CN101127926A - Image quality evaluation method based on multi-scale geometric analysis - Google Patents

Abstract

The utility model discloses an image quality evaluation method based on multi-scale geometric analysis, mainly solving the problem that consistency is poor between image quality objective evaluation value and subjective evaluation value, comprising (1) The scale and direction sub-band decomposition is performed on reference image and evaluated image using multi-scale analysis method; (2) All decomposed sub-bands are weighted using contrast sensitivity function; (3) The vision perception threshold value of all coefficients are determined based on human eye vision perception property; (4) The proportion of coefficient which is more than the vision perception threshold value in sub-bands of every direction of reference image and evaluated image is calculated respectively; (5) The absolute value difference of the proportion of coefficient which is more than the vision perception threshold value in sub-bands of every direction of reference image and evaluated image is recorded, and then the evaluation measurement of image quality is defined based on the proportion absolute value difference. The utility model has the advantages of simple structure, small transmission data size, low calculating complexity, good consistency with subjective evaluation, and applicability to validity evaluation for image processing method.

Description

Image quality evaluating method based on multi-scale geometric analysis

Technical field

The invention belongs to technical field of image processing, particularly a kind of evaluating method that relates to picture quality can be used for the storage of image compression, image, Image Communication, in the fields such as image detection to the evaluation process of picture quality.

Background technology

Image, can directly or indirectly act on human eye and and then produce the entity of vision with multi-form and means observation objective world and obtain with various observation systems.The human information that obtains from the external world has 75% to obtain from image approximately.Along with the development of signal processing and computer science and technology, Image Engineering also become one abundant in content and develop subject rapidly.A picture system comprises collection, demonstration, storage, communication, processing and the analysis of image.It is widely used in the every field in the national economy, as: scientific research, industrial production, health care, education, amusement, management and field such as communicate by letter, to promoting social development, improve people's living standard and all play important effect.Though image technique has been obtained development rapidly, but some is compromise still making in the design of image processing algorithm and equipment under the present technical merit, as trading off between the scope of compromise, the brightness of compromise, the spatial resolution between temporal resolution and the noise sensitivity and picture size and the exponent number.After making certain selection therein, will have influence on the sense organ of reconstructed image.In order to obtain optimum selection, what kind of influence is the result who is necessary to know these selections how cause can for the sense organ of reconstructed image.By method for objectively evaluating image quality, can effectively assess some image processing methods, finally obtain a better image effect.

The research of present digital picture quality evaluation can be divided into two kinds of diverse methods: subjective assessment and objective evaluation.

First kind mainly is to come assess image quality by subjective experiment.A typical method is the subjective evaluation method of the television image of International Telecommunications Union (ITU, International Telecommunications Union) proposition.Subjective assessment experiment is meant, under certain condition (image source, display device and watch condition etc.), two width of cloth pictures is provided simultaneously for the beholder, and wherein a width of cloth is an original image, and another width of cloth is a distorted image.Original image is without any damage, and distorted image has distortion and also may not have, i.e. distortion is zero.Should comprise ordinary people and image professional and layman for the beholder.Also to a large amount of score data be added up (average, standard deviation, 95% confidence interval etc.) at last.The result of subjective assessment has two kinds of method for expressing: a kind of is that MOS (Mean Opinion Score) is expressed in absolute scoring, promptly represents the absolute mass of distorted image; Another kind is that difference is expressed DMOS (Difference Mean Opinion Score), promptly represents the absolute difference of distorted image and original image evaluation achievement.

Image is for the people watches, thereby the subjective experiment evaluation method is the most accurate and effective method of assess image quality, but also has important disadvantages, i.e. subjective assessment experiment is very consuming time.In the reality, need the data volume of experiment very big, and all will again experimentize when doing the design alternative that makes new advances at every turn, and the subjective assessment experiment can only be tested the image pattern of limited quantity.Therefore, this method is difficult to use in practice.People press for the objective image quality evaluating method of design and be similar to and reflect subjective feeling, evaluation method---the digital picture quality method for objectively evaluating of Here it is second kind of digital picture quality.

