CN107071447A - A kind of correlated noise modeling method based on two secondary side information in DVC - Google Patents
A kind of correlated noise modeling method based on two secondary side information in DVC Download PDFInfo
- Publication number
- CN107071447A CN107071447A CN201710221263.9A CN201710221263A CN107071447A CN 107071447 A CN107071447 A CN 107071447A CN 201710221263 A CN201710221263 A CN 201710221263A CN 107071447 A CN107071447 A CN 107071447A
- Authority
- CN
- China
- Prior art keywords
- frame
- side information
- noise
- frames
- secondary side
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/30—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using hierarchical techniques, e.g. scalability
- H04N19/395—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using hierarchical techniques, e.g. scalability involving distributed video coding [DVC], e.g. Wyner-Ziv video coding or Slepian-Wolf video coding
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
The invention discloses the correlated noise modeling method based on two secondary side information in a kind of DVC, the correlated noise source that the model is used in distributed video coding is：On the basis of the traditional once online noise and side information that decoding end pre-compensation frame is generated with after to compensation frame, using forward frame and while information frame, while information frame and backward frame between obtain the secondary online noises of two kinds of forms respectively, the fusion correlated noise that once online noise and secondary online noise fusion are obtained is used as correlated noise.Fusion correlated noise is modeled with the gauss hybrid models of parameter Estimation and the on-fixed window width Density Estimator model of non-parametric estmation, designs KL divergences judging module to determine the correlated noise model form of each frequency band.Model proposed by the present invention more can accurately be fitted the statistical property of the correlated noise in " pseudo channel " between WZ frames and side information, so as to effectively improve the distortion performance of Transform Domain Distribution formula Video coding.
Description
Technical field
It is specifically one kind the present invention relates to a kind of method of distributed video coding in video signal treatment technique field
Correlated noise modeling method based on two secondary side information in DVC.
Background technology
H.264/AVC, traditional video encoding standard such as PEG-4, waits and uses asymmetric coded system, coding side is hidden
Containing a decoder and comprising complicated motion estimation module, this asymmetric coding and decoding mode adapts to first encoding,
The application field repeatedly decoded, but in some application fields in recent years, as the wireless video detecting head in monitoring system,
Portable video camera, wireless pc camera etc., the equipment of this class often energy and resource-constrained, it is desirable to encoding device
Simply, relatively more resources and energy can be possessed to carry out complexity by being in the decoding device of the central server of terminal
Calculating is handled.
In the epoch of eighties of last century 70, Slepain, Wolf, Wyner, Ziv et al. theoretically demonstrate multiple related letters
Source can also reach the code efficiency of traditional combined coding combined decoding in the case of absolute coding and combined decoding., this
Beginning of the century just starts to have successively foreign scholar to set about carrying out video to realize algorithm research in absolute coding and combined decoding, and takes
Obtain certain achievement and gradually cause concern.This new distributed video coding scheme is by originally complicated excavation video
The motion compensation of sequence time and spatial redundancy and inter prediction have been put into decoding end, so as to greatly reduce the complexity of coding side
Degree.
DVC code efficiency is realized typically by the associated statistical information using source information and side information.Decoding end
The statistical information of excavation correlated noise is needed, the probability log-likelihood for calculating estimation is used for the initialization input of decoding algorithm.
Particularly, high coding efficiency is largely dependent upon capability of fitting of the model used to correlated noise.However, to correlated noise
Carry out the reason for Accurate Model faces many challenges and be that decoding end can not obtain primitive frame, and can not also be obtained in coding side
Side information.In addition, the time domain and the non-stationary property in spatial domain of vision signal, and block and can influence correlated noise with illumination variation
Statistical information.In order to improve coding efficiency, correlated noise statistical estimate between source information and side information should will be as far as possible
It is accurate, in order to realize this point, the present invention is from the aspect of the obtaining and optimize noise modeling two of online noise is optimized, more accurately
The characteristic of " pseudo channel " is described, more accurate input is provided for decoder, Transform Domain Distribution formula video volume can be effectively improved
The distortion performance of code.
