CN102158703B - Distributed video coding-based adaptive correlation noise model construction system and method - Google Patents

Abstract

The invention discloses a distributed video coding-based adaptive correlation noise model construction system and a distributed video coding-based adaptive correlation noise model construction method, which belong to the technical field of distributed video coding. The system comprises a DC coefficient tape correlation noise (DCN) mean value calculation module, an original information-assisted residual error processing module, a judgment module and a mode establishment module. A coder transmits a DC coefficient mean value of an original Wyner-Ziv frame to a decoder; and the decoder constructs models of a DC coefficient tape correlation noise area and an AC coefficient tape correlation noise area respectively, and dynamically regulates the models mainly by utilizing the positioning and parameter calculation of an original information assistance model provided by the coder in combination with the motion conditions of an image so as to realize adaptive model construction. By the system and the method, the main technical problems of low accuracy and low flexibility of the prior art are solved, the correlation noise model is more accurate on the basis of not increasing the coding complexity and a transmission code rate, and the rate-distortion performance of the whole system is improved; and the system and the method can be used for a distributed video communication system.

Description

Based on distributed video coding self adaptation correlated noise modeling system and method

Technical field

The invention belongs to technical field of video coding, relate generally to distributed video coding system and correlated noise modeling system and method, specifically a kind of based on distributed video coding self adaptation correlated noise modeling system and method, can be used for the distributed video coding system.

Background technology

Traditional video encoding standard such as MPEG-X and H.26X serial mainly rely on encoder to utilize the statistical property of code signal to realize compressed encoding, are more than 5 to 10 times of decoder thereby cause the computational complexity of encoder.In recent years, some new Video Applications such as wireless video sensor network, mobile video telephone and wireless video monitoring etc. are in the life that incorporates rapidly and changing people.Yet, these emerging multimedia application have proposed and in the past different demands the coding/decoding system of video, be that encoder must be simple as much as possible because of resource-constrained, and decoder has the decoding computing that more resource can be carried out complexity, and this structural design to video coding and decoding system has proposed new challenge.

The distributed video coding system, namely the Wyner-Ziv video coding system has the system configuration opposite with traditional coding standard, fits like a glove with above-mentioned application demand.It is theoretical that 20 century 70s, Slepian and Wolf have proposed the distributed source lossless coding, and Wyner and Ziv have proposed to diminish distributed source coding subsequently, and these two theories have been established the basis of Wyner-Ziv Video coding.Compare with traditional video coding system, the distributed video coding system has transferred to decoding end with main high computational complexity module from coding side, utilizes the correlation of information source to realize efficient compressed encoding by decoder.Distributed video coding has been widely used in reality, the modal wireless video sensor network that is applied as, the supervisory control system that namely is seen everywhere, the camera that is distributed in everywhere is equivalent to encoding device, they carry out absolute coding, Surveillance center then is equivalent to decoding device, carries out combined decoding.

Referring to Fig. 2, the distributed video coding system is divided into key frame and original Wyner-Ziv frame with input video sequence, and both carry out absolute coding at coding side, the decoding end combined decoding.Wherein key frame adopts H.264 intraframe coding and decoding, and for the Wyner-Ziv frame, at first carries out discrete cosine transform and extraction coefficient band, and coefficient tape is carried out Zigzag scanning ordering; Then to the coefficient tape uniform quantization, quantization parameter is extracted bit-plane, bit-plane is sent into respectively low-density checksum LDPC encoder encode.During decoding, at first key frame is carried out H.264 intraframe decoder, utilize former and later two key frames that decode, produce side information SI frame by motion compensated interpolation; Then the SI frame is carried out DCT identical with coding side, scanning, quantification and bit-plane and extract, and send into the LDPC decoder; Dash area is correlated noise MODEL C NM tectonic system among Fig. 2, and this system utilizes the motion compensated residual frame Z computation model parameter that produces in the SI frame generative process and constructs CNM, and CNM decodes for LDPC and inverse quantization reconstruct provides information; The check digit that the utilization of LDPC decoder is received, corresponding sides information coefficient band and CNM begin to decode successively from the highest order plane to each coefficient tape; Next, LDPC decoder Output rusults is obtained successively the decoding and reconstituting image of Wyner-Ziv frame by merge bit plane, inverse quantization and inverse transformation.

In the distributed video coding system, correlated noise refers to the residual error between the DCT coefficient tape of the DCT coefficient tape of Wyner-Ziv frame and SI frame, existing distributed video coding system all adopts 4 * 4DCT conversion, so the DCT coefficient tape comprises 1 direct current DC coefficient tape and exchange the AC coefficient tape with 15, corresponding correlated noise comprises DC coefficient tape correlated noise DCN and AC coefficient tape correlated noise ACN.Correlated noise MODEL C NM is the Distribution Statistics that correlated noise is obeyed.

CNM is a key factor that affects the distributed video coding systematic function, if CNM can accurately simulate the distribution of residual error, then the bit-plane confidence calculations is accurate, and then decoding and reconstituting is more accurate, and system's rate distortion will increase; Otherwise system's distortion performance will descend to some extent.

