CN105592312A - Reference-free video image quality estimating method based on reconstruction - Google Patents

Abstract

The invention belongs to the field of video image signal processing and especially relates to improvement and optimization of a video image quality estimating algorithm. A reference-free video image quality estimating method based on reconstruction is characterized by reconstructing a part of statistical magnitude of an original image by using the spatial-frequency domain characteristic of a video image to be measured, correcting a mean square error estimating model, and greatly decreasing the influences of image complexity and sharpness on the accuracy of the mean square error estimating model; reconstructing the sub-band zero-coefficient value of each DCT of the area of the video image to be measured; separately computing the mean square error value of each DCT sub-band according to a zero coefficient and a nonzero coefficient so as to obtain a peak signal-to-noise ratio of the area of the video image to be measured. The method does not need the information and the coding information of any original image, is completely independent of an encoding and decoding algorithm, and may greatly increase the accuracy of video image quality estimation compared with a conventional algorithm.

Description

A kind of based on reconstruct without reference video image quality estimation method

Technical field

The invention belongs to video signal process field, relate in particular to the improvement of video image quality algorithm for estimating and excellentChange.

Background technology

Video image quality estimates it is a process calculating testing image quality, and its main purpose is to solve video imageQuality evaluation problem. Video image quality estimation technique has been widely used in video broadcasting, video request program, network video at presentFrequently, video monitoring and video image compression are processed and social networks application. In video on-demand system, server is by connecingReceiving end is carried out film source adjustment and coding parameter adjustment to the assessment feedback of video quality, ensures user's view reception effect; Look in real timeFrequently, in/Image Communication application, transmitting terminal improves communication according to the adjustment in real time of quality evaluation feedback, the Optimized Coding Based parameter of receiving terminalQuality; Encoder, by the quality of predictive code rear video image, is adjusted coding parameter and is reached the most optimized parameter selection. TotalIt, video image quality estimation technique is deep into each application of encoding video pictures, communication and monitoring already, arrives greatly matterAmount evaluating system controlled encoder parameter, communication bandwidth distribution, memory device distribution etc., little of a quality as encoderPrediction module is for optimum block size and the intra/inter-predictive mode selection of frame etc., and its range of application is very wide.

Video image quality estimation technique is according to there being non-reference picture can be divided into full reference mass estimation, partial reference qualityEstimate and without reference mass estimation technique. Full reference mass estimation technique is to calculate between testing image and original image pixelsMean square error and (MSE) and calculate Y-PSNR (PSNR) and weigh the actual mass of testing image. This Technology Need is originalImage is as with reference to standard, and the scope of application is the narrowest, is only applicable to coding side or server end the video image quality sending is enteredRow prediction, is mainly used in compression coding parameter adjustment. Partial reference quality estimation technique is to utilize partial original image information to estimateMean square error between meter testing image and original image pixels and (MSE) and calculate Y-PSNR (PSNR) and estimate to be measuredThe actual mass of image. The information of this Technology Need partial original image is as reference, and the more full reference mass of the scope of application is estimatedExtensively, not being only applicable to that coding side or server end predict also the video image quality sending can be for receiving terminal toThe quality that receives image is estimated. But the necessary translator unit raw video image data of this technology, the mass data of certainly will uprushing passesDefeated, occupy more that multi-band is wide maybe needs to set up another dedicated transmissions link, be not suitable for network video image communication and transmission bandwidthLimited video image communication. Not have any raw information of raw video image to go to estimate completely without reference mass estimation techniqueThe quality of meter testing image. Because the information that does not need raw video image is removed estimated quality, the range of application of this technology is obviousExtensively in full reference and partial reference quality estimation technique, applicable in transmitting terminal, transmission network and receiving terminal computed imageQuality and without extra other original reference information that transmit. Without reference mass estimation technique be divided into again algorithm based on bit stream withBased on the algorithm of pixel. Wherein, based on bit stream without reference mass estimation technique be utilize from encoding stream, extractVarious parameters put into that prior model is estimated the quality of video image to be measured and without encoding stream complete decoding is obtainedThe pixel point value of testing image. Is pixel value, the DCT coefficient that utilizes testing image based on pixel without reference mass estimation techniqueThereby the features such as distribution are set up the quality of calculation model of mass estimation testing image, without any ginseng extracting in encoding streamNumber.

