CN102685548B - The nothing ginseng appraisal procedure of video quality - Google Patents

Abstract

Present invention is disclosed a kind of nothing ginseng appraisal procedure of video quality, comprise the following steps: step 1: from compressed video data stream, extract quantization parameter and the DCT coefficient value after quantizing, by the distributed area estimating original DCT coefficient in quantization parameter; Step 2: by the sample comparison in the DCT coefficient extracted in data flow and quantization parameter and training set, and then the probability distribution parameters of DCT coefficient when estimating uncompressed; Step 3: integration is carried out to DCT coefficient PDF distribution, produces the objective estimation to compression noise with this.The present invention is not when having original video sequence reference, the method is based on the study to video priori, the calculating of distributed constant, fuzzy C average grader, the matching algorithm in the nearest field of macro block and the integral algorithm of snr computation, achieve printenv assessment MPEG2 video compression sequence being carried out to video quality.

Description

The nothing ginseng appraisal procedure of video quality

Technical field

The present invention relates to a kind of nothing ginseng appraisal procedure of video quality, more particularly, relate to a kind of method of MPEG2 video compression sequence being carried out to the printenv assessment of video quality.

Background technology

Video compression mostly is lossy compression method, and the detailed information of losing in compression process cannot be recovered completely in decoding end, so the compression noise of video estimates it is a typical ill-conditioning problem.At present, insider adopts Y-PSNR (PSNR) to carry out the video quality of metric contraction video usually.But Y-PSNR is one ginseng estimated amount of damage: with former video as a reference, the compressed video through decoding is compared with it, describe crushing loss with its difference signal.Have ginseng estimated amount of damage algorithm simply and comparatively accurate, but its service condition is harsh, is only applicable to the active video occasions as a reference such as encoder performance assessment.For the video monitoring system of reality, be to obtain the original scene information without compression at all; So, also just cannot use ginseng loss appraisal algorithm.

Therefore, provide a kind of innovation, efficient nothing ginseng evaluation system, namely when not having former video reference, it is the technical problem that this area needs solution badly that the parameter of foundation compressed video and priori measure the compressed error of video in video monitoring system.

Summary of the invention

Object of the present invention aims to provide a kind of nothing ginseng appraisal procedure of video quality, solves the various deficiencies existed in prior art.

According to the present invention, a kind of nothing ginseng appraisal procedure of video quality is provided, comprise the following steps: step 1: from compressed video data stream, extract quantization parameter and discrete cosine transform (DCT) coefficient value after quantizing, by the distributed area estimating original DCT coefficient in quantization parameter; Step 2: by the sample comparison in the DCT coefficient extracted in data flow and quantization parameter and training set, and then the probability distribution parameters of DCT coefficient when estimating uncompressed; Step 3: integration is carried out to DCT coefficient PDF distribution, produces the objective estimation to compression noise with this.

According to one embodiment of the invention, DCT coefficient meets laplacian distribution $f\left(x\right)=\frac{\mathrm{\λ}(u,v)}{2}\exp (-\mathrm{\λ}(u,v)|x\left|\right).$

According to one embodiment of the invention, with the distributed constant block of 64 λ (u, v) coefficient compositions for target function, will have multiple data block clusters of identical DCT coefficient distribution characteristics together, clustering algorithm is: 1≤m < ∞; After setting cluster numbers m, clustering algorithm, by recursive calculation, is asked for and is made J _mminimum optimal classification combination.

According to one embodiment of the invention, first take out multiple uncompressed test pattern, the true λ of DCT coefficient is calculated by uncompressed image, uncompressed image quantization is compressed, produce different compression samples, estimated the probability distribution of DCT coefficient again by compression samples, and the λ estimated and true λ is contrasted, calculate the mean and variance of the difference of λ and the true λ estimated.

According to one embodiment of the invention, by the image block after decompression and Sample Storehouse contrast, to extract the lambda parameter of described DCT coefficient distribution, the image block that compressed picture blocks is consistent with compression parameters in the sample of priori storehouse contrasts, first the coding mode of read block and described quantization parameter, read in corresponding described priori library information subsequently.

According to one embodiment of the invention, the sample of the described DCT coefficient after de-quantization directly and in priori storehouse is contrasted.

According to one embodiment of the invention, first by predictive frame complete decoding, then call dct transform function, convert thereof into as described DCT coefficient, quantization error in described DCT coefficient is produced by Inter quantization table, when module is mated, adopts the sample in corresponding inter priori storehouse to mate.