Existing distributed video coding system all adopt location parameter be zero laplacian distribution as CNM, concrete model is as follows:

$f\left(n\right)=\frac{1}{2b}{e}^{-\frac{\left|n\right|}{b}}---\left(1\right)$

Wherein n represents the correlated noise value, the probability density function of f (n) expression correlated noise n, and b is the scale parameter of laplacian distribution.This model has following shortcoming: the first, because decoding end does not have the information of original image, this model is to utilize the estimated information of decoding end to obtain fully, therefore accuracy is limited; The second, because the importance of DCN and ACN is different with statistical property, this model is not distinguished the two and is treated, but adopts identical model; The 3rd, because the distribution of DCN is relevant with the image motion situation, this model adopts the distribution of fixing distribution simulation DCN, fails dynamically to adjust very flexible according to the image motion situation.These shortcomings have all affected the accuracy of CNM, cause the distortion performance of whole system to descend.

Summary of the invention

The object of the invention is to overcome the shortcoming of above-mentioned prior art, propose a kind of based on distributed video coding self adaptation correlated noise modeling system and method, to improve the accuracy of correlated noise model, for decoding end bit-plane confidence calculations and decoding and reconstituting provide more accurately information, and then improve the distortion performance of whole system.

For achieving the above object, the invention provides a kind of based on distributed video coding self adaptation correlated noise modeling system, comprise the residual noise reduction module, this module is output as to be set up the correlated noise model information is provided, it is characterized in that: described residual noise reduction module is the residual noise reduction module that has raw information auxiliary, this tectonic system also includes: DCN mean value computation module, determination module and set up model module, wherein DCN mean value computation module will calculate the average μ of residual frame direct current DC coefficient, the average that is DC coefficient tape correlated noise DCN sends to the auxiliary residual noise reduction module of raw information and sets up model module, there is the auxiliary residual noise reduction module of raw information first residual frame Z to be carried out DCT and extraction coefficient band, calculate again the quadratic sum SQC of all coefficients, and at lower variance and the scale parameter that calculates coefficient of assisting of average μ, the variance of coefficient and scale parameter are admitted to sets up model module, SQC is admitted to determination module, determination module carries out whether violent judgement of image motion in conjunction with SQC, result of determination outputs to sets up model module, set up model module in conjunction with average μ, the variance of result of determination and coefficient and scale parameter are according to the Formula self adaptation correlated noise model of Gaussian Profile and laplacian distribution; Above-mentioned DCN mean value computation module comprises coding side mean value computation unit ECU, decoding end mean value computation cells D CU and side-play amount computing unit BCU, and wherein ECU calculates the average μ of the DC coefficient of original Wyner-Ziv frame at coding side _WZ, DCU calculates the average μ of the DC coefficient of side information SI frame in decoding end _SI, μ _WZWith μ _SIBe sent to BCU, BCU makes poor obtain DCN average μ, i.e. μ=μ to the two _WZ-μ _SI

Prior art is owing to utilizing the estimated information structure CNM of decoding end fully, causes between the distribution of CNM and actual correlated noise error larger, and the structure that the present invention utilizes coding side raw information to assist CNM, improved the accuracy of CNM.Also according to the relation between image motion situation and the DC coefficient tape correlated noise, dynamic adjustment model has been realized adaptive Construction of A Model in the present invention, has overcome the shortcoming of prior art very flexible.

Realization of the present invention is that also the described residual noise reduction module that has raw information to assist comprises: converter unit, the SQC computing unit, B parameter calculation unit and C parameter calculation unit, residual frame Z obtains 16 coefficient tapes through the processing of converter unit, the B parameter calculation unit is calculated first variance and the scale parameter of each AC coefficient tape, calculate again variance and the scale parameter of DC coefficient tape in conjunction with average μ, then the variance of all coefficient tapes and scale parameter are sent to the C parameter calculation unit and calculate, the variance of the coefficient that obtains and scale parameter are sent to sets up model module, process 16 coefficient tapes that obtain by converter unit and also be sent to the SQC computing unit, the SQC that calculates is sent to determination module.

Prior art utilizes the estimated information of decoding end to calculate variance and scale parameter fully, there be not the auxiliary of raw information, and the present invention has utilized the raw information of coding side to assist the operation of B parameter calculation unit and C parameter calculation unit, therefore improved DC coefficient tape correlated noise model parameter accuracy of computation.

Realization of the present invention also be to set up model module in conjunction with the result of determination of the average μ of DCN mean value computation module output, determination module output, variance and the scale parameter of the coefficient of the auxiliary residual noise reduction module output of raw information are arranged, according to the Formula correlated noise model of Gaussian Profile and laplacian distribution.

Prior art adopts fixing and single model, can't be according to the concrete condition of video sequence adjustment model flexibly, and the adaptive tectonic model of motion conditions of the comprehensive raw information of the present invention and image, so that model can dynamically be adjusted according to concrete video content, improve flexibility and the accuracy of model.