The content of the invention
The online noise that the present invention is obtained for existing method is not accurate enough, and fitting of the existing model to online noise
The problem of ability has much room for improvement.In order to obtain more accurate correlated noise and raising model pair between original WZ frames and side information
A kind of capability of fitting of online noise, it is proposed that the correlated noise modeling method based on two secondary side information in DVC.The model is incited somebody to action
The fusion noise arrived uses gauss hybrid models and the on-fixed window width of non-parametric estmation as residual sample (correlated noise)
Density Estimator (KDE, kernel density estimation) is common to be modeled to fusion noise, and passes through KL divergences
(KLD, Kullback-Leibler divergence) judging module come adaptively determine a certain frequency band be use based on ginseng
The gauss hybrid models or nonparametric estimation model of number estimation.Correlated noise preparation method and the correlated noise modeling of the present invention
Scheme can more accurately describe the characteristic of " pseudo channel ", more accurate input be provided for decoder, while also effectively improving
The distortion performance of Transform Domain Distribution formula Video coding (Distributed Video Coding, DVC).
The present invention is achieved through the following technical solutions.
A kind of correlated noise modeling method based on two secondary side information in DVC, this method comprises the following steps：
(1) the secondary generation of online noise and side information；
(2) the merging of secondary online noise and once online noise, the fusion of side information；
(3) training KDE models and estimation mixed Gauss model parameter；
(4) KL divergence judging modules are utilized, adaptively the correlated noise of each frequency band is modeled.
Further, step (1) is specifically included：Video sequence is passed through into image grouping module GOP=2 (Group of
Pictures), it is divided into even frame and odd-numbered frame, even frame is that Wyner-Ziv frames i.e. WZ frames and odd-numbered frame is key frame i.e. K frames,
Then WZ frames include WZ in 2m frames video sequence2、WZ4、…、WZ2n、…、WZ2mFrame, K frames include K1、K3、…、K2n-1、…、K2m-1
Frame；K frames use traditional intraframe coding, and the K frames that decoding end is decoded includeFrame, n=1
~m；(1.1) decoding end,WithSearched between frame, obtain preceding backward motion vector Wherein dxfRepresent the component of forward motion vector in the horizontal direction, dxbRepresent backward motion vector
Component in the horizontal direction, dyfForward motion vector is represented in vertically-oriented component, dybBackward motion vector is represented vertical
The component in direction, upper target implication represents searching motion vector, following dx between corresponding two framef dxb, dyf dybIt is also
Same implication, dxf dxb, dyf dybIn f be represent before to Front the meaning, b is the meaning for representing backward Back, WZ2nFrame
Pre-compensation frame beX represents the abscissa of pixel, and y represents vertical
Coordinate, is constructedFrame be2n-1 frames are utilized on the basis of frame, the propulsion arrow that 2n+1 frame search goes out
AmountObtain, backward compensation frame isIt is following
Alphabetical implication in each frame that step (1.2) is constructed is also referring to this step；One secondary side information of generationForSubscript 1 represent be a secondary side information, subscript 2n represents 2n frames, i.e. WZ2nFrame
Corresponding side information, once online residual error is that correlated noise is Subscript 1 represent be one
Secondary online residual error, subscript 2n represents 2n frames, i.e. WZ2nThe corresponding online residual error of frame；All it is
The frame constructed；
(1.2) on the one hand exist) andBetween search for, obtain moving forward and backward vector Backward frame be's
Forward frameOn the other hand existWithBetween search for, before obtaining
Backward motion vector Backward frame be Forward frame beDx hereinf dxb, dyf dybImplication is as abovementioned steps；
(1.3) two secondary side information of generation are as follows：Wherein subscript
21 and subscript 22 represent respectively generation two secondary side information the first form and second of form；The quadratic residue of generation is phase
Close noise as follows:
Further, step (2) is specifically included：
Fusion treatment is carried out respectively to secondary correlated noise and two secondary side information, obtains merging online noiseFusion side information is taken as：Subscript 2n and step (1) herein is described
N in 2n-1,2n+1 is meant that identical, and what 2n was meant that is 2n frames, and what 2n-1 was meant that is 2n-1 frames, and 2n+1 is with regard to table
What is shown is 2n+1 frames, so SI2nIt is 2n frames i.e. WZ2nThe final corresponding fusion side information of frame, R2nIt is 2n frames i.e. WZ2nFrame
The final corresponding online noise of fusion.
