, TL] interval, and still can maintain its heavy-tailed nature. In this way, λ can be derived.

令拉普拉斯概率密度在[-TL,TL]内的积分值为P_L(TL)，它可以按下式计算：Let the integral value of the Laplace probability density in [-TL,TL] be P _L (TL), it can be calculated as follows:

${P P}_{L L} ((TL TL)) = = P P ((TL TL \leq \leq x x \leq \leq TL TL)) = = {&Integral; &Integral;}_{- - TL TL}^{TL TL} \frac{{α α}_{{b b}_{k k}}}{22} exp exp ((- - {α α}_{{b b}_{k k}} | | x x | |)) dx dx = = C C$

其中按 $α_{b_{k}} = \sqrt{2 / σ_{b_{k}}^{2}}, σ_{b_{k}}^{2} = E ({| C_{R_{XY}}^{b_{k}} |}^{2}) - E {(| C_{R_{XY}}^{b_{k}} |)}^{2}$ 计算。in according to $α_{b_{k}} = \sqrt{2 / σ_{b_{k}}^{2}}, σ_{b_{k}}^{2} = E. ({| C_{R_{X Y}}^{b_{k}} |}^{2}) - E. {(| C_{R_{X Y}}^{b_{k}} |)}^{2}$ calculate.

令P_C(LH)表示柯西概率密度在[-TL,TL]内的积分值，按下式计算：Let P _C (LH) represent the integral value of the Cauchy probability density in [-TL,TL], and it can be calculated according to the following formula:

$\begin{matrix} {P P}_{C C} ((TL TL)) = = P P ((- - TL TL \leq \leq x x \leq \leq TL TL)) = = {&Integral; &Integral;}_{- - TL TL}^{TL TL} \frac{11}{π π} \frac{λ λ}{{λ λ}^{22} + + {x x}^{22}} dx dx \\ = = \frac{22}{π π} {tan the tan}^{- - 11} ((\frac{TL TL}{λ λ})) = = {P P}_{L L} ((TL TL)) = = C C \end{matrix}$

这样就能得到： $λ = \frac{TL}{\tan (πC / 2)} .$ This gives: $λ = \frac{TL}{the tan (πC / 2)} .$

利用下式计算属于S₁残差系数集合的拉普拉斯参数：The Laplace parameter belonging to the set of S ₁ residual coefficients is calculated using the following formula:

${α α}_{11} = = \sqrt{\frac{22}{{σ σ}_{11}^{22}}}$

是S₁残差系数集合的方差。 is the variance of the set of S ₁ residual coefficients.

最终可以这样来表示相关噪声模型：Finally, the correlated noise model can be represented as follows:

$p p (({C C}_{{R R}_{XY X Y}}^{{b b}_{k k}} ((μ μ,, &upsi; &upsi;)))) = = \{\begin{matrix} \frac{{α α}_{00}}{22} exp exp ((- - {α α}_{00} | | {C C}_{{R R}_{XY X Y}}^{{b b}_{k k}} ((μ μ,, &upsi; &upsi;)) | |)),, {C C}_{{R R}_{XY X Y}}^{{b b}_{k k}} ((μ μ,, &upsi; &upsi;)) &Element; &Element; {S S}_{11} \\ \frac{11}{π π} \frac{λ λ}{{λ λ}^{22} + + {[[{C C}_{{R R}_{XY X Y}}^{{b b}_{k k}} ((μ μ,, &upsi; &upsi;))]]}^{22}},, {C C}_{{R R}_{XY X Y}}^{{b b}_{k k}} ((μ μ,, &upsi; &upsi;)) &Element; &Element; {S S}_{22},, {S S}_{33} \end{matrix}$

有益效果：本发明与现有技术相比，具有以下优点：Beneficial effect: compared with the prior art, the present invention has the following advantages:

a)本发明方案提出了一种混合相关噪声模型建模方法，在该方法中，采用K-Medoids聚类算法来提高分类的准确性与鲁棒性。使用拉普拉斯分布描述小残差的分布，使用柯西分布描述大残差的分布。提高了相关噪声模型的准确性与自适应性能。a) The solution of the present invention proposes a mixed correlation noise model modeling method, in which the K-Medoids clustering algorithm is used to improve the accuracy and robustness of classification. Use the Laplace distribution to describe the distribution of small residuals, and the Cauchy distribution to describe the distribution of large residuals. Improved accuracy and adaptive performance of correlated noise models.

b)本文方法通过在解码端的改进，有效提高了DVC系统的率失真性能。b) The method in this paper effectively improves the rate-distortion performance of the DVC system by improving the decoder.