The present invention or a kind of based on distributed video coding self adaptation correlated noise model construction method, applicable system shown in the method be the present invention propose based on distributed video coding self adaptation correlated noise modeling system, the method comprises the steps:

Step 1. is calculated the average μ of residual frame DC coefficient, the i.e. average of DC coefficient tape correlated noise DCN;

Step 2. utilizes the residual frame Z after decoded former and later two key frame motion compensation to replace residual frame between original Wyner-Ziv frame and the SI frame, and Z is carried out 4 * 4 discrete cosine transforms and extraction coefficient band, obtains 16 residual error coefficient band T _i, i=1～16;

Step 3. is calculated first the quadratic sum SQC of all coefficients, calculates variance and the scale parameter of each AC coefficient tape again, then in conjunction with average μ, calculates variance and the scale parameter of DC coefficient tape, calculates at last variance and the scale parameter of each coefficient;

Step 4. is determined thresholding N, and compares SQC and thresholding N, if SQC＞N then judges current Wyner-Ziv frame motion acutely, and upgrades thresholding with α N+ (1-α) SQC, otherwise, judge that current Wyner-Ziv frame motion is mild, do not upgrade thresholding;

Step 5. is for the AC coefficient tape of residual frame, be AC coefficient tape correlated noise ACN, the chosen position parameter is that zero laplacian distribution is as model, DC coefficient tape for residual frame, be that DC coefficient tape correlated noise DCN is according to the adaptive tectonic model of court verdict, if current Wyner-Ziv frame motion is violent, the laplacian distribution of chosen position parameter non-zero is as model, if current Wyner-Ziv frame motion is mild, select the Gaussian Profile of average non-zero as model, and the average μ that obtains according to step 1 proofreaies and correct the skew that DCN occurs, and realizes the accurate location of correlated noise MODEL C NM, and is specific as follows:

Wherein n represents the correlated noise value, and f (n) expression correlated noise n is at the probability density function of transform domain, μ _lBe respectively location parameter and the scale parameter of laplacian distribution, μ with b _gAnd σ ²Be respectively average and the variance of Gaussian Profile.

Contrasting prior art and model of the present invention can find, the present invention not only is penetrated into raw information parameter b and the σ of model ²Calculating in, also utilize raw information to come model is accurately located, the present invention also carries out refinement according to the importance of coefficient tape and the motion conditions of image to model.

Realization of the present invention is that also the calculating of step 1 average μ comprises:

5.1) coding side calculates the average μ of original Wyner-Ziv frame DC coefficient tape _WZAnd be sent to decoding end;

5.2) decoding end calculates the average μ of SI frame DC coefficient tape _SI

5.3) μ that sends of coding side _WZμ with decoding end _SIDiffer from, obtain the average μ of DCN, i.e. μ=μ _WZ-μ _SI

Because average μ obtains according to the average of the original Wyner-Ziv frame of coding side DC coefficient tape and the mean value computation of decoding end SI frame DC coefficient tape, so it is the average of actual correlated noise, it is an exact value, it is not estimated value, average μ is used for follow-up model parameter calculating and location, is conducive to improve the accuracy of model.

Realization of the present invention is that also step 3 residual noise reduction comprises:

6.1) calculate the quadratic sum SQC of all coefficients;

6.2) calculate the average of each AC coefficient tape of residual frame

Variance

And scale parameter

I=2～16 wherein, M is coefficient tape length;

6.3) utilize average μ, calculate the DC coefficient tape T of residual frame ₁Variance

And scale parameter $b_1=\sqrt{{\mathrm{\σ}}_1^2/2};$

6.4) do not do any correction for the variance of coefficient tape, directly with its worthwhile variance of doing each coefficient in this coefficient tape, be used for the structure of model, revise the scale parameter b that obtains each coefficient for the coefficient tape scale parameter _i(m), concrete modification method is as follows: i=1～16

$b_i\left(m\right)=\left\{\begin{array}{cc}{b}_i& {\left[D_i\left(m\right)\right]}^2\≤{{\mathrm{\σ}}_i}^2\\ \sqrt{\frac{{\left[D_i\left(m\right)\right]}^2}{2}}& {\left[D_i\left(m\right)\right]}^2>{{\mathrm{\σ}}_i}^2\end{array}\right.---\left(3\right)$

Wherein i represents the sequence number of coefficient tape, i=1～16, b _i(m) be the scale parameter of m coefficient of i coefficient tape, b _iThe scale parameter that represents i coefficient tape,

The variance that represents i coefficient tape, D _i(m)=| T _i(m) |-E[T _i], | T _i(m) | represent the absolute value of m coefficient of i coefficient tape, E[T _i] expression i coefficient tape average, D _i(m) expression | T _i(m) | with the average E[T of coefficient tape i under it _i] between distance;

The present invention has designed based on distributed video coding self adaptation correlated noise modeling system, provides based on distributed video coding self adaptation correlated noise model construction method.