The objective evaluation of digital picture quality is more and more paid attention to by people, and has formed many complete and effective algorithm systems.From existing document, according to the reference degree to original image, the objective evaluation of picture quality can be divided into three kinds: full reference type (FR, Full Reference), partial reference type (RR, Reduced Reference) and no reference type (NR, No Reference).

1) full reference type

So-called full reference type that is to say that the prerequisite of algorithm is that original image is fully known, and think without any distortion, be used as the reference of estimating the distorted image quality.Present most of image quality evaluation algorithm all belongs to full reference type.

2) partial reference type

Because the reference image data amount is often bigger, be not easy to transmission and storage in some cases, particularly in the communications field, very high to the requirement of bandwidth.For can be in this case the quality of evaluation map picture preferably still, the image quality evaluation of RR type becomes one of focus of people's research.Partial reference type is between full reference type and does not have a kind of image quality evaluating method between the reference type.The visually-perceptible quality that this method only utilizes the information of the original image of part to come the distortion estimator image.VQEG (Video Quality Experts Group) is with its direction as future development at present.Very useful at some application RR type image quality evaluating method.For example, in the real-time video communication system, can come the control stream resource, thereby satisfy different needs by the monitoring image quality.The RR image quality evaluating method of a success must be obtained good balance on the precision of prediction of RR characteristic rate and picture quality.This be because, if RR characteristic rate is big more, the information that can comprise reference picture is just many more, the prediction that obtains will be accurate more, but this also can give and transmits these parameters and cause very big burden; On the contrary, more little being easy to more of data volume transmitted, but final prediction also can be poor more.

3) no reference type

The common ground of full reference type and partial reference type method is, they all rely on an original and undistorted image as a reference whole or in part.No reference type method is a kind of original image that do not need, directly the method that distorted image is estimated.In the last few years, the image quality evaluation of NR type was paid close attention to by more and more scholars.Thereby this to be one be very significant research direction.

Full reference type method biggest advantage is exactly accurate to the distorted image prediction of quality, yet full reference type method requires to have or not the original image of any distortion, and this all is difficult to realize in a lot of practical applications.And concerning no reference type method, on the one hand, no reference type method is a kind of very difficult work, and existing method effect is mostly not ideal enough, on the other hand, existing no reference type method all has very strong precondition: promptly will know distortion mode or types such as image in advance, this also is difficult to realize in actual applications.The partial reference type method has been utilized the Partial Feature information of original image, when not knowing the image fault type, still can make evaluation and test more accurately to distorted image.

Traditional image quality evaluating method Y-PSNR (PSNR, Peak Signal-to-Noise-Ratio) etc. is that performance has significant limitation from the angle of the pure mathematics statistics to the error between the pixel of image.Recently human visual system (HVS, HumanVisual System) is introduced into the new method of studying image quality evaluation, and the main effect of human eye is in order to extract the structural information in the visual field, and the human visual system also is highly suitable for this purpose.Z Wang etc. has proposed in view of the above based on structural similarity (SSIM, Mean Structural Similarity) full-reference image quality evaluating method " Z Wang; A C Bovik; H RSheikh; and E P Simoncelli.Image quality assessment:from error visibility to structural similarity[J] .IEEE Trans.on Image Processing.2004,13 (4): 600-612. " experimental result shows that this method and subjective assessment have good correlation.But this method is to carry out at the correlation of local pixel in the image, and therefore the image feature information that extracts just is not sufficiently complete.And then, in " Z Wang.and Simoncelli E.P.Reduced-reference image qualityassessment using a wavelet-domain natural image statistic model.Human Vision and ElectronicImaging X.Proc.Jan.2005; 5666:149-159. ", utilize the statistical model of natural image to carry out image quality evaluation (RR-WISM, Reduced-Reference Image Quality Assessment Using A Wavelet-Domain NaturalImage Statistic Model).This method satisfies the characteristic of generalized Gaussian distribution at the wavelet sub-band coefficient of natural image, wavelet sub-band coefficient with distorted image approaches this distribution, according to error of fitting the distorted image quality is estimated then, obtained good evaluation result, but this method structure is complicated, the calculation cost height, and, solution evaluation and test effect that this method is not optimum and the contradiction between the transmitted data amount cost.Only utilize the statistical independence of sub-band coefficients because the coefficient of wavelet sub-band distributes, and nonindependence is present in all directions, yardstick, position between the coefficient, so its effect and not ideal enough.Characteristic information in the most effective extraction image of the key problem of reference image quality evaluation method comes being evaluated and tested by altimetric image by the situation of change of characteristic information in the movement images.So how effectively the characteristic information of extraction and statistical picture becomes the key to image quality evaluation.