附图说明Description of drawings

图1是现有的基于变换域的分布式视频编码框架图。FIG. 1 is a frame diagram of an existing distributed video coding based on a transform domain.

图2是本发明改进的视频编码框架图。Fig. 2 is a frame diagram of the improved video coding of the present invention.

图3是本发明的混合相关噪声模型建模方法流程图。Fig. 3 is a flow chart of the mixed correlation noise model modeling method of the present invention.

具体实施方式Detailed ways

下面结合附图与具体实施方式对本发明作进一步详细描述。The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

图1是现有的基于变换域的分布式视频编码框架图，本发明提出了一种分布式视频编码中建立相关噪声模型及估计模型参数的方法，该混合模型首先利用K-Mediods将残差系数分为小残差和大残差，利用改进的拉普拉斯分布描述小残差系数的分布，采用柯西分布描述大残差系数。该文提出的混合模型(Hybrid Distribution CorrelationNoise Model，HDCNM)能较精确地描述WZ帧和边信息间的残差系数分布，从而有效地改善了变换域分布式视频编码的率失真性能，并减少系统解码端计算复杂度。改进后的视频编码框架如图2所示。Fig. 1 is the frame diagram of the existing distributed video coding based on the transform domain. The present invention proposes a method for establishing a correlated noise model and estimating model parameters in distributed video coding. The hybrid model first utilizes K-Mediods to convert the residual The coefficients are divided into small residuals and large residuals. The improved Laplace distribution is used to describe the distribution of the small residual coefficients, and the Cauchy distribution is used to describe the large residual coefficients. The Hybrid Distribution Correlation Noise Model (HDCNM) proposed in this paper can more accurately describe the distribution of residual coefficients between WZ frames and side information, thereby effectively improving the rate-distortion performance of distributed video coding in the transform domain and reducing the system Computational complexity of the decoder. The improved video coding framework is shown in Figure 2.

下面给出本发明方法的具体实施例：Provide the specific embodiment of the inventive method below:

2)利用K-Medoids聚类算法将残差特征向量分成大残差类与小残差类2) Use the K-Medoids clustering algorithm to divide the residual feature vector into large residual class and small residual class

3)使用改进的拉普拉斯分布描述小残差类，使用柯西分布描述大残差类。分别计算相应的分布参数，最终得到混合分布式相关噪声模型。3) Use the improved Laplace distribution to describe the small residual class, and use the Cauchy distribution to describe the large residual class. The corresponding distribution parameters are calculated respectively, and finally a mixed distributed correlated noise model is obtained.

柯西分布可以表示为：The Cauchy distribution can be expressed as:

${α α}_{11} = = \sqrt{\frac{22}{{σ σ}_{11}^{22}}}$

Claims

1. a correlated noise model modelling approach in distributed video coding, it is characterized in that, the method includes the steps of:

1) for current decoded sub-band b _k, the absolute value of the distance of each coefficient and it and this subband residual mean square (RMS) forms 2 dimensional feature vectors, i.e. subband b _kin the n-th residual error characteristic vector be expressed as wherein

D_{b_{k}} (μ, v) = C_{R_{XY}}^{b_{k}} (μ, v) - E (C_{R_{XY}}^{b_{k}});

2) K-Medoids clustering algorithm is utilized residual error characteristic vector to be divided into large residual error class and little residual error class;

3) use the laplacian distribution improved to describe little residual error class, use Cauchy to distribute and describe large residual error class; Calculate corresponding distributed constant respectively, finally obtain mixed distribution formula correlated noise model.