Further, step (3) is specifically included：
(3.1) gauss hybrid models parameter initialization：
Gauss hybrid modelsY represents some residual sample value, and θ is gauss hybrid models
Parameter set, θ=(θ1,θ2,...θS), S is the sub-model number of gauss hybrid models, θi=(αi,ui,σi), i=1,2 ... S,
I represents it is which Gaussian Profile, αiIt is the weight of i-th of Gaussian Profile, the probability density function of i-th of Gaussian Profile isμiIt is the weight of i-th of Gaussian Profile,It is the variance of i-th of Gaussian Profile,
To the fusion noise R of band level2nK-means clusters are done, the cluster centre result gathered for S classes are regard as θ in gauss hybrid models
=(θ1,θ2,...θS) parameter Estimation initial value；
(3.2) basis,
Update gauss hybrid models parameter set；yjIt is residual sample collection R2nIn j-th of residual values；N is residual sample collection R2nSample
Amount of capacity, S is the sub-model number of gauss hybrid models, αiIt is the weight of i-th of Gaussian Profile, μiIt is i-th of Gaussian Profile
Weight,It is the variance of i-th of Gaussian Profile,Represent yjBelong to the renewal estimate of the degree of membership of i-th of Gaussian Profile,It is the renewal estimate of the weight of i-th of Gaussian Profile,It is the renewal estimate of the weight of i-th of Gaussian Profile,It is
The renewal estimate of the variance of i-th of Gaussian Profile；
(3.3) when the log-likelihood function changing value of front and rear gauss hybrid models twice is less than given threshold, Gauss
The parameter set of mixed model, which updates, to be terminated；
(3.4) KDE model parameters are initialized：
KDE modelsX is the residual values to be estimated, XiIt is that residual sample is concentrated
I-th sample value, N is residual sample collection R2nSample size, h (x) * are the required optimization of step (3.5) to be estimated
Residual values x adaptive bandwidth varying, K () uses kernel function,Calculate initial fixing band
Wide h0, It is sample variance；
(3.5) the adaptive bandwidth varying of calculation optimizationWherein,h0It is initial required by step (3.4)
Fixed-bandwidth, x is the residual values to be estimated, XiIt is i-th of sample value that residual sample is concentrated, N is residual sample collection R2nSample
This capacity,It is to use h0For under bandwidth situation to the Multilayer networks value for the residual values x to be estimated.
Further, step (4) is specifically included：
(4.1) the residual error R under two kinds of correlated noise models of Gaussian Mixture and KDE is obtained using step (2)2nProbability it is close
Degree, does the calculating of KL divergences with merging the probability density of noise residual sample respectively；Selection can obtain the mould of smaller KL divergence values
Type as present band optimal models；
(4.2) optimal models selected based on step (4.1), calculate after residual error binary representation each for 0 and be 1
Probability log-likelihood ratio, you can obtain LDPCA decoders and more accurately input, and send into LDPCA decoder modules.
Further, two secondary side information and secondary online noise refer to：In generation forward frameWith side information SI2n
Between, side information SI2nWith backward frameBetween continue search for more further motion vector；Obtain it is more accurate before
To frameFront and rear is to frameBy these newly-generated forward frames and backward frame, obtain
Two kinds of two new secondary side information and secondary online noiseAnd fusion treatment is carried out to it；Melting after processing
Close the superposition that noise is original once online noise sample collection and the secondary online noise sample collection of two kinds of forms；After processing
Fusion side information is being averaged for two secondary side information of an original secondary side information and two kinds of forms.
Further, for fusion noise R in step (3.5)2nKDE models in window width estimation no longer use fixed window
Width, but for the different window width of different Variable selections, better adapts to close quarters in sample set or sparse region
Independent variable.
Further, in step (3.1)-(3.5), incorporate for fusion noise R2nTwo kinds of different statistical methods
Multilayer networks；To fusion noise R2nThe mixed Gauss model of parameter Estimation and the KDE models of non-parametric estmation are carried out simultaneously
Estimation, obtain residual error R2nMultilayer networks under two kinds of models, then again respectively with fusion noise R2nProbability density
The calculating of KL divergences is done, selects that in the case where present band merges the residual sample of noise the model conduct of smaller KL divergence values can be obtained
The optimal models of present band.
The present invention compared with the prior art, has the following advantages that and beneficial effect：
1st, conventional noise preparation method is innovated on source from obtaining for online noise, the fusion noise of acquisition
The true correlation noise that can more press close in " pseudo channel ".