The present invention compared with prior art has following advantage:

(1) the present invention utilizes the information of coding side original image to come the calculating of auxiliary decoder end model parameter, has solved prior art owing to the information that lacks original image makes the larger technical problem of Errors;

(2) the present invention adopts different distributions as model according to the importance of DCN and ACN and the difference of statistical property to DCN and ACN, has solved prior art DCN and ACN are carried out the low technical problem of accuracy that nothing differentiation Construction of A Model brings;

(3) the present invention is according to the distribution of DCN and the relation between the image motion situation, dynamic adjustment model, if current Wyner-Ziv frame motion is violent, adopt the laplacian distribution of location parameter non-zero as model, if current Wyner-Ziv frame motion is mild, adopt the Gaussian Profile of average non-zero as model, solved prior art and adopted fixing and the very flexible that single model brings, the technical problem that model can not dynamically be adjusted with the image motion situation; In addition, the present invention utilizes the location of the auxiliary DC coefficient tape correlated noise model of information of original image, has further improved the accuracy of model.

Description of drawings

Fig. 1 is the formation block diagram that the present invention is based on distributed video coding self adaptation correlated noise modeling system;

Fig. 2 is without feedback distributed video coding system block diagram;

Fig. 3 is algorithm flow schematic diagram of the present invention;

Fig. 4 is the correlated noise MODEL C NM accuracy comparison diagram of the present invention and prior art, and wherein Fig. 4 (a) is the CNM accuracy comparison diagram take foreman as cycle tests, and Fig. 4 (b) is the CNM accuracy comparison diagram take coastguard as cycle tests;

Fig. 5 is the system's distortion performance comparison diagram that adopts the present invention and prior art, wherein Fig. 5 (a) is the system's distortion performance comparison diagram take foreman as cycle tests, and Fig. 5 (b) is the system's distortion performance comparison diagram take coastguard as cycle tests.

Embodiment

Embodiment 1

With reference to Fig. 1, the present invention is based on distributed video coding self adaptation correlated noise modeling system, comprise the residual noise reduction module, this module is output as to be set up the correlated noise model information is provided, residual noise reduction module of the present invention is the residual noise reduction module that has raw information auxiliary, and prior art is directly to utilize the estimated information tectonic model of decoding end, does not have the auxiliary of raw information.Tectonic system of the present invention also includes: DCN mean value computation module, determination module and set up model module, wherein DCN mean value computation module will calculate the average μ of residual frame direct current DC coefficient, the average that is DC coefficient tape correlated noise DCN sends to the auxiliary residual noise reduction module of raw information and sets up model module, there is the auxiliary residual noise reduction module of raw information first residual frame Z to be carried out DCT and extraction coefficient band, calculate again the quadratic sum SQC of all coefficients, and at lower variance and the scale parameter that calculates coefficient of assisting of average μ, the variance of coefficient and scale parameter are admitted to sets up model module, SQC is admitted to determination module, determination module carries out whether violent judgement of image motion in conjunction with SQC, result of determination outputs to sets up model module, set up model module in conjunction with average μ, the variance of result of determination and coefficient and scale parameter are according to the Formula self adaptation correlated noise model of Gaussian Profile and laplacian distribution; Above-mentioned DCN mean value computation module comprises coding side mean value computation unit ECU, decoding end mean value computation cells D CU and side-play amount computing unit BCU, and wherein ECU calculates the average μ of the DC coefficient of original Wyner-Ziv frame at coding side _WZ, DCU calculates the average μ of the DC coefficient of side information SI frame in decoding end _SI, μ _WZWith μ _SIBe sent to BCU, BCU makes poor obtain DCN average μ, i.e. μ=μ to the two _WZ-μ _SI

Look in the coded system distributed, decoding end does not have original Wyner-Ziv frame, only has side information SI frame, so can't obtain the average μ of actual DC coefficient tape correlated noise DCN, but in order to improve the CNM accuracy, the present invention requires decoding end to utilize the structure of average μ submodel, requires coding side to transmit the DC Coefficient Mean μ of original Wyner-Ziv frame for this reason _WZTo decoding end, decoding end is utilized μ _WZDC Coefficient Mean μ with the SI frame _SIMake the poor μ that obtains, μ is used to the auxiliary residual noise reduction module of raw information and sets up model module.Because coding side originally will calculate μ _WZTherefore the present invention does not increase the coding side computational complexity, just each frame of video is transmitted an average more, and the transmission code rate of increase can be ignored.

Of the present invention have the auxiliary residual noise reduction module of raw information to replace residual error between original Wyner-Ziv frame and the SI frame with the residual frame Z of the motion compensation of decoded former and later two key frames, residual frame Z obtains 16 coefficient tapes through the processing of converter unit, the B parameter calculation unit is calculated first variance and the scale parameter of each AC coefficient tape, calculate again variance and the scale parameter of DC coefficient tape in conjunction with average μ, then the variance of all coefficient tapes and scale parameter are sent to the C parameter calculation unit and calculate, the variance of the coefficient that obtains and scale parameter are sent to sets up model module, process 16 coefficient tapes that obtain by converter unit and also be sent to the SQC computing unit, the SQC that calculates is sent to determination module.