According to one embodiment of the invention, estimate stochastic variable x _ibe quantified as X _iprobability, and the probability-distribution function obtaining original described DCT coefficient x is

Have employed technical scheme of the present invention, for the blank of current such technology immature, domestic of prior art in the world, and provide a kind of nothing ginseng appraisal procedure of video quality.When there is no original video sequence reference, the method is based on the study to video priori, the calculating of distributed constant, fuzzy C average grader, the matching algorithm in the nearest field of macro block and the integral algorithm of snr computation, achieve printenv assessment MPEG2 video compression sequence being carried out to video quality.

Accompanying drawing explanation

In the present invention, identical Reference numeral represents identical feature all the time, wherein:

Fig. 1 is the flow chart of video decode of the present invention;

Fig. 2 is the laplacian distribution figure under the λ value that the present invention is different;

Fig. 3 is the comparison diagram of λ evaluated error of the present invention and number of categories;

Fig. 4 is the comparison diagram of PSNR evaluated error of the present invention and number of categories;

Fig. 5 is the flow chart of priori schema extraction of the present invention;

Fig. 6 is the flow chart of signal-to-noise ratio (SNR) estimation of the present invention.

Embodiment

Technical scheme of the present invention is further illustrated below in conjunction with drawings and Examples.In order to realize estimating without participating in evaluation and electing of video quality, the present invention adopts following technical scheme:

Signal to noise ratio is in the process of video decode, an important symbol of evaluates video quality.As shown in Figure 1, video decode of the present invention comprises the following steps:

S101: start video decode;

S102: player initialization;

S103: reading flow media file analysis type;

S104: files in stream media is split into audio stream and video flowing;

S105: audio stream;

S106: decoding video stream;

S107: sound, audio video synchronization export;

S108: event message circulates;

S109: file terminates? if so, S110 is gone to step; If not, S104 is gone to step;

S110: removing work is also exited.

When signal to noise ratio is assessed, need the prior information of image, this information contributes to the order of accuarcy improving compressed video signal-to-noise ratio (SNR) estimation.Original image, i.e. uncompressed image, also exist strong correlation between its content, so can be able to be estimated current by the probability distribution of investigation coefficient by the distribution of neighborhood DCT coefficient.But video compression can cause image impairment, destroy the coherence of adjacent block in image, so be difficult to the regularity of distribution accurately calculating original DCT coefficient from compressed image.The key addressed this problem sets up associating between compression samples with original image samples.If such contact can be set up, comparatively accurate DCT probability distribution just can be estimated.

Therefore, the present invention realizes estimating without participating in evaluation and electing of video quality by 3 key steps:

Step 1: extract quantization parameter and the DCT coefficient value after quantizing from compressed video data stream, by the distributed area estimating original DCT coefficient in quantization parameter;

Step 2: by the sample comparison in the DCT coefficient extracted in data flow and quantization parameter and training set, and then the probability distribution parameters of DCT coefficient when estimating uncompressed;

Step 3: integration is carried out to DCT coefficient probability-distribution function (Probability Density Function, PDF) distribution, produces the objective estimation to compression noise with this.

Describe technical scheme corresponding to each step below in detail.

step 1:

distributed constant calculates

Pertinent literature result of study shows, original DCT coefficient meets laplacian distribution,

$f\left(x\right)=\frac{\mathrm{\&lambda;}(u,v)}{2}\exp (-\mathrm{\&lambda;}(u,v)|x\left|\right)$

Wherein (u, v) is DCT coefficient coordinated indexing, and x is the value of DCT coefficient, and λ (u, v) is Laplce's controling parameters.For independent identically distributed sample x=(x ₁..., x _n), the maximum a posteriori probability of their statistical parameter λ (u, v) is

$L\left(\mathrm{\&lambda;}\right)=\underset{i=1}{\overset{n}{\mathrm{\&Pi;}}}\mathrm{\&lambda;}\&CenterDot;\exp (-\mathrm{\&lambda;}\&CenterDot;x_i)={\mathrm{\&lambda;}}^n\exp (-\mathrm{\&lambda;}\&CenterDot;\stackrel{\&OverBar;}{x})$

Wherein $\stackrel{\&OverBar;}{x}={\mathrm{\&Sigma;}}_{i=1}^n{x}_i$ For sample average.