Summary of the invention

The objective of the invention is to overcome the deficiency of above-mentioned prior art, a kind of image quality evaluating method based on multi-scale geometric analysis is provided, to solve in traditional partial reference type image quality evaluating method the contradiction of evaluation and test accuracy and required transmission amount of information, realize simpler, more effective, image quality evaluating more accurately.

The technical scheme that realizes the object of the invention is: according to the partial reference type image quality evaluation system, adopt the transform method of multi-scale geometric analysis, utilize human-eye visual characteristic, the sub-band coefficients that conversion obtains is carried out the vision mask, define corresponding visually-perceptible threshold value, add up in each subband situation of change, obtain the evaluation of image is estimated greater than the coefficient of visually-perceptible threshold value.

Detailed process is as follows:

(1) utilize the multi-scale geometric analysis method that reference picture is become different yardstick and the subband on the direction with tested picture breakdown;

(2) each sub-band coefficients is carried out the normalization weighting, and sets visually-perceptible threshold value T according to the sub-band coefficients of reference picture,

$T=\frac{\mathrm{\α}}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\sqrt{\frac{1}{N-1}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(x_{i,j}-{\stackrel{\‾}{x}}_i)}^2}$

In the formula, x _{I, j}Be j coefficient of i direction subband on the fine dimension after the subband weighting,

The subband weighting

The average of i the direction sub-band coefficients in back; M is the number of selected direction subband after the subband weighting, and N is total number of the coefficient in each subband;

(3) calculate reference picture and accounted for the proportion P of all coefficients in each corresponding yardstick of altimetric image and the direction subband greater than the coefficient of described visually-perceptible threshold value T respectively _R(n) and P _D(n),

$P_R\left(n\right)=\frac{R_T\left(n\right)}{R\left(n\right)}$

$P_D\left(n\right)=\frac{D_T\left(n\right)}{D\left(n\right)}$

In the formula, R _T(n) and D _T(n) be respectively reference picture and by in n the subband of altimetric image greater than the number of visually-perceptible threshold value T, R (n) and D (n) are the number of all coefficients in its corresponding subband;

(4) statistical-reference image and by the absolute difference total amount S of visually-perceptible coefficient proportion in each corresponding subband of altimetric image,

$S=\underset{n=1}{\overset{L}{\mathrm{\Σ}}}|P_R\left(n\right)-P_D\left(n\right)|$

L is the total number that is selected subband;

(5), determine to be estimated Q by the quality evaluation of altimetric image according to described absolute difference total amount S

$Q=\frac{1}{1+{\log }_2(\frac{S}{Q_0}+1)}$

In the formula, Q ₀Be the regulatory factor of Q dynamic range, the Q span between 0～1,

Along with the increase of Q value, picture quality is also corresponding will to be improved.

The present invention utilizes the variation of sub-band coefficients that picture quality is estimated owing to adopted the multi-scale geometric analysis method that image is decomposed on yardstick and direction, compared with prior art, has following advantage:

A) the evaluation and test performance has a distinct increment, and is better with the consistency of human visual perception.Be applied to the present invention and be example with the HWD conversion with the jpeg image, it is evaluated and tested accuracy, monotonicity, is respectively 0.9702,0.9473,0.03 from going out rate, the evaluating method of existing relatively RR-WISM, its accuracy, monotonicity, improved 0.05,0.04 respectively, reduced by 0.11 from going out rate,, its accuracy, monotonicity, improved 0.01 respectively with respect to existing SSIM evaluating method, 0.01, reduced by 0.01 from going out rate.