2nd, fusion noise is to be merged once online noise and secondary online noise, the sample size of noise sample collection
Doubled, the increase of sample size, be favorably improved the estimation of gauss hybrid models parameter set, the accuracy of Density Estimator.
When the 3rd, modeling, the gauss hybrid models for belonging to parameter Estimation in statistical analysis and non-parametric estmation mould have been considered
The advantage of both different classes of models of type, and the fitting degree of both model samples is judged by objective KL divergences, it is adaptive
Ground selection is answered to model correlated noise.
Brief description of the drawings
Fig. 1 is the general frame figure of existing transform domain coding and decoding video.
Fig. 2 is the frame diagram of correlated noise modeling of the present invention based on two secondary side information.
Fig. 3 is the flow chart of the invention based on two secondary side information and secondary online noise.
Specific implementation method
With reference to embodiment and accompanying drawing, the present invention is described in further detail, but embodiments of the present invention are not limited
In this.
Fig. 1 is the frame diagram of the distributed video encoding and decoding of existing transform domain.The present invention proposes one kind as shown in Figure 2
Correlated noise modeling method based on two secondary side information in DVC, modeling module incorporates the Gaussian Mixture mould for belonging to parameter Estimation
Type and the KDE models of non-parametric estmation, using KL divergence judging modules, are adaptively selected most for the residual sample of each frequency band
Good model, so as to more accurately excavate the statistical property of " pseudo channel " that belongs to the frequency band.For the input of LDPCA decoders more
Input is used for for reliable probability log-likelihood.So as to improve the distortion performance of DVC systems.
The specific embodiment of the present invention is given below.
(1) such as Fig. 3, the secondary generation of online noise and side information comprises the following steps：
(1.1) by video sequence " foreman ", " news ", " soccer " is by image grouping module (GOP, Group of
Pictures), GOP=2 point is even frame and odd-numbered frame, and even frame is Wyner-Ziv frames (WZ frames) and odd-numbered frame is key frame
WZ frames include in (K frames), such as 2m frame video sequences：WZ2、WZ4、…、WZ2n、…、WZ2mFrame, K frames include K1、K3、…、K2n-1、…、
K2m-1Frame.K frames use traditional intraframe coding, and the K frames that decoding end is decoded includeFrame, n=
1~m.
(1.2) decoding end,WithSearched between frame, obtain preceding backward motion vector Wherein dxfRepresent the component of forward motion vector in the horizontal direction, dxbRepresent backward motion vector
Component in the horizontal direction, dyfForward motion vector is represented in vertically-oriented component, dybBackward motion vector is represented vertical
The component in direction, upper target implication represents searching motion vector, following dx between corresponding two framef dxb, dyf dybIt is also
Same implication, dxf dxb, dyf dybIn f be represent before to Front the meaning, b is the meaning for representing backward Back, WZ2nFrame
Pre-compensation frame beX represents the abscissa of pixel, and y represents vertical
Coordinate, is constructedFrame be2n-1 frames are utilized on the basis of frame, the propulsion arrow that 2n+1 frame search goes out
AmountObtain, backward compensation frame isIt is following
Alphabetical implication in each frame that step (1.2) is constructed is also referring to this step；One secondary side information of generationForSubscript 1 represent be a secondary side information, subscript 2n represents 2n frames, i.e. WZ2nFrame
Corresponding side information, once online residual error is that correlated noise is Subscript 1 represent be one
Secondary online residual error, subscript 2n represents 2n frames, i.e. WZ2nThe corresponding online residual error of frame；All it is
The frame constructed；
(1.3) on the one hand existWithBetween search for, obtain moving forward and backward vector Backward frame be's
Forward frameOn the other hand existWithBetween search for, obtain before and after
To motion vector Backward frame beForward frame beDx hereinf dxb, dyf dybImplication is as abovementioned steps；
(1.4) two secondary side information of generation are as follows：