To above formula differential, its maximum a posteriori probability solution is

${\mathrm{\&lambda;}}_{\mathrm{ML}}=\frac{n}{{\mathrm{\&Sigma;}}_{i=1}^n{x}_i}$

In above formula, laplacian distribution of the present invention as shown in Figure 2, DCT coefficient is distributed near 0 value very precipitous, different λ (u, v) value differs greatly in the integrated value in this interval, so we must to DCT coefficient, especially λ (u, v) the accurately valuation of fractional value DCT coefficient, the estimated accuracy of guarantee PSNR.

step 2:

fuzzy C-mean algorithm grader

The compression process of image/video decreases the comentropy of image, is characterized by the less characteristic quantity of the image varied.This mapping relations are many-to-one, namely have countless multiple DCT coefficient distribution will be quantized into as a certain specific distribution form.Therefore can not from compression samples, unique finds corresponding DCT coefficient distribution matrix.Suppose that N number of DCT coefficient block is quantized and be mapped to a certain specific DCT block, although these original DCT coefficient are not quite similar, they have certain general character, and namely the distribution of its coefficient is positioned to quantize centered by rear DCT coefficient, in the quantizing range of ± 1/2.

DCT coefficient matrix is the matrix of 8*8, namely has 64 DCT coefficient in matrix.Each DCT coefficient has oneself distribution function, i.e. λ (u, v).In the design, multiple data blocks with identical DCT coefficient distribution characteristics, for target function, flock together by the distributed constant block that we form with 64 λ (u, v) coefficients.C means clustering algorithm is as follows:

$J_m={\mathrm{\&Sigma;}}_{i=1}^N{\mathrm{\&Sigma;}}_{j=1}^C{u}_{i,j}^m{\left|\right|x_i-c_j\left|\right|}^2,$ 1≤m＜∞

Wherein, m for being greater than 1 integer, i.e. cluster numbers, u _{i, j}for x _ibe under the jurisdiction of the degree of membership of a jth set.X _ifor the measurement data of multidimensional, c _jfor the center of individual bunch of jth, || || be distance measure function, measure x _iwith c _jgeometric distance.After setting cluster numbers m, clustering algorithm, by recursive calculation, is asked for and is made J _mminimum optimal classification combination.

Experiment shows, clusters number is not The more the better.After clusters number exceedes a certain thresholding, the increase of clusters number reduces the contribution of Y-PSNR (PSNR) estimated accuracy.Due to when estimating signal to noise ratio, need decoded data block and sorted center of a sample to compare, sample space determines comparison speed, significantly can lower the execution speed of video signal-to-noise ratio estimating algorithm.In addition, Sample Storehouse is excessive also can propose very high requirement to system storage.

The present invention absorbs several sections of videos, from wherein intercepting the image of 4 frame 720*576 as original image samples.

These videos adopt MPEG2 coding, carry out shallow compression, so above-mentioned sequence maintains the correlative character of natural scene image substantially with the speed of 25Mbps.Because MPEG2 and JPEG compress mode is similar, we change JPEG quantization table, adopt Intra and Inter quantization parameter table and different compression ratios in MPEG2 to compress this four two field picture respectively.The selection of quantization parameter defines identical with MPEG quantization parameter, namely has 64 quantization parameter collection.2 kinds of quantization parameter tables and 64 quantization parameter collection form 128 kinds altogether and compress and combine, and produce 128 kinds of compression samples, then carry out cluster respectively to these 128 samples respectively.

The present invention sets optimal classification number by experiment.First take out some uncompressed test patterns (haveing nothing to do with training set), calculated the true λ of DCT coefficient by these images.These image quantizations are compressed, produces different compression samples, then estimate the probability distribution of wherein DCT coefficient by these compression samples.The λ estimated and true λ is contrasted, calculates the mean and variance of their differences.

As shown in Figure 3, Figure 4, after number of categories is greater than 100, nicety of grading and PSNR evaluated error reduce no longer merely, so the present invention gets 200 classification during estimated snr.

macro block nearest field matching algorithm

Image block after decompression needs and Sample Storehouse contrast, to extract the lambda parameter of DCT coefficient distribution.

In MPEG2, Video coding has four kinds of patterns: Intra, Inter, Direct, Skip.Quantization error under Intra and Inter pattern causes the main cause of image quality loss.Also there is encoding error in the macro block under Direct, Skip pattern-coding, its error is general less, far below the error of encoding block under Intra and Inter pattern, and under these two kinds of patterns the quantity of coded macroblocks also much smaller than, so analyze time, do not consider Direct, Skip mode macro.