Estimate mainly by setting up DCT coefficient distributed model without reference mass, then estimate by the DCT coefficient value of testing imageMeter obtains Y-PSNR (PSNR). Therefore, the accuracy of DCT coefficient distributed model becomes without reference mass evaluated error sizeKey point. According to statistical theory, the mean square error of sample can be passed through formula by the distribution characteristics of sample value ${\mathrm{\ϵ}}^2={\∫}_{-\mathrm{\∞}}^{+\mathrm{\∞}}{f}_X\left(x\right|X){(X-x)}^{2}dx=\frac{{\∫}_{-\mathrm{\∞}}^{+\mathrm{\∞}}{f}_X\left(x\right){(X-x)}^{2}dx}{{\∫}_{-\mathrm{\∞}}^{+\mathrm{\∞}}{f}_X\left(x\right)dx}$ Obtain, wherein, f_X(x) be DCT coefficient probability-distribution function, normalDistribute and be similar to Laplace distribution or Cauchy, Laplace distribution function isCauchy distributesFunction isAfter calculating mean square error (MSE), pass throughCalculating peak value letterThe ratio of making an uproar, wherein, n represents the bit number of each pixel.

But two major defects of DCT coefficient distributed model have had a strong impact on the accuracy of estimating:

Laplace distributes or Cauchy distributes is only all the approximate of Image DCT coefficient probability-distribution function, its accuracyBecause often there is the PSNR evaluated error up to 5～6dB in the up and down fluctuations different from image definition of picture material complexity.

In video image clarity lower (conventionally at large quantization step) situation, the zero coefficient ratio of its DCT coefficient willGreatly increase, even part dct transform sub-band coefficients is all zero. Model (1) estimation error in the case will lose efficacy, because ofAnd conventional quantization step and the distortion upper limit are estimated its error, so too high distortion estimator error is caused to testing image qualityEstimate greatly to depart from actual value.

Summary of the invention

The object of the invention is to for the deficiencies in the prior art, provide a kind of based on reconstruct without reference video image matterAmount estimation method, the method utilizes reconfiguration technique to recover to greatest extent the statistical nature of raw video image, and round-off error is estimatedModel makes up the deficiency of conventional algorithm, without any need for coding parameter and original image information, is totally independent of codec, energySignificantly promote compared with conventional method the accuracy that quality is estimated.

Technical scheme of the present invention is:

Utilize the part statistic of sky-frequency domain character reconstituting initial image of video image to be measured, revise mean square error and estimateMeter model, greatly reduces the impact on the model of error estimate degree of accuracy of image complexity and definition; And then reconstruct is to be measured looksFrequently the each subband zero DCT coefficient value of image; Last each DCT subband calculates respectively mean square error by zero coefficient and nonzero coefficient two partsValue, thus the Y-PSNR of video image to be measured obtained.

Based on reconstruct without a reference video image quality estimation method, concrete steps are as follows:

S1, video image is divided into some regions, to process one by one, current processing region is video image to be measured districtTerritory. Video image region to be measured is made to N × N point dct transform and obtain K × N × N DCT coefficient, by this video image region to be measuredSynthetic N × N the subband of all DCT coefficient sets, wherein, N is dct transform size, 4≤N≤16 and be positive even numbers,NT is this region total pixel number;

S2, according to formulaThe quantization step of video image region to be measured described in estimation S1, wherein, X_lTableShow l DCT coefficient value in video image region to be measured, Qs is quantization step, l=1, and 2,3 ..., K × N × N;

Described in S3, statistics S1, N × N subband is reconstructed intoProbability, be designated asCalculate the modulation of each subbandParameterAccording to described R_kCalculate each subband corrected parameterWherein,For the maximum reconstructed value in region to be measured described in S1,Be the deflection of k Subband DCT coefficient,Represent the highest subbandNonzero coefficient ratio, ρ₁Represent lowest sub-band nonzero coefficient ratio, φ_kRepresent k the normalized average DCT coefficient of subbandValue, j is the sequence number of j reconstruction value in subband,Represent that k subband is reconstructed intoProbability,Represent X_jWeightStructure value k=1,2,3 ..., N × N.