Wherein subscript 21 and subscript 22 represent the first form and second of form of two secondary side information of generation respectively.What is generated is secondary
Residual error (correlated noise) is as follows:
(1.5) fusion treatment is carried out to secondary correlated noise and two secondary side information, merging online noise isMerging side information is：
(2) utilize newly-generated fusion noise, training KDE models and estimation mixed Gauss model parameter method include with
Lower step：
(2.1) gauss hybrid models parameter initialization：
Gauss hybrid modelsY represents some residual sample value, and θ is gauss hybrid models
Parameter set, θ=(θ1,θ2,...θS), S is the sub-model number of gauss hybrid models, θi=(αi,ui,σi), i=1,2 ... S,
I represents it is which Gaussian Profile, αiIt is the weight of i-th of Gaussian Profile, the probability density function of i-th of Gaussian Profile isμiIt is the weight of i-th of Gaussian Profile,It is the variance of i-th of Gaussian Profile,
To the fusion noise R of band level2nK-means clusters are done, the cluster centre result gathered for S classes are regard as θ in gauss hybrid models
=(θ1,θ2,...θS) parameter Estimation initial value；
(2.2) basis,
Update gauss hybrid models parameter set；yjIt is residual sample collection R2nIn j-th of residual values；N is residual sample collection R2nSample
Amount of capacity, S is the sub-model number of gauss hybrid models, αiIt is the weight of i-th of Gaussian Profile, μiIt is i-th of Gaussian Profile
Weight,It is the variance of i-th of Gaussian Profile,Represent yjBelong to the renewal estimate of the degree of membership of i-th of Gaussian Profile,It is the renewal estimate of the weight of i-th of Gaussian Profile,It is the renewal estimate of the weight of i-th of Gaussian Profile,It is
The renewal estimate of the variance of i-th of Gaussian Profile；
(2.3) when the log-likelihood function changing value of front and rear gauss hybrid models twice is less than given threshold 0.001,
The parameter set of gauss hybrid models, which updates, to be terminated.
(2.4) KDE model parameters are initialized：
KDE modelsX is the residual values to be estimated, XiIt is that residual sample is concentrated
I-th sample value, N is residual sample collection R2nSample size, h (x)*It is step (2.5) required optimization to be estimated
Residual values x adaptive bandwidth varying, K () uses kernel function,Calculate initial fixing band
Wide h0, It is sample variance；
(2.5) the adaptive bandwidth varying of calculation optimizationWherein,h0It is initial required by step (2.4)
Fixed-bandwidth, x is the residual values to be estimated, XiIt is i-th of sample value that residual sample is concentrated, N is residual sample collection R2nSample
This capacity,It is to use h0For under bandwidth situation to the Multilayer networks value for the residual values x to be estimated.
(3) KL divergence judging modules are utilized, the adaptively correlated noise modeling to each frequency band comprises the following steps：
(3.1) the residual error R under two kinds of correlated noise models of Gaussian Mixture and KDE is obtained using step (2)2nProbability it is close
Degree, does the calculating of KL divergences with merging the probability density of noise residual sample respectively.Selection can obtain the mould of smaller KL divergence values
Type as present band optimal models.
(3.2) optimal models selected based on step (3.1), calculate after residual error binary representation each for 0 and be 1
Probability log-likelihood ratio, you can obtain " the soft input " of LDPCA decoders.And send into LDPCA decoder modules.
Above-described embodiment is preferably embodiment, but embodiments of the present invention are not by above-described embodiment of the invention
Limitation, other any Spirit Essences without departing from the present invention and the change made under principle, modification, replacement, combine, simplification,
Equivalent substitute mode is should be, is included within protection scope of the present invention.
Claims (8)
1. the correlated noise modeling method based on two secondary side information in a kind of DVC, it is characterised in that this method comprises the following steps：
(1) the secondary generation of online noise and side information；
(2) the merging of secondary online noise and once online noise, the fusion of side information；
(3) training KDE models and estimation mixed Gauss model parameter；
(4) KL divergence judging modules are utilized, adaptively the correlated noise of each frequency band is modeled.