Because the variance of coefficient and the calculating of scale parameter have raw information to participate in, so that the accuracy of the two is improved, use such variance and scale parameter tectonic model, have also just improved the accuracy of model.

Determination module of the present invention is used for determining that the Wyner-Ziv two field picture changes thresholding N, and the SQC and the thresholding N that relatively have the auxiliary residual noise reduction unit of raw information to export, judge whether current Wyner-Ziv two field picture changes obvious, result of determination delivered to set up model module, then upgrade thresholding N according to court verdict.

The model module of setting up of the present invention is in conjunction with the average μ of DCN mean value computation module output, the result of determination of determination module output, variance and the scale parameter that has the auxiliary residual noise reduction module of raw information to export, according to the Formula correlated noise model of Gaussian Profile and laplacian distribution; It is that zero laplacian distribution is as model that ACN is selected average, DCN is then set up model according to the result of determination of determination module output, when image change is violent, the laplacian distribution of chosen position parameter non-zero is as the model of DCN, when image change is mild, select the Gaussian Profile of average non-zero as the model of DCN, model parameter is by DCN mean value computation module and have the auxiliary residual noise reduction module of raw information to provide.

What the present invention proposed is between the generation and decoding and reconstituting of side information of decoding end based on distributed video coding self adaptation correlated noise modeling system, on the basis that does not increase coding side complexity and transmission code rate, improved the accuracy of CNM, for the decoding and reconstruct more reliable information is provided.

Embodiment 2

Based on distributed video coding self adaptation correlated noise modeling system with embodiment 1, referring to Fig. 1, specialized designs of the present invention DCN mean value computation module and the auxiliary residual noise reduction module of raw information is arranged, DCN mean value computation module comprises coding side mean value computation unit ECU, decoding end mean value computation cells D CU and side-play amount computing unit BCU, wherein:

Coding side mean value computation unit ECU: the average μ that is used for the DC coefficient of the original Wyner-Ziv frame of calculation code end _WZ, and with μ _WZBe sent to the side-play amount computing unit of decoding end;

Decoding end mean value computation cells D CU: the average μ that is used for the DC coefficient of calculating decoding end side information SI frame _SI, and with μ _SIBe sent to the side-play amount computing unit;

Side-play amount computing unit BCU: the μ that utilizes input _WZWith μ _SIMake the poor side-play amount μ that obtains DC coefficient tape noise DCN;

Have the auxiliary residual noise reduction module of raw information to comprise: converter unit, SQC computing unit, B parameter calculation unit and C parameter calculation unit, wherein:

Converter unit: replace residual error between original Wyner-Ziv frame and the SI frame with the residual frame Z after decoded former and later two key frame motion compensation, DCT and extraction coefficient band to Z carries out 4 * 4 obtain coefficient tape T _i, i=1～16;

The SQC computing unit: to first square of all coefficients again summation obtain coefficient quadratic sum SQC;

B parameter calculation unit: calculate first variance and the scale parameter of each AC coefficient tape, calculate again variance and the scale parameter of DC coefficient tape in conjunction with average μ;

C parameter calculation unit: variance and scale parameter that variance and the scale parameter correction of coefficient tape obtained coefficient.

These two modules have effectively incorporated raw information, so that the parameter that is used for model construction is more near actual value.

Embodiment 3

Referring to Fig. 3, institute's applicable system is with embodiment 1-2 based on distributed video coding self adaptation correlated noise model construction method as a kind of in the present invention, and this building method comprises the steps:

Step 1. is calculated the average μ of residual frame DC coefficient, i.e. the average of DC coefficient tape correlated noise DCN, and concrete mean value computation step is as follows:

5.1) coding side calculates the average μ of Wyner-Ziv frame DC coefficient tape _WZAnd be sent to decoding end;

5.2) decoding end calculates the average μ of SI frame DC coefficient tape _SI

5.3) the average μ of the Wyner-Ziv frame DC coefficient tape that will send from coding side _WZAverage μ with corresponding edge information D C coefficient tape _SISubtract each other, obtain the side-play amount μ of relevant correlated noise distributed model=μ _WZ-μ _SI

Step 2. utilizes the residual frame Z after decoded former and later two key frame motion compensation to replace residual frame between original Wyner-Ziv frame and the SI frame, and Z is carried out 4 * 4 discrete cosine transforms and extraction coefficient band, obtains 16 residual error coefficient band T _i, i=1～16, the expression formula of residual frame Z is as follows:

$Z(x,y)=\frac{f_f(x+{\mathrm{dx}}_f,y+{\mathrm{dy}}_f)-f_b(x+{\mathrm{dx}}_b,y+{\mathrm{dy}}_b)}{2}---\left(4\right)$

(x, y) expression location of pixels in the formula, f _fAnd f _bDecoded two key frames before and after being, dx _f, dy _fAnd dx _b, dy _bThe motion vector that obtains during to motion compensation before and after being respectively is horizontal, ordinate; Residual frame Z carried out 4 * 4 dct transform and extraction coefficient band, obtain 16 residual error coefficient band T _i, i=1～16.