B) simple in structure, amount of calculation is little.With the wavelet transformation is example, utilizes wavelet transformation that image is carried out three grades of decomposition, shown in Fig. 2 (a), under the same conditions, to identical a pair of 488 * 610 images, evaluates and tests the required time with the present invention and only evaluates and tests 1/20 of the time for existing RR-WISM method;

C) required information transmitted amount is less.Being applied to the present invention with the wavelet conversion is example, and the amount of information that the present invention need transmit to arbitrary width of cloth image only is 84bit, is about half of existing RR-WISM method.

Description of drawings

Fig. 1 is that the present invention evaluates and tests the process schematic diagram;

Fig. 2 is a picture breakdown schematic diagram of the present invention, wherein

Fig. 2 (a) is applied to when of the present invention when wavelet transformation, and picture breakdown and subband are chosen schematic diagram;

Fig. 2 (b) is applied to when of the present invention when the contourlet conversion, and picture breakdown and subband are chosen schematic diagram;

Fig. 3 is the nonlinear fitting curve chart of several image quality evaluating methods and subjective MOS value, wherein:

Figure (a1) when adopting PSNR evaluation and test jpeg image to the prediction curve figure of MOS;

Figure (a2) when adopting PSNR evaluation and test JPEG2000 image to the prediction curve figure of MOS;

Figure (b1) when adopting classical partial reference type method RR-WISM evaluation and test jpeg image to the prediction curve figure of MOS;

Figure (b2) when adopting classical partial reference type method RR-WISM evaluation and test JPEG2000 image to the prediction curve figure of MOS;

Figure (c1) when adopting classical full reference type method MSSIM evaluation and test jpeg image to the prediction curve figure of MOS;

Figure (c2) when adopting classical full reference type method MSSIM evaluation and test JPEG2000 image to the prediction curve figure of MOS;

Figure (d1) evaluates and tests the prediction curve figure of jpeg image to MOS when utilizing WBCT for the present invention;

Figure (d2) evaluates and tests the prediction curve figure of JPEG2000 image to MOS when utilizing WBCT for the present invention.

Embodiment

Core concept of the present invention is to utilize effective multi-scale geometric analysis method, obtains the rarefaction representation of image, extracts the texture information and the directional information of image.By contrasting the perceptibility mask to sensitivity coefficient normalization, set visually-perceptible threshold value T according to the sub-band coefficients of reference picture, calculate reference picture and by in the altimetric image all directions subband greater than the proportion of the shared whole coefficients of the coefficient of visually-perceptible threshold value T, and statistics and its difference of comparison, the quality evaluation that obtains image is estimated.

As shown in Figure 1, place transmitting terminal to handle reference picture, to be placed receiving terminal to handle by altimetric image, reference picture is handled the visually-perceptible threshold value obtain and is transferred to receiving terminal greater than the proportion that the coefficient of visually-perceptible threshold value accounts for all coefficients by auxiliary channel, obtain by the evaluation result of altimetric image by reference picture relatively with by the variation of altimetric image sub-band coefficients at receiving terminal.

Key step of the present invention is:

1) image preliminary treatment

Utilize the multi-scale geometric analysis method to reference picture with carried out the sub-band division of yardstick and direction by altimetric image, this decomposes employing Beamlet conversion, the Wedgelet conversion, the adaptive analysis method of Bandelet conversion, perhaps Wavelet conversion, Steerable Pyramid conversion, the Curvelet conversion, the Contourlet conversion, Contourlet conversion based on Wavelet, the non-self-adapting analytical method of Wavelets and directional filter group mixing transformation is carried out, and obtains the point of image respectively, straight line, smooth closed curve, the smooth contoured zone, straightway, wedge shape, the cross linear feature.Be transformed to example with contourlet, carrying out three layers of Laplacian Pyramid decomposes, shown in Fig. 2 (b), each layer adopts 4 grades, 3 grades, 3 grades directional filters to decompose respectively from fine to coarse, character according to directional filter, only select the director of half to bring extraction graphical rule information and directional information, i.e. the subband part that marks with white dashed line frame and numeral among Fig. 2 (a), Fig. 2 (b).