Because image impairment is relevant with the quantization matrix of compression factor and employing, the residual signals of same image under different compression factor and amount template quantize is variant.So when doing Block-matching, must the image block that compressed picture blocks is consistent with compression parameters in the sample of priori storehouse be carried out Bizet the most effective.

Priori library information adopts Classification Management, is divided into intra and inter two class from large class.It is different that quantization parameter m quant is pressed in this two classes priori storehouse, is divided into 64 character libraries further, i.e. 128 priori storehouses altogether.When doing Block-matching, we are the coding mode of read block and quantization parameter first, reads in corresponding priori library information subsequently.

As shown in Figure 5, the priori schema extraction in priori storehouse of the present invention comprises following step:

S501: gather image pattern;

S502: image compression;

The true probability statistical computation of S503:DCT coefficient, and result of calculation is supplied to step S506;

S504: by λ cluster;

S505: the geometric center calculating DCT block sample set;

S506: vector combination, provides set of eigenvectors.

Sample entries in priori storehouse is made up of two parts, and leading portion is the DCT coefficient of cluster centre, and the latter is the DCT distributed constant of its correspondence.Intra coding lattice is similar to JPEG, so the sample of the DCT coefficient after de-quantization directly and in priori storehouse can be contrasted.Image block referenced by Inter block may cross over several lattice, so cannot directly calculate.The present invention, first by predictive frame complete decoding, is calling dct transform function, is converting thereof into as DCT coefficient.Because the quantization error in these DCT coefficient produced by Inter quantization table, so during module coupling, the sample in corresponding inter priori storehouse is adopted to mate.

signal-to-noise ratio (snr) estimation

As shown in Figure 6, signal-to-noise ratio (SNR) estimation of the present invention comprises the following steps:

S601: start video decode;

S602: information extraction, and quantization parameter is supplied to step S607;

S603: by formed data frame, by DCT computing with words;

S604:DCT data block is extracted, and DCT system is supplied to step S607;

S605: priori schema extraction;

S606: block mode mates, and λ estimated value is supplied to step S607;

S607: the integral operation of signal-to-noise ratio (SNR) estimation, thus the estimated value of the PSRN asked.

According to above-mentioned step, when after the distributed constant drawing out DCT coefficient, we can estimate stochastic variable x _ibe quantified as X _iprobability

$P(X_i=Q\left(X_i\right))={\&Integral;}_a^b\frac{\mathrm{\&lambda;}}{2}{e}^{-\mathrm{\&lambda;}\left|x\right|}\mathrm{dx}$

Wherein (a, b) is quantized interval, and DCT coefficient is wherein quantized, and uses X _irepresent.Given go quantize after DCT coefficient X _i,

The probability-distribution function of original DCT coefficient x is

$P\left(x\right|X_i)=\frac{P\left(X_i\right|x)P\left(x\right)}{P\left(X_i\right)}$

Wherein $P\left(x\right|X_i)=\left\{\begin{array}{cc}1& \mathrm{ifx}\&Element;[a,b]\\ 0& \mathrm{otherwise}\end{array}\right..$ Quantization parameter X _ion mean square error be

${\mathrm{\&epsiv;}}_i^2=\frac{1}{P\left(X_i\right)}{\&Integral;}_a^{b}P\left(x\right|X_i){(X_i-x)}^2\mathrm{dx}$

Intra and Inter block adopts different quantization tables, so the quantized interval of correspondence is not identical yet.Interval (a, b) depends on the product of quantization parameter m_quant and quantization parameter table.

I frame is reference frame, and macro block wherein all adopts intraframe coding, and we are by the mean square error of all pieces be added, be the MSE of this two field picture.B frame and P frame are encoded predicted frame, and wherein macro block has four kinds of coded systems.The quantization error 1 of our cumulative wherein intra and inter coded macroblocks and ε 2, the macroblock number N1 of Intra coding mode, Inter coded macroblocks block number N2, then total quantization error approximate evaluation is

ε＝N*(ε1+ε2)/(N1+N2)

Wherein N is the number of macro block in B or P frame.