Step 3 is calculated the quadratic sum SQC of all coefficients, the variance of each coefficient and scale parameter, and concrete calculation procedure is as follows:

6.1) calculate the quadratic sum SQC of all coefficients;

6.2) calculate the average of each AC coefficient tape of residual frame

Variance

And scale parameter I=2～16 wherein, M is coefficient tape length;

6.3) utilize average μ, calculate the DC coefficient tape T of residual frame ₁Variance

And scale parameter

6.4) do not do any correction for the variance of coefficient tape, directly with its worthwhile variance of doing each coefficient in this coefficient tape, be used for the structure of model, revise the scale parameter b that obtains each coefficient for the coefficient tape scale parameter _i(m), concrete correction formula is as shown in the formula shown in (3);

Step 4. is determined thresholding N, and compares SQC and thresholding N, if SQC＞N then judges current Wyner-Ziv frame motion acutely, and upgrades thresholding with α N+ (1-α) SQC, otherwise, judge that current Wyner-Ziv frame motion is mild, do not upgrade thresholding;

The AC coefficient tape of step 5. pair residual frame, be AC coefficient tape correlated noise ACN, the chosen position parameter is that zero laplacian distribution is as model, DC coefficient tape to residual frame, be DC coefficient tape correlated noise DCN, as a result preference pattern according to judging module, as SQC＞N, the laplacian distribution of chosen position parameter non-zero is as the DCN model, as SQC＜N, select the Gaussian Profile of non-zero as model, and according to the average μ that step 1 obtains the skew that DC coefficient tape correlated noise occurs is proofreaied and correct, the accurate location of implementation model, the present invention finds that by a large amount of experiments and analysis the distribution of DCN is to have skew, be not to be distributed near the null value, so the present invention takes this factor of side-play amount into account when setting up CNM.Concrete relevant correlated noise model is shown in (2) formula.

Embodiment 4

With embodiment 3, the system that is suitable for is with embodiment 1-2 based on distributed video coding self adaptation correlated noise model construction method.Because the distribution situation of DC coefficient tape correlated noise DCN is relevant with the motion conditions of corresponding Wyner-Ziv frame, the motion conditions of Wyner-Ziv frame is different, the distribution situation of corresponding DCN is also different, so need to judge the motion conditions of current Wyner-Ziv frame.

Consider that original Wyner-Ziv frame is in the decoding end non-availability, and side information SI frame can be regarded as the version of original Wyner-Ziv frame " band noise ", therefore the present invention is similar to the motion conditions of weighing the Wyner-Ziv frame with the content change degree between SI frame and adjacent two key frames.Again because the movement compensating frame f of previous key frame _f(x+dx _f, y+dy _f) and the SI frame between residual error f _f(x+dx _f, y+dy _f)-SI (x, y) has reflected the content change between SI frame and previous key frame; In like manner, the movement compensating frame f of SI frame and a rear key frame _b(x+dx _f, y+dy _f) between residual error SI (x, y)-f _b(x+dx _b, y+dy _b) reflected the content change between SI frame and a rear key frame; Content change to this two aspect is averaged and asks quadratic sum, and the motion conditions that obtains current Wyner-Ziv frame is

Because adopt 4 * 4 integer DCT based on the distributed video coding system, and this conversion is unitary transformation, follows Paasche Wa Er theorem, namely the pixel domain ability equates with the transform domain ability, is formulated to be following formula:

$\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{\left[\frac{f_f(x+{\mathrm{dx}}_f,y+{\mathrm{dy}}_f)-f_b(x+{\mathrm{dx}}_f,y+{\mathrm{dy}}_f)}{2}\right]}^2$

$=\underset{x}{\mathrm{\Σ}}\underset{y}{\mathrm{\Σ}}{Z}^2(x,y)---\left(5\right)$

$=\underset{j=1}{\overset{M}{\mathrm{\Σ}}}\underset{i=1}{\overset{15}{\mathrm{\Σ}}}{\mathrm{AC}}_{i,j}^2+\underset{l=1}{\overset{M}{\mathrm{\Σ}}}{\mathrm{DC}}_j^2$

AC in the formula _{I, j}J coefficient of expression Z i AC coefficient tape on transform domain, DC _jExpression Z j DC coefficient on transform domain,

Quadratic sum for coefficient represents with SQC.By above analysis as can be known, the quadratic sum SQC of decoding end availability coefficient weighs the motion conditions of Wyner-Ziv frame, and variation that like this can the perception video content does not increase again the extra computation amount.If SQC＞N judges that current Wyner-Ziv frame motion is violent; Otherwise, represent that current Wyner-Ziv frame motion is mild.In order to guarantee the accuracy of CNM, the N value should the adaptive change along with the variation of video image motion situation, when judging that current Wyner-Ziv frame motion is violent, just upgrade thresholding with α N+ (1-α) SC, when judging that current Wyner-Ziv frame motion is mild, then do not upgrade thresholding, so that thresholding can dynamically be adjusted, be unlikely to again fluctuation too large like this.Wherein α is empirical value, and its span is [0,1], and this routine value is 0.3.