2) subband normalization weighting:

(a) all subbands after decomposing are utilized the normalized frequency f of image _nSample frequency f with image _s, obtain the spatial frequency f of signal, promptly

f＝f _n·f _s(1)

In the formula, f _n, f _sObtain by following formula respectively

f _n＝3/2 ⁿ⁺²(2)

In the formula, n=1,2 ... for being weighted subband place yardstick,

V is an observed range, and unit is a rice, here v get height of display 2-2.5 doubly.

R is an exploration on display resolution ratio, and unit is a pixel/inch, and display is of a size of 21 inches, and resolution is 1024 * 768, then has $r=\sqrt{{1024}^2+{768}^2}/21=61$ Pixel/inch.

(b) utilize spatial frequency f, and utilize contrast sensitivity function,

$H\left(f\right)=2.6(0.192+0.114f)e^{[-{\left(0.114f\right)}^{1.1}]}$

Subband after decomposing is carried out the normalization weighting, make the coefficient in the different frequency domain subbands identical visually-perceptible degree be arranged human eye.

3) set the visually-perceptible threshold value

Human eye is had under the prerequisite of same perceived at different frequency coefficients,, determines the visually-perceptible threshold value T of all coefficients according to the following procedure according to the human eye vision apperceive characteristic:

(a) each yardstick that utilizes following formula to calculate after multi-scale geometric analysis and subband weighting, to obtain and the standard deviation T of direction subband _i,

$T_i=\sqrt{\frac{1}{N-1}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(x_{i,j}-{\stackrel{\‾}{x}}_i)}^2}---\left(5\right)$

In the formula, x _{I, j}Be j coefficient of selected i direction subband after the subband weighting,

Average for i direction sub-band coefficients after the subband weighting; N is total number of the coefficient in each subband;

(b) choose that M director band on the fine dimension carries out weighting after the subband weighting, and calculate all T _iAverage Th,

$\mathrm{Th}=\frac{1}{M}\underset{j=1}{\overset{M}{\mathrm{\Σ}}}{T}_i---\left(6\right)$

(c) the average Th that will calculate gained is set at visually-perceptible threshold value T

T＝α·Th(7)

In the formula, α is a weighting parameters.

4) to coefficient normalization greater than visually-perceptible threshold value T

Calculate in reference picture and the distorted image all directions subband the shared proportion of coefficient respectively greater than visually-perceptible threshold value T.

$P_R\left(n\right)=\frac{R_T\left(n\right)}{R\left(n\right)}---\left(8\right)$

$P_D\left(n\right)=\frac{D_T\left(n\right)}{D\left(n\right)}---\left(9\right)$

In the formula, R _T(n) and D _T(n) be respectively the number of visually-perceptible coefficient in n the subband of reference picture and distorted image, R (n) and D (n) are the number of all coefficients in its corresponding subband;

For reducing volume of transmitted data, only with visually-perceptible threshold value T and P _R(n) characteristic information as the reference image is transferred to tested image processing end, and tries to achieve P at tested image processing end _D(n);

5) the absolute value difference of statistics proportion:

Add up P in the selected subband by following formula _R(n) and P _D(n) absolute difference and S

$S=\underset{n=1}{\overset{L}{\mathrm{\Σ}}}|P_R\left(n\right)-P_D\left(n\right)|---\left(10\right)$

L is the total number that is selected subband, gets L=24 in the experiment.

6) image quality evaluation is estimated

Q is estimated in evaluation according to absolute value difference S definition picture quality

$Q=\frac{1}{1+{\log }_2(\frac{S}{Q_0}+1)}---\left(11\right)$

In the formula, Q ₀Be the regulatory factor of Q dynamic range, the Q span is between 0～1, and along with the increase of Q value, picture quality is also corresponding will to be improved, and gets Q in the experiment ₀=0.1.