2. bit rate control method in a kind of distributed video coding according to claim 1, is characterized in that, described step 2) in, carry out K-Medoids cluster in accordance with the following steps:

1) initialization cluster centre: select three the residual error characteristic vectors started as initial cluster center corresponding class is make k=0;

2) sample clustering: by set of eigenvectors to be sorted the a certain class in three classes is allocated to one by one, if that is: by minimal distance principle (i=1,2 ..., N), j=(1,2,3), then in formula represent and class center distance, superscript k represents iterations, and distance here selects Euclidean distance, so produce new cluster

3) cluster centre is recalculated: different from K-Means, K-Means selects the mean value of all data points in current cluster to be new central point, and in K-Medoids, the minimum point of the distance sum of (in current cluster) point to other are all will be chosen as central point from current cluster;

4) end condition is judged: if (j=1,2,3), then terminate, and obtains 3 clusters, and DCT residual error coefficient is divided into three S set ₁, S ₂, S ₃, otherwise k=k+1, goes to 2).

3. in a kind of distributed video coding according to claim 1 and 2, side information is improved one's methods, and it is characterized in that, step 3) in, the idiographic flow calculating the parameter of laplacian distribution and the parameter of Cauchy's distribution is:

After cluster, calculate three residual error coefficient S set _jthe variance that (j=1,2,3) are respective, then by order from small to large, is designated as S respectively by the residual error coefficient set of its correspondence ₁, S ₂, S ₃, wherein S ₁be designated as little coefficient set, and S ₂, S ₃then be designated as large coefficient set, S ₁substantially symmetrical about 0, in order to calculate the convenience of the parameter of Cauchy's distribution, to S ₁revise, note S ₁the upper bound and the minimum value of absolute value of lower bound be TL, large residual error coefficient collection will be described with Cauchy's distribution afterwards, and describe little residual error coefficient collection by laplacian distribution;

Cauchy's distribution can be expressed as:

p (x) = \frac{1}{π} \frac{λ}{λ^{2} + {(x - μ)}^{2}}

Wherein λ is form parameter, and μ is location parameter, because in video compression coding, residual error coefficient is substantially symmetrical about 0, so make μ=0; Can also meet to keep the probability density finally represented by laplacian distribution and Cauchy's distributed combination so for Cauchy's distribution, a λ must be found to make the integrated value of its probability density in [-TL, TL] interval equal the integrated value of probability density in [-TL, TL] interval of Laplce, and still can maintain its heavy-tailed property; So just λ can be derived;

The integrated value of Laplacian probability density in [-TL, TL] is made to be P _l(TL), it can be calculated as follows:

P_{L} (TL) = P (- TL \leq x \leq TL) = {&Integral;}_{- TL}^{TL} \frac{α_{b_{k}}}{2} \exp ({- α}_{b_{k}} | x |) dx = C

Wherein press

α_{b_{k}} = \sqrt{2 / σ_{b_{k}}^{2}}, σ_{b_{k}}^{2} = E ({| C_{R_{XY}}^{b_{k}} |}^{2}) - E {(| C_{R_{XY}}^{b_{k}} |)}^{2}

Calculate;

Make P _c(LH) represent the integrated value of Cauchy probability density in [-TL, TL], be calculated as follows:

\begin{matrix} P_{C} (TL) = P (- TL \leq x \leq TL) = {&Integral;}_{- TL}^{TL} \frac{1}{π} \frac{λ}{λ^{2} + x^{2}} dx \\ = \frac{2}{π} \tan^{- 1} (\frac{TL}{λ}) = P_{L} (TL) = C \end{matrix}

So just can obtain:

λ = \frac{TL}{\tan (πC / 2)};

Utilize following formula to calculate and belong to S ₁the Laplacian parameter of residual error coefficient set:

α_{1} = \sqrt{\frac{2}{σ_{1}^{2}}}

s ₁the variance of residual error coefficient set.

Final expression correlated noise model:

p (C_{R_{XY}}^{b_{k}} (μ, &upsi;)) = \{\begin{matrix} \frac{α_{0}}{2} \exp ({- α}_{0} | C_{R_{XY}}^{b_{k}} (μ, &upsi;) |), C_{R_{XY}}^{b_{k}} (μ, &upsi;) &Element; S_{1} \\ \frac{1}{π} \frac{λ}{λ^{2} + {[C_{R_{XY}}^{b_{k}} (μ, &upsi;)]}^{2}}, C_{R_{XY}}^{b_{k}} (μ, &upsi;) &Element; S_{2}, S_{3} \end{matrix}