Embodiment 5

With reference to Fig. 4, based on distributed video coding self adaptation correlated noise model construction method with embodiment 3-4, the present invention and prior art are respectively applied to nothing feedback distributed video coding system shown in Figure 2, estimate as the accuracy of CNM with the error sum of squares SSE between DC coefficient tape correlated noise model and the distribution of actual DC coefficient tape correlated noise, the accuracy that compares the two, concrete experiment condition is as follows: image sets GOP length is 2, be that sequence number is that the image of even number is key frame, sequence number is that the image of odd number is the Wyner-Ziv frame, block size is 4 * 4 pixels, cycle tests is foreman and coastguard, video format is QCIF, be that resolution is 176 * 144, experimental result as shown in Figure 4, the SSE expression formula is as follows:

$\mathrm{SSE}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(f\left(x_i\right)-y_i)}^2---\left(6\right)$

X in the formula _iBe the DC coefficient tape correlated noise value of reality, f (x _i) be that CNM is at x _iThe probability density at place, and y _iFor actual DC coefficient tape correlated noise is distributed in x _iThe probability density at place.

Fig. 4 (a) is the CNM accuracy comparative result take foreman as cycle tests, as seen from the figure, and compared with prior art, image for 90%, the SSE of the present invention obviously SSE than prior art is low, and for remaining 10% image, SSE of the present invention is a little less than the SSE of prior art.Fig. 4 (b) is with the CNM accuracy comparative result of coastguard as cycle tests, as seen from the figure, and compared with prior art, image for 95%, the SSE of the present invention obviously SSE than prior art is low, and for remaining 5% image, SSE of the present invention is a little less than the SSE of prior art.

The foreman sequence is the more violent sequence of typical motion, the coastguard sequence is the milder sequence of typical motion, by above experimental result as can be known, no matter be that motion is violent or move mild sequence, CNM accuracy of the present invention all will be higher than the CNM accuracy of prior art.

Embodiment 6

With reference to Fig. 5, based on distributed video coding self adaptation correlated noise model construction method with embodiment 3-4, the present invention and prior art are respectively applied to nothing feedback distributed video coding system shown in Figure 2, recover Y-PSNR PSNR the estimating as system's distortion performance of Wyner-Ziv frame with decoding, the system's distortion performance that relatively adopts the present invention and adopt prior art to obtain, concrete experiment condition is identical with embodiment 5, and experimental result as shown in Figure 5.

Fig. 5 (a) is the system's distortion performance comparison diagram take foreman as cycle tests, as known in the figure, compared with prior art, system's distortion performance of using the present invention to obtain has improved approximately 0.5dB, Fig. 5 (b) is that the system's distortion performance take coastguard as cycle tests compares, as known in the figure, compared with prior art, system's distortion performance of using the present invention to obtain has improved approximately 1dB.

The foreman sequence is the more violent sequence of typical motion, the coastguard sequence is the milder sequence of typical motion, by above experimental result as can be known, no matter be that motion is violent or move mild sequence, system's distortion performance of using the present invention to obtain all is better than system's distortion performance of using prior art.

The present invention mainly solves the technical problem of the low and very flexible of prior art accuracy, do not increasing on encoder complexity and the transmission code rate basis, make the correlated noise model more accurate, improved the distortion performance of whole system, can be used for the distributed video communication system.

Claims (2)