Advantage of the present invention can further specify by following experiment:

This experiment is carried out on the U.S. LIVE of TEXAS university image quality measure database, this database comprises the RGB image of 29 panel height resolution and a series of distorted images that obtain by JPEG and JPEG2000 compression, jpeg image 175 width of cloth wherein, JPEG2000 image 169 width of cloth give the information such as subjective assessment MOS of every width of cloth image simultaneously.In order to test the picture quality objective evaluation result and the consistency of subjective perception that the present invention proposes, we have selected following three kinds of measurement criterions: (1) coefficient correlation, reflected the accuracy that method for objectively evaluating is predicted; (2) Spearman rank order correlation coefficient has reflected the monotonicity of objective evaluation result prediction; (3) from going out rate, reflected the method for objectively evaluating estimation stability.Table 1 has provided the present invention and PSNR, the MSSIM of full reference type and the contrast and experiment of RR-WISM method.

Table 1PSNR, MSSIM, RR-WISM and subjective and objective consistency evaluation ratio of the present invention

As can be seen from Table 1, the existing relatively method of the present invention has good superiority: 1) higher accuracy of forecast is arranged, and promptly coefficient correlation is 0.9728 and 0.9565, and is bigger than existing method coefficient correlation; 2) stricter prediction monotonicity is arranged, promptly rank order correlation coefficient 0.9527 and 0.9390 is bigger than existing method rank order correlation coefficient; 3) better stability is arranged, promptly, lower from going out rate than existing method from going out rate 0.0457 and 0.0414.

Fig. 3 shows that the present invention approaches effect than the nonlinear fitting curve that existing method and subjective MOS value have better.Curve is the optimum Match curve of subjective observation value MOS and objective evaluating value, the objective evaluating value of abscissa presentation video, ordinate is the subjectivity evaluation and test value of image, JPEG or JPEG2000 image in "+" data representing image storehouse, the distribution of "+" shows that more near the curve among the figure this method is effective more.

Figure (a1) is to the prediction curve of subjective assessment value MOS during with existing method PSNR evaluation and test jpeg image, figure (a2) is to the prediction curve figure of subjective assessment value MOS during with existing method PSNR evaluation and test JPEG2000 image, the distribution of "+" is bigger as can be seen from (a1), (a2), and deflection curve is distant.

Figure (b1) is to the prediction curve figure of subjective assessment value MOS when evaluating and testing jpeg image with classical partial reference type method RR-WISM.Figure (b2) is to the prediction curve figure of subjective assessment value MOS when evaluating and testing the JPEG2000 image with classical partial reference type method RR-WISM.From figure (b1), (b2) as can be seen "+" be distributed in afterbody the time approximating curve relatively, but whole trend still is that deflection curve is distant.

Figure (c1) is to the prediction curve figure of subjective assessment value MOS during with classical full reference type method MSSIM evaluation and test jpeg image.Figure (c2) is to the prediction curve figure of subjective assessment value MOS during with classical full reference type method MSSIM evaluation and test JPEG2000 image.From figure (c1), figure (c2) as can be seen "+" be distributed in relatively approximating curve of some regional areas, but that other departs from is very big, therefore says still to be that deflection curve is distant on the whole.

Figure (d1) is that the present invention is when utilizing the WBCT conversion to carry out image quality evaluation, at the prediction curve figure of jpeg image to subjective assessment value MOS, figure (d2) is that the present invention is when utilizing the WBCT conversion to carry out image quality evaluation, at the prediction curve figure of JPEG2000 image to subjective assessment value MOS.From figure (d1), figure (d2) the distribution approximating curve relatively on the whole of "+" as can be seen, only have that extremely other deflection curve is bigger, but still approximating curve very on the whole.From the contrast of above figure as can be seen, the present invention is significantly improved than existent method, and has consistency preferably with the subjective assessment value.