Comparing with traditional video quality evaluation method without ginseng appraisal procedure of the video quality that the present invention proposes, have the following advantages:

1, parameterless assessment technology is adopted.When assessing video quality, do not need, with reference to original video sequence, to be applicable to the application scenarios that great majority can not get original video sequence, such as, the highway video monitoring system used in public security industry, net bar video frequency monitor system etc.;

2, owing to adopting objective algorithm to carry out estimating without participating in evaluation and electing of video quality, without the need to manual intervention, evaluation result not being affected by artificial subjective factor, has more objectivity;

3, owing to adopting computer generation for estimator, can be very large to data volume faster, chronic video sequence carries out comprehensively, the assessment of statistics.

Those of ordinary skill in the art will be appreciated that, above specification is only one or more execution modes in the numerous embodiment of the present invention, and not uses limitation of the invention.Any equalization for the above embodiment changes, modification and the equivalent technical scheme such as to substitute, as long as spirit according to the invention, all will drop in scope that claims of the present invention protect.

Claims (7)

1. a nothing ginseng appraisal procedure for video quality, it is characterized in that, described appraisal procedure is by setting up associating between compression samples with original image samples, and estimate accurate DCT probability distribution, it comprises the following steps:

Step 1: extract quantization parameter and the DCT coefficient value after quantizing from compressed video data stream, by the distributed area estimating original DCT coefficient in described quantization parameter;

Step 2: by the original DCT coefficient that extracts in described data flow and the sample comparison consistent with compression parameters in training set of described quantization parameter, when comparing, first cluster is carried out to the sample in training set, and then estimate the probability distribution parameters of original DCT coefficient;

Step 3: integration is carried out to the PDF distribution of original DCT coefficient, produces the objective estimation to compression noise with this;

Carry out signal-to-noise ratio (SNR) estimation in this step to comprise the following steps:

S601: start video decode;

S602: information extraction, and quantization parameter is supplied to step S607;

S603: by video data frame arrangement, original DCT coefficient is recombinated;

S604:DCT data block is extracted, and original DCT coefficient is supplied to step S607;

S605: priori schema extraction;

S606: block mode mates, and distribution function λ estimated value is supplied to step S607;

S607: the integral operation of signal-to-noise ratio (SNR) estimation, thus the estimated value of trying to achieve PSNR.

2. the nothing ginseng appraisal procedure of video quality as claimed in claim 1, is characterized in that:

Described original DCT coefficient meets laplacian distribution

Wherein (u, v) is DCT coefficient coordinated indexing, and x is the value of DCT coefficient, and λ (u, v) is Laplce's controling parameters.

3. the nothing ginseng appraisal procedure of video quality as claimed in claim 1, is characterized in that:

With the distributed constant block of 64 distribution function λ (u, v) coefficient compositions for target function, will there are multiple data block clusters of identical described original DCT coefficient distribution characteristics together,

Described clustering algorithm is:

Wherein m is for being greater than 1 integer, i.e. cluster numbers, for x _ibe under the jurisdiction of the degree of membership of a jth set; x _ifor the measurement data of multidimensional, c _jfor the center of individual bunch of jth, || || be distance measure function, measure x _iwith c _jgeometric distance, N, C are corresponding variable parameter;

After setting cluster numbers m, clustering algorithm, by recursive calculation, is asked for and is made described J _mminimum optimal classification combination.

4. the nothing ginseng appraisal procedure of video quality as claimed in claim 3, is characterized in that:

First take out multiple original image samples, the true λ of original DCT coefficient is calculated by multiple original image samples, original image samples is quantized compression, produce different compression samples, the probability distribution of original DCT coefficient is estimated again by described compression samples, and the λ estimated and true λ is contrasted, the mean and variance of the difference of the λ estimated described in calculating and described true λ.

5. the nothing ginseng appraisal procedure of video quality as claimed in claim 1, is characterized in that:

The image block that compressed picture blocks is consistent with compression parameters in the sample of priori storehouse contrasts, to extract the lambda parameter of original DCT coefficient distribution, first read the coding mode of decompressed data block and described quantization parameter, read in corresponding described priori library information subsequently.

6. the nothing ginseng appraisal procedure of video quality as claimed in claim 5, is characterized in that:

First by predictive frame complete decoding, then call dct transform function, convert thereof into as original DCT coefficient, the quantization error on original DCT coefficient is produced by Inter quantization table, when module is mated, adopts the sample in corresponding inter priori storehouse to mate.

7. the nothing ginseng appraisal procedure of video quality as claimed in claim 1, is characterized in that:

Estimate stochastic variable x _ibe quantified as X _iprobability, and the probability-distribution function obtaining original DCT coefficient x is

Wherein (a, b) is quantized interval, X _ifor quantization parameter.