S4, calculate the square mean error amount of each subband nonzero coefficient $\begin{array}{c}{\mathrm{\ϵ}}_k^2=\underset{j}{\Σ}\left(C_k\right({\hat{X}}_j){\mathrm{\γ}}_{k}Qs+C_k({\hat{X}}_k\left)\right({\hat{X}}_j^2+{\mathrm{\γ}}_k)\arctan \frac{{\hat{X}}_j+\mathrm{\α}Qs}{{\mathrm{\γ}}_k}\\ -C_k\left({\hat{X}}_k\right)({\hat{X}}_j^2+{\mathrm{\γ}}_k)\arctan \frac{{\hat{X}}_j-(1-\mathrm{\α})Qs}{{\mathrm{\γ}}_k}\\ +C_k\left({\hat{X}}_k\right){\mathrm{\γ}}_k{\hat{X}}_j\log \left(\frac{{({\hat{X}}_j-(1-\mathrm{\α}\left)Qs\right)}^2+{\mathrm{\γ}}_k^2}{{({\hat{X}}_j+\mathrm{\α}Qs)}^2+{\mathrm{\γ}}_k^2}\right))\end{array},$ Wherein, ${\mathrm{\γ}}_k=\frac{\mathrm{\α}\·QS}{tan\frac{{\mathrm{\πP}}_0}{2}},$ P₀For the ratio of each subband zero coefficient,Be that k subband is reconstructed intoIndex modulation correctionParameter, α controls parameter empirical value for quantizing blind area;

S5, estimate the square mean error amount of each subband zero coefficient with Reconstruction Method, concrete steps are as follows:

S51, video image region to be measured is carried out to down-sampling until DCT coefficient zero coefficient ratio slip is less than Φ turnsEnter S52, wherein, Φ is empirical value;

S52, the video image region to be measured that carries out down-sampling described in S51 is carried out to dct transform, calculate each sub-band coefficientsThe probability of value between 0～α Qs, is designated asReconstruct as the each subband zero coefficient in testing image region between 0～α Qs is generalRate, t ∈ (0, α Qs);

S53, according to described in S52Calculate the zero coefficient error amount of each subbandWherein, M represents total number of testing image region discrete cosine transform block;

S6, according to formulaCalculate the mean square error (Mean of m video image region to be measuredSquareError，MSE)；

S7, traversing operation is carried out in remaining region described in S1, repeat S1-S6 until all video image regions are estimated completeAfter try to achieve entire imageSubstitutionObtain evaluating objective quality PSNR value, itsIn, the bit number that n is each pixel.

Further, Φ=10% described in S51.

Further, down-sampling is carried out to minimum wide high size in testing image region described in S51, under stopping, adoptingSample, proceeds to S52.

Further, described in S1, N is 4,8 or 16 any one number.

Further, α=0.5 described in S4.

The invention has the beneficial effects as follows:

Excavate the feature of video image to be measured, the DCT coefficient of reconstituting initial image distributes and each subband nonzero coefficientReconstruction value existsBetween probability, revise MSE estimation model, the value of approaching of the each subband zero coefficient of reconstruct, finally realizes nothingMore accurately estimating of reference video image quality. The present invention is without information and the coded message of any original image, completely independentIn code decode algorithm, more existing algorithm can improve the degree of accuracy that video image quality is estimated greatly.

Brief description of the drawings

Fig. 1 is system block diagram of the present invention.

Detailed description of the invention

Below in conjunction with embodiment and accompanying drawing, describe technical scheme of the present invention in detail.