Claims (4)

1. image quality evaluating method based on multi-scale geometric analysis, by reference picture and tested image comparison are carried out, detailed process is as follows:

(1) utilize the multi-scale geometric analysis method that reference picture is become subband on different scale and the direction with tested picture breakdown;

(2) each sub-band coefficients is carried out the normalization weighting, and is according to the sub-band coefficients setting visually-perceptible threshold value of reference picture,

$T=\frac{\mathrm{\α}}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}\sqrt{\frac{1}{N-1}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(x_{i,j}-{\stackrel{\‾}{x}}_i)}^2}$

X in the formula _{I, j}Be j coefficient of i direction subband on the fine dimension after the subband weighting,

Be the average of i direction sub-band coefficients after the subband weighting,

M is the number of selected direction subband, and N is total number of the coefficient in each subband;

(3) calculate reference picture and accounted for the proportion P of all coefficients in each corresponding yardstick of altimetric image and the direction subband greater than the coefficient of described visually-perceptible threshold value T respectively _R(n) and P _D(n), promptly to coefficient normalization greater than visually-perceptible threshold value T,

$P_R\left(n\right)=\frac{R_T\left(n\right)}{R\left(n\right)}$

$P_D\left(n\right)=\frac{D_T\left(n\right)}{D\left(n\right)}$

In the formula, R _T(n) and D _T(n) be respectively reference picture and by in n the subband of altimetric image greater than the number of visually-perceptible threshold value T, R (n) and D (n) are the number of all coefficients in its corresponding subband;

(4) statistical-reference image and by the absolute difference total amount S of visually-perceptible coefficient proportion in each corresponding subband of altimetric image,

$S=\underset{n=1}{\overset{L}{\mathrm{\Σ}}}|P_R\left(n\right)-P_D\left(n\right)|$

In the formula, L is the total number that is selected subband;

(5), determine to be estimated Q by the quality evaluation of altimetric image according to described absolute difference total amount S

$Q=\frac{1}{1+{\log }_2(\frac{S}{Q_0}+1)}$

In the formula, Q ₀Be the regulatory factor of Q dynamic range, the Q span is between 0～1.

2. method according to claim 1, it is characterized in that utilizing the multi-scale geometric analysis method that picture breakdown is become different yardstick and the subband on the direction, adopt the Beamlet conversion, the Wedgelet conversion, the adaptive analysis method of Bandelet conversion, perhaps Wavelet conversion, Steerable Pyramid conversion, the Curvelet conversion, the Contourlet conversion, Contourlet conversion based on Wavelet, the non-self-adapting analytical method of Wavelets and directional filter group mixing transformation, decompose, obtain the point of image respectively, straight line, smooth closed curve, the smooth contoured zone, straightway, wedge shape, the cross linear feature.

3. method according to claim 1, it is as follows to it is characterized in that the antithetical phrase band carries out the process of normalization weighting:

1) utilizes the normalized frequency f of image _nSample frequency f with image _s, the space that obtains signal is f frequently, promptly

f＝f _n·f _s

In the formula, f _n, f _sObtain by following formula respectively

f _n＝3/2 ⁿ⁺²

Wherein, n=1,2 ... for being weighted subband place yardstick, v is an observed range, and unit is a rice,

R is an exploration on display resolution ratio, and unit is a pixel/inch;

2) utilize spatial frequency f, and utilize contrast sensitivity function,

$H\left(f\right)=2.6(0.192+0.114f)e^{[-{\left(0.114f\right)}^{1.1}]},$ Subband after decomposing is carried out the normalization weighting.

4. method according to claim 1 is characterized in that the assignment procedure of visually-perceptible threshold value T is as follows:

1) each yardstick that utilizes following formula to calculate after multi-scale geometric analysis and subband weighting, to obtain and the standard deviation T of direction subband _i,

$T_i=\sqrt{\frac{1}{N-1}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(x_{i,j}-{\stackrel{\‾}{x}}_i)}^2}$

In the formula, x _{I, j}Be j coefficient of selected i direction subband after the subband weighting,

Average for i direction sub-band coefficients after the subband weighting;

N is total number of the coefficient in each subband;

2) choose that M director band on the fine dimension carries out weighting after the subband weighting, and calculate all T _iAverage be,

$\mathrm{Th}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{T}_i$

3) the average Th that will calculate gained is set at visually-perceptible threshold value T

T＝α·Th

In the formula, α is a weighting parameters.

Images