2. the correlated noise modeling method based on two secondary side information in a kind of DVC according to claim 1, its feature exists
In：Step (1) is specifically included：
(1.1) video sequence is passed through into image grouping module GOP=2 (Group of Pictures), is divided into even frame and odd number
Frame, even frame is that Wyner-Ziv frames i.e. WZ frames and odd-numbered frame is key frame i.e. K frames, then WZ frames include in 2m frames video sequence
WZ2、WZ4、...、WZ2n、...、WZ2mFrame, K frames include K1、K3、...、K2n-1、...、K2m-1Frame；K frames are compiled using traditional frame in
Code, the K frames that decoding end is decoded includeFrame, n=1~m；
(1.2) decoding end,WithSearched between frame, obtain preceding backward motion vector Wherein dxfRepresent the component of forward motion vector in the horizontal direction, dxbRepresent backward motion vector
Component in the horizontal direction, dyfForward motion vector is represented in vertically-oriented component, dybBackward motion vector is represented vertical
The component in direction, upper target implication represents searching motion vector, following dx between corresponding two framef dxb, dyf dybIt is also
Same implication, dxf dxb, dyf dybIn f be represent before to Front the meaning, b is the meaning for representing backward Back, WZ2nFrame
Pre-compensation frame beX represents the abscissa of pixel, and y represents vertical
Coordinate, is constructedFrame be2n-1 frames are utilized on the basis of frame, the propulsion arrow that 2n+1 frame search goes out
AmountObtain, backward compensation frame isIt is following
Alphabetical implication in each frame that step (1.3) is constructed is also referring to this step；One secondary side information of generationFor Subscript 1 represent be a secondary side information, subscript 2n represents 2n frames, i.e. WZ2nFrame pair
The side information answered, once online residual error is that correlated noise is Subscript 1 represent be once
Online residual error, subscript 2n represents 2n frames, i.e. WZ2nThe corresponding online residual error of frame；All it is structure
The frame created；
(1.3) on the one hand existWithBetween search for, obtain new preceding backward motion vector And calculateBackward frame be Forward frameOn the other hand existWithBetween search for, obtain
To preceding backward motion vector Backward frame be Forward frame beDx hereinf dxb, dyf dybImplication is as abovementioned steps；
(1.4) two secondary side information of generation are as follows：Wherein
Subscript 21 and subscript 22 represent the first form and second of form of two secondary side information of generation respectively；The quadratic residue of generation
I.e. correlated noise is as follows:
3. the correlated noise modeling method based on two secondary side information in a kind of DVC according to claim 2, its feature exists
In：Step (2) is specifically included：
Fusion treatment is carried out respectively to secondary correlated noise and two secondary side information, obtains merging online noise
Fusion side information is taken as：N in subscript 2n herein and step (1) described 2n-1,2n+1
Identical is meant that, what 2n was meant that is 2n frames, what 2n-1 was meant that is 2n-1 frames, that 2n+1 is meant that is 2n+1
Frame, so SI2nIt is 2n frames i.e. WZ2nThe final corresponding fusion side information of frame, R2nIt is 2n frames i.e. WZ2nFrame is finally corresponding to be melted
Close online noise.
4. the correlated noise modeling method based on two secondary side information in a kind of DVC according to claim 1, its feature exists
In：Step (3) is specifically included：
(3.1) gauss hybrid models parameter initialization：
Gauss hybrid modelsY represents some residual sample value, and θ is the parameter of gauss hybrid models
Collection, θ=(θ1,θ2,...θS), S is the sub-model number of gauss hybrid models, θi=(αi,ui,σi), i=1,2 ... S, i table
Show it is which Gaussian Profile, αiIt is the weight of i-th of Gaussian Profile, the probability density function of i-th of Gaussian Profile isμiIt is the weight of i-th of Gaussian Profile,It is the variance of i-th of Gaussian Profile,
To the fusion noise R of band level2nK-means clusters are done, the cluster centre result gathered for S classes are regard as θ in gauss hybrid models
=(θ1,θ2,...θS) parameter Estimation initial value；
(3.2) basis,Update
Gauss hybrid models parameter set；yjIt is residual sample collection R2nIn j-th of residual values；N is residual sample collection R2nSample size
Size, S is the sub-model number of gauss hybrid models, αiIt is the weight of i-th of Gaussian Profile, μiIt is the power of i-th of Gaussian Profile
Weight,It is the variance of i-th of Gaussian Profile,Represent yjBelong to the renewal estimate of the degree of membership of i-th of Gaussian Profile,It is
The renewal estimate of the weight of i-th of Gaussian Profile,It is the renewal estimate of the weight of i-th of Gaussian Profile,It is i-th
The renewal estimate of the variance of Gaussian Profile；
(3.3) when the log-likelihood function changing value of front and rear gauss hybrid models twice is less than given threshold, Gaussian Mixture
The parameter set of model, which updates, to be terminated；
(3.4) KDE model parameters are initialized：
KDE modelsX is the residual values to be estimated, and Xi is that residual sample is concentrated
I sample value, N is residual sample collection R2nSample size, h (x) * are the required optimization of step (3.5) to the residual error to be estimated
Value x adaptive bandwidth varying, K () uses kernel function,Calculate initial fixed-bandwidth h0, It is sample variance；
(3.5) the adaptive bandwidth varying of calculation optimizationWherein,h0It is initial required by step (3.4)
Fixed-bandwidth, x is the residual values to be estimated, XiIt is i-th of sample value that residual sample is concentrated, N is residual sample collection R2nSample
This capacity,It is to use h0For under bandwidth situation to the Multilayer networks value for the residual values x to be estimated.