1. one kind based on distributed video coding self adaptation correlated noise modeling system, comprise the residual noise reduction module, this module is output as to be set up the correlated noise model information is provided, it is characterized in that: described residual noise reduction module is the residual noise reduction module that has raw information auxiliary, and this tectonic system also includes DCN mean value computation module, determination module and sets up model module; Wherein DCN mean value computation module will calculate the average μ of residual frame direct current DC coefficient, namely the average of DC coefficient tape correlated noise DCN sends to the auxiliary residual noise reduction module of raw information and sets up model module; There is the auxiliary residual noise reduction module of raw information first the residual frame Z after decoded former and later two key frame motion compensation to be carried out DCT and extraction coefficient band, calculate again the quadratic sum SQC of all coefficients, and at lower variance and the scale parameter that calculates coefficient of assisting of average μ, the variance of coefficient and scale parameter are admitted to sets up model module, and SQC is admitted to determination module; Determination module carries out whether violent judgement of image motion in conjunction with SQC, and result of determination outputs to sets up model module; Set up model module in conjunction with average μ, the variance of result of determination and coefficient and scale parameter, Formula self adaptation correlated noise model according to Gaussian Profile and laplacian distribution, wherein AC coefficient tape correlated noise ACN being selected average is that zero laplacian distribution is as model, DCN is then set up model according to the result of determination of determination module output, when image motion is violent, the laplacian distribution of chosen position parameter non-zero is as the model of DCN, when image motion is mild, select the Gaussian Profile of average non-zero as the model of DCN, model parameter is by DCN mean value computation module and have the auxiliary residual noise reduction module of raw information to provide; Above-mentioned DCN mean value computation module comprises coding side mean value computation unit ECU, decoding end mean value computation cells D CU and side-play amount computing unit BCU, and wherein ECU calculates the average μ of the DC coefficient of original Wyner-Ziv frame at coding side _WZ, DCU calculates the average μ of the DC coefficient of side information SI frame in decoding end _SI, μ _WZWith μ _SIBe sent to BCU, BCU makes poor obtain DCN average μ, i.e. μ=μ to the two _WZ-μ _SIDescribed have the auxiliary residual noise reduction module of raw information to comprise: converter unit, the SQC computing unit, B parameter calculation unit and C parameter calculation unit, residual frame Z obtains 16 coefficient tapes through the processing of converter unit, the B parameter calculation unit is calculated first variance and the scale parameter of each AC coefficient tape, calculate again variance and the scale parameter of DC coefficient tape in conjunction with average μ, then the variance of all coefficient tapes and scale parameter are sent to the C parameter calculation unit and calculate, the variance of the coefficient that obtains and scale parameter are sent to sets up model module, process 16 coefficient tapes that obtain by converter unit and also be sent to the SQC computing unit, the SQC that calculates is sent to determination module.

2. one kind based on distributed video coding self adaptation correlated noise model construction method, is applicable to system claimed in claim 1, and it is characterized in that: the method comprises:

Step 1. is calculated the average μ of residual frame DC coefficient, i.e. the average of DC coefficient tape correlated noise DCN, and the calculating of average μ comprises:

1.1) coding side calculates the average μ of Wyner-Ziv frame DC coefficient _WZAnd be sent to decoding end;

1.2) decoding end calculates the average μ of side information SI frame DC coefficient _SI

1.3) μ that sends of coding side _WZμ with decoding end _SIDiffer from, obtain the average μ of DCN, i.e. μ=μ _WZ-μ _SI

Step 2. utilizes the residual frame Z after decoded former and later two key frame motion compensation to replace residual frame between original Wyner-Ziv frame and the SI frame, and Z is carried out 4 * 4 discrete cosine transforms and extraction coefficient band, obtains 16 residual error coefficient band T _i, i=1～16;

Step 3. is calculated the quadratic sum SQC of all coefficients, the variance of each coefficient and scale parameter, and calculating comprises:

3.1) calculate the quadratic sum SQC of all coefficients;

3.2) calculate the average of each AC coefficient tape of residual frame Z

Variance

And scale parameter

I=2～16 wherein, M is coefficient tape length;

3.3) utilize average μ, calculate the DC coefficient tape T of residual frame Z ₁Variance

And scale parameter

3.4) do not do any correction for the variance of coefficient tape, directly with its worthwhile variance of doing each coefficient in this coefficient tape, be used for the structure of model, revise the scale parameter b that obtains each coefficient for the coefficient tape scale parameter _i(m), concrete modification method is as follows: i=1～16

Wherein i represents the sequence number of coefficient tape, i=1～16, b _i(m) be the scale parameter of m coefficient of i coefficient tape, b _iThe scale parameter that represents i coefficient tape,

The variance that represents i coefficient tape, D _i(m)=| T _i(m) |-E[T _i], | T _i(m) | represent the absolute value of m coefficient of i coefficient tape, E[T _i] expression i coefficient tape average, D _i(m) expression | T _i(m) | with the average E[T of coefficient tape i under it _i] between distance;

Step 4. is determined thresholding N, and comparison SQC and thresholding N, if SQC＞N judges that then current Wyner-Ziv frame motion is violent, and upgrades thresholding with α N+ (1-α) SQC, otherwise, judge that current Wyner-Ziv frame motion is mild, do not upgrade thresholding, wherein α is empirical value, its span is [0,1];

Step 5. is for the AC coefficient tape of residual frame, be AC coefficient tape correlated noise ACN, the chosen position parameter is that zero laplacian distribution is as model, DC coefficient tape for residual frame, be DC coefficient tape correlated noise DCN, according to the adaptive tectonic model of court verdict, if current Wyner-Ziv frame motion is violent, the laplacian distribution of chosen position parameter non-zero is as model, if current Wyner-Ziv frame motion is mild, selects the Gaussian Profile of average non-zero as model, and according to the average μ that step 1 obtains the skew that DCN occurs is proofreaied and correct, realize the accurate location of correlated noise MODEL C NM, specific as follows:

Wherein n represents the correlated noise value, and f (n) expression correlated noise n is at the probability density function of transform domain, μ _lBe respectively location parameter and the scale parameter of laplacian distribution, μ with b _gAnd σ ²Be respectively average and the variance of Gaussian Profile.

Images