Embodiment,

As shown in Figure 1, regard view picture video image as a video image region to be measured, concrete steps are as follows:

S1, video image to be measured is made to N × N point dct transform, then by synthetic the DCT coefficient sets of view picture video image N × NIndividual subband, N=8.

S2, according to formulaEstimate quantization step, X in formula_kRepresent any DCT coefficient in testing imageValue.

Described in S3, statistics S1, N × N subband is reconstructed intoProbability, be designated as

Calculate the modulation parameter of each subband

Calculate each subband corrected parameter $\begin{array}{cc}{C}_k\left({\hat{X}}_j\right)={\mathrm{\ρ}}_k\left({\hat{X}}_j\right)R_k,& {\hat{X}}_J=round\left(\frac{\underset{i}{\max }\left(X_i\right)}{Qs}\right)\·Qs.\end{array}$ Wherein,It is k subbandThe deflection of DCT coefficient,Represent the nonzero coefficient ratio of high subband, ρ₁Represent lowest sub-band nonzero coefficient ratio, φ_kTableShow the normalized average DCT coefficient value of k subband.

S4, press formula $\begin{array}{c}{\mathrm{\ϵ}}_k^2=\underset{j}{\Σ}\left(C_k\right({\hat{X}}_j){\mathrm{\γ}}_{k}Qs+C_k({\hat{X}}_k\left)\right({\hat{X}}_j^2+{\mathrm{\γ}}_k)\arctan \frac{{\hat{X}}_j+\mathrm{\α}Qs}{{\mathrm{\γ}}_k}\\ -C_k\left({\hat{X}}_k\right)({\hat{X}}_j^2+{\mathrm{\γ}}_k)\arctan \frac{{\hat{X}}_j-(1-\mathrm{\α})Qs}{{\mathrm{\γ}}_k}\\ +C_k\left({\hat{X}}_k\right){\mathrm{\γ}}_k{\hat{X}}_j\log \left(\frac{{({\hat{X}}_j-(1-\mathrm{\α}\left)Qs\right)}^2+{\mathrm{\γ}}_k^2}{{({\hat{X}}_j+\mathrm{\α}Qs)}^2+{\mathrm{\γ}}_k^2}\right))\end{array}$ Calculate each subband nonzero coefficientSquare mean error amount, wherein,P₀For the ratio of each subband zero coefficient.

S5, the zero coefficient error amount of each subband is reconstructed. Concrete steps are as follows:

S51, to testing image carry out down-sampling until DCT coefficient zero coefficient ratio slip be less than 10% or image under adoptSample is to minimum wide high size, and wherein, the wide height of described minimum is of a size of 44 pixels.

S52, down-sampled images is carried out to dct transform, and calculate the probability of each sub-band coefficients value between 0～α Qs, noteFor

S53, pressCalculate the zero coefficient error amount of each subbandWherein, M representsTotal number of entire image discrete cosine transform block.

S6, press formulaCalculate overall mean square error (MSE), rear substitution formulaObtain evaluating objective quality PSNR value.

S7, input CIF (176 × 144) video image to be measured, carry out quality estimation. As shown in table 1, several nothings are with reference to matterAmount estimation method objective examination degree of accuracy comparison, the present invention all has the degree of accuracy more than 2dB to improve compared with Eden and TBrandao.

Table 1

Claims (6)

1. Based on reconstruct without a reference video image quality estimation method, it is characterized in that, comprise the steps:

   S1, video image is divided into some regions, current processing region is video image region to be measured, to video image to be measured districtTerritory is made N × N point dct transform and is obtained K × N × N DCT coefficient, by all DCT coefficient combinations of described video image region to be measuredBecome N × N subband, wherein, N is dct transform size, 4≤N≤16 and be positive even numbers,NT is the total pixel in this regionNumber;

   S2, according to formulaThe quantization step of video image region to be measured described in estimation S1, wherein, X_lExpression is treatedSurvey l DCT coefficient value in video image region, Qs is quantization step, l=1, and 2,3 ..., K × N × N;