5. the correlated noise modeling method based on two secondary side information in a kind of DVC according to claim 1, its feature exists
In：Step (4) is specifically included：
(4.1) the residual error R under two kinds of correlated noise models of Gaussian Mixture and KDE is obtained using step (2)2nProbability density, point
The calculating of KL divergences is not done with merging the probability density of noise residual sample；Selection can obtain the model conduct of smaller KL divergence values
The optimal models of present band；
(4.2) based on step (4.1) select optimal models, calculate residual error binary representation after each for 0 and be 1 it is general
Rate log-likelihood ratio, you can obtain LDPCA decoders and more accurately input, and send into LDPCA decoder modules.
6. the correlated noise modeling method based on two secondary side information in a kind of DVC according to claim 1, its feature exists
In：Two secondary side information and secondary online noise refer to：In generation forward frameWith side information SI2nBetween, side information SI2n
With backward frameBetween continue search for more further motion vector；Obtain more accurate forward frameFront and rear is to frameBy these newly-generated forward frames and backward frame, two kinds are obtained
Two new secondary side information and secondary online noiseAnd fusion treatment is carried out to it；Fusion after processing is made an uproar
Sound is the superposition of original once online noise sample collection and the secondary online noise sample collection of two kinds of forms；Fusion after processing
Side information is being averaged for two secondary side information of an original secondary side information and two kinds of forms.
7. the correlated noise modeling method based on two secondary side information in a kind of DVC according to claim 1, its feature exists
In：For fusion noise R in step (3.5)2nKDE models in window width estimation no longer using fixed window width, but be directed to
The different window width of different Variable selections, better adapts to close quarters or the independent variable of sparse region in sample set.
8. the correlated noise modeling method based on two secondary side information in a kind of DVC according to claim 1, its feature exists
In：In step (3.1)-(3.5), incorporate for fusion noise R2nThe probability density of two kinds of different statistical methods estimate
Meter；To fusion noise R2nThe estimation of the mixed Gauss model of parameter Estimation and the KDE models of non-parametric estmation is carried out simultaneously, is obtained
Residual error R2nMultilayer networks under two kinds of models, then again respectively with fusion noise R2nProbability density make KL divergence meters
Calculate, select that the model of smaller KL divergence values can be obtained as present band in the case where present band merges the residual sample of noise
Optimal models.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710221263.9A CN107071447B (en) | 2017-04-06 | 2017-04-06 | Correlated noise modeling method based on secondary side information in DVC |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710221263.9A CN107071447B (en) | 2017-04-06 | 2017-04-06 | Correlated noise modeling method based on secondary side information in DVC |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107071447A true CN107071447A (en) | 2017-08-18 |
| CN107071447B CN107071447B (en) | 2020-06-19 |
Family
ID=59601642
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710221263.9A Expired - Fee Related CN107071447B (en) | 2017-04-06 | 2017-04-06 | Correlated noise modeling method based on secondary side information in DVC |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107071447B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108833920A (en) * | 2018-06-04 | 2018-11-16 | 四川大学 | A kind of DVC side information fusion method based on light stream and Block- matching |
| CN112560832A (en) * | 2021-03-01 | 2021-03-26 | 腾讯科技（深圳）有限公司 | Video fingerprint generation method, video matching method, video fingerprint generation device and video matching device and computer equipment |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103096085A (en) * | 2013-02-04 | 2013-05-08 | 山东科技大学 | Method of real-timely generating high-quality digital video camera (DVC) decoding side edge information and based on orthogonal discrete wavelet transform (ODWT) sub-band |
| CN103561269A (en) * | 2013-08-14 | 2014-02-05 | 广西大学 | Distributed video coding correlated noise model construction method based on multi-probability distribution |
| CN103618907A (en) * | 2013-11-08 | 2014-03-05 | 天津大学 | Multi-viewpoint distributed type video encoding and frame arranging device and method based on compressed sensing |