   Described in S3, statistics S1, N × N subband is reconstructed intoProbability, be designated asCalculate the modulation parameter of each subbandAccording to described R_kCalculate each subband corrected parameterWherein,ForThe maximum reconstructed value in region to be measured described in S1,Be the deflection of k Subband DCT coefficient,Represent the non-zero of high subbandCoefficient ratio, ρ₁Represent lowest sub-band nonzero coefficient ratio, φ_kRepresent k the normalized average DCT coefficient value of subband, j isThe sequence number of j reconstruction value in subband,Represent that k subband is reconstructed intoProbability,Represent X_jReconstruction value k=1,2,3,...,N×N,

   S4, calculate the square mean error amount of each subband nonzero coefficient

   $\begin{array}{c}{\mathrm{\ϵ}}_k^2=\underset{j}{\mathrm{\Σ}}\left(C_k\right({\hat{X}}_j){\mathrm{\γ}}_{k}Qs+C_k({\hat{X}}_k)({\hat{X}}_j^2+{\mathrm{\γ}}_k)\arctan \frac{{\hat{X}}_j+\mathrm{\α}Qs}{{\mathrm{\γ}}_k}\\ -C_k\left({\hat{X}}_k\right)({\hat{X}}_j^2+{\mathrm{\γ}}_k)\arctan \frac{{\hat{X}}_j-(1-\mathrm{\α})Qs}{{\mathrm{\γ}}_k}\\ +C_k\left({\hat{X}}_k\right){\mathrm{\γ}}_k{\hat{X}}_j\log \left(\frac{{({\hat{X}}_j-(1-\mathrm{\α}\left)Qs\right)}^2+{\mathrm{\γ}}_k^2}{{({\hat{X}}_j+\mathrm{\α}Qs)}^2+{\mathrm{\γ}}_k^2}\right))\end{array},$ Wherein, ${\mathrm{\γ}}_k=\frac{\mathrm{\α}\·Qs}{tan\frac{{\mathrm{\πP}}_0}{2}},$ P₀For each subband zero coefficientRatio,Be that k subband is reconstructed intoIndex modulation corrected parameter, α controls parameter empirical value for quantizing blind area;

   S5, estimate the square mean error amount of each subband zero coefficient with Reconstruction Method, concrete steps are as follows:

   S51, video image region to be measured is carried out to down-sampling until DCT coefficient zero coefficient ratio slip is less than Φ proceeds toS52, wherein, Φ is empirical value;

   S52, the video image region to be measured that carries out down-sampling described in S51 is carried out to dct transform, calculate each sub-band coefficients value and existProbability between 0～α Qs, is designated asReconstruct probability as the each subband zero coefficient in testing image region between 0～α Qs, t∈(0,αQs)；

   S53, according to described in S52Calculate the zero coefficient error amount of each subbandWherein,M represents total number of testing image region discrete cosine transform block;

   S6, according to formulaCalculate the mean square error (MeanSquare of m video image region to be measuredError，MSE)；

   S7, traversing operation is carried out in remaining region described in S1, repeat S1-S6 until all video image regions are asked after estimatingEntire imageSubstitutionObtain evaluating objective quality PSNR value, wherein, nFor the bit number of each pixel.
2. According to claim 1 a kind of based on reconstruct without reference video image quality estimation method, it is characterized in that:Φ=10% described in S51.
3. According to claim 1 a kind of based on reconstruct without reference video image quality estimation method, it is characterized in that:Described in S51, down-sampling is carried out to minimum wide high size in testing image region, stop down-sampling, proceed to S52.
4. According to claim 1 a kind of based on reconstruct without reference video image quality estimation method, it is characterized in that:Described in S51, the minimum wide height of image down sampling is of a size of 44 pixels.
5. According to claim 1 a kind of based on reconstruct without reference video image quality estimation method, it is characterized in that:Described in S52, use the probability of the each subband zero coefficient in down-sampling technology reconstruct testing image region between 0～α Qs
6. According to claim 1 a kind of based on reconstruct without reference video image quality estimation method, it is characterized in that:α=0.5 described in S4.