| CN104702961A (en) * | 2015-02-17 | 2015-06-10 | 南京邮电大学 | Code rate control method for distributed video coding |
| JP2015154198A (en) * | 2014-02-13 | 2015-08-24 | 沖電気工業株式会社 | Moving image decoding device and program |
-
2017
- 2017-04-06 CN CN201710221263.9A patent/CN107071447B/en not_active Expired - Fee Related
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103096085A (en) * | 2013-02-04 | 2013-05-08 | 山东科技大学 | Method of real-timely generating high-quality digital video camera (DVC) decoding side edge information and based on orthogonal discrete wavelet transform (ODWT) sub-band |
| CN103561269A (en) * | 2013-08-14 | 2014-02-05 | 广西大学 | Distributed video coding correlated noise model construction method based on multi-probability distribution |
| CN103618907A (en) * | 2013-11-08 | 2014-03-05 | 天津大学 | Multi-viewpoint distributed type video encoding and frame arranging device and method based on compressed sensing |
| JP2015154198A (en) * | 2014-02-13 | 2015-08-24 | 沖電気工業株式会社 | Moving image decoding device and program |
| CN104702961A (en) * | 2015-02-17 | 2015-06-10 | 南京邮电大学 | Code rate control method for distributed video coding |
Non-Patent Citations (1)
| Title |
|---|
| 刘杰平等: "DVC中基于广义高斯分布的WZ帧重构算法", 《华南理工大学学报（自然科学版）》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108833920A (en) * | 2018-06-04 | 2018-11-16 | 四川大学 | A kind of DVC side information fusion method based on light stream and Block- matching |
| CN108833920B (en) * | 2018-06-04 | 2022-02-11 | 四川大学 | DVC side information fusion method based on optical flow and block matching |
| CN112560832A (en) * | 2021-03-01 | 2021-03-26 | 腾讯科技（深圳）有限公司 | Video fingerprint generation method, video matching method, video fingerprint generation device and video matching device and computer equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107071447B (en) | 2020-06-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107105278B (en) | The video coding and decoding system that motion vector automatically generates | |
| CN102137263B (en) | Distributed video coding and decoding methods based on classification of key frames of correlation noise model (CNM) | |
| CN107463989B (en) | A kind of image based on deep learning goes compression artefacts method | |
| CN110533044B (en) | Domain adaptive image semantic segmentation method based on GAN | |
| Qi et al. | Reduced reference stereoscopic image quality assessment based on binocular perceptual information | |
| CN110135386B (en) | Human body action recognition method and system based on deep learning | |
| CN110889895B (en) | Face video super-resolution reconstruction method fusing single-frame reconstruction network | |
| CN110852964A (en) | Image bit enhancement method based on deep learning | |
| CN110087087A (en) | VVC interframe encode unit prediction mode shifts to an earlier date decision and block divides and shifts to an earlier date terminating method | |
| CN109948721B (en) | Video scene classification method based on video description | |
| CN105338357B (en) | A kind of distributed video compressed sensing decoding method | |
| CN108924558B (en) | Video predictive coding method based on neural network | |
| CN103501441A (en) | Multiple-description video coding method based on human visual system | |
| CN105933711B (en) | Neighborhood optimum probability video steganalysis method and system based on segmentation | |
| CN110312138B (en) | High-embedding-capacity video steganography method and system based on time sequence residual convolution modeling | |
| CN109949217A (en) | Video super-resolution method for reconstructing based on residual error study and implicit motion compensation | |
| CN108257098A (en) | Video denoising method based on maximum posteriori decoding and three-dimensional bits matched filtering | |
| Zha et al. | The power of triply complementary priors for image compressive sensing | |
| CN102970510A (en) | Method for transmitting human face video | |
| CN107071447A (en) | A kind of correlated noise modeling method based on two secondary side information in DVC | |
| Wu et al. | Memorize, then recall: a generative framework for low bit-rate surveillance video compression | |
| Hoogeboom et al. | High-fidelity image compression with score-based generative models | |
| Persson et al. | Mixture model-and least squares-based packet video error concealment | |
| Xiong et al. | High quality image reconstruction via non-local collaborative estimation for wireless image/video softcast | |
| CN108833920A (en) | A kind of DVC side information fusion method based on light stream and Block- matching |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200619 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |