CN105844640A - Color image quality evaluation method based on gradient - Google Patents

Abstract

The invention discloses a color image quality evaluation method based on gradient, wherein the color image quality evaluation method mainly settles a problem of low color distortion effect on the color images in a current image quality evaluation method. The color image quality evaluation method comprises the steps of 1), performing color perception change on an original image and a distortion image by means of an S-CIELAB color appearance model, and respectively decomposing the two images to a brightness channel and two chroma channels; 2), performing gradient calculation on each channel by means of linear convolution filtering, obtaining the brightness edges and the chroma edges of the original image and the distortion image; 3), calculating the brightness edge difference and the chroma edge brightness between the original image and the distortion image; and 4) performing linear fusion on the brightness edge difference and the chroma edge difference, thereby obtaining a final quality evaluation value CGBM. The color image quality evaluation method can be used for performing image evaluation on the color image more effectively and more accurately and can be used for processing the color image in color image compression, storage and transmission.

Description

Color image quality evaluation method based on gradient

Technical field

The invention belongs to technical field of image processing, particularly a kind of color image quality evaluation method, can be used for coloured image Compression, storage, process to coloured image in transmitting procedure.

Background technology

Along with the fast development of colour image technique, color digital image by large-scale application in data visualization field. Compared with gray level image, coloured image contains higher level of information, the description objective world that it can be true and lively. But when being digitized coloured image processing, such as at information gathering, conversion process, compression storage, transmission, end During end display grade, it is impossible to the meeting avoided introduces some distortions, and cause the colouring information of image to be distorted or loss etc. shows As, make the quality of coloured image that decline in various degree to occur.Degrade coloured image can be caused to occur between object, color is mixed, The phenomenons such as object edge color bulk obfuscation, it is difficult to obtaining effective information from image, this recognizes objective world band to people Carry out the biggest puzzlement, also bring obstacle for follow-up color picture processing system and analysis.It is thus desirable to it is reasonable in design Color image quality evaluation algorithm.

Existing image quality evaluation algorithm majority is aimed at gray level image and designs, first that coloured image is the most empty from RGB Between transform to gray scale territory after evaluate and test again.Owing to the pixel of gray level image is represented by scalar, and the pixel of coloured image It is to represent with vector, if this kind of algorithm is applied to chromatic distortion image, then have ignored the color component information in image, The evaluation result obtained is poor with subjective perception concordance.Research shows, observes the initial stage of piece image, human eye a people The colouring information that visual information 80% is image received, even if after observing a few minutes, this percentage ratio can also keep About 50%.It thus is seen that color information plays important role during human perception image.Accordingly, it would be desirable to The color impact for image fault is considered in image quality evaluation.Multiformity and human vision pair due to color change The complexity of color-aware, this makes the quantitatively evaluating to color image quality more difficult, for image color distortion feature It is the most necessary that design is applicable to the quality evaluating method of coloured image.

Wang et al. analyzes the local variance variable of image RGB channel, first with the coefficient of its probability distribution structure quaternary number, Further according to the structural information of Quaternion Matrix phenogram picture, finally matrix is decomposed, using the similarity between singular value as Evaluation measurement.This type of method take into account the relation between each chrominance component, but draws being changed by brightness drift and contrast In the cross-color type risen, results contrast is poor.

Xie et al. from the mankind to the perception of picture quality the most relevant with the brightness of image, quantity of information, contrast and noise this Angle is set out, and is extracted the correlated characteristic of three passages, is finally estimated after comprehensive fusion.The feature master that the method is used Want or be made up of the monochrome information of image, not extracting feature, and the distortion range that the method is suitable for from color itself Limited, it is only applicable to evaluation and adds and make an uproar and broad image.As can be seen here, existing most of color image quality evaluation methods exist Relatively large distance is still had with actually used demand in performance indications.

Summary of the invention

Present invention aims in current image quality evaluation methodology the best to color image color distortion evaluation effect Problem, it is proposed that a kind of color image quality evaluation method based on gradient, more effectively, more accurately to coloured image to enter Row quality evaluation, meets the use demand of more multilevel color image.

The technical scheme realizing the object of the invention is: combine phenogram image distortion by the luminance edges of image and chroma edge Degree, the i.e. marginal information according to coloured image are sensitive to distortion variations, extract each channel luminance edge based on gradient characteristics And chroma edge, then merged by reference picture and distorted image Edge difference and difference, coloured image is carried out quality evaluation.

Implementation step includes the following:

(1) utilize the S-CIELAB colored quantum noise that International Commission on Illumination CIE proposes respectively to original image R and distortion map As D carries out color-aware conversion, both images are separately disassembled into luminance channel L and two chrominance channels (a, b), Wherein, L represents that the brightness of color, the color of a passage are to dark green from redness, and the color of b passage is from yellow to blueness；

(2) utilize linear convolution filtering that each passage is carried out gradient calculation, it is thus achieved that original image and the brightness of distorted image Edge and chroma edge:

$g_R^L(x,y)=\sqrt{{(L_R\⊗f_x)}^2(x,y)+{(L_R\⊗f_y)}^2(x,y)}$

$g_D^L(x,y)=\sqrt{{(L_D\⊗f_x)}^2(x,y)+{(L_D\⊗f_y)}^2(x,y)}$

$g_R^a(x,y)=\sqrt{{(a_R\⊗f_x)}^2(x,y)+{(a_R\⊗f_y)}^2(x,y)}$

$g_D^a(x,y)=\sqrt{{(a_D\⊗f_x)}^2(x,y)+{(a_D\⊗f_y)}^2(x,y)}$

$g_R^b(x,y)=\sqrt{{(b_R\⊗f_x)}^2(x,y)+{(b_R\⊗f_y)}^2(x,y)}$

$g_D^b(x,y)=\sqrt{{(b_D\⊗f_x)}^2(x,y)+{(b_D\⊗f_y)}^2(x,y)}$

Wherein, x represents the horizontal direction of wave filter, and y represents wave filter vertical direction, L_RRepresent the luminance channel of original image R, L_DRepresent the luminance channel of distorted image D,Represent the gradient amplitude of original image luminance channel,Represent distortion The gradient amplitude of brightness of image passage；a_RRepresent red to dark green color, a in original image R_DRepresent red in distorted image D Color to dark green color,Represent the gradient amplitude of a chrominance channel in original image and distorted image respectively；b_R Represent in original image R that yellow is to the color of blueness, b_DRepresent that in distorted image D, yellow is to blue color, Represent the gradient amplitude of b chrominance channel in original image and distorted image respectively；

(3) the Edge difference parameter between original image and distorted image is calculated:

(3a) brightness in structural similarity algorithm SSIM is utilized

Similarity evaluation index

Derive luminance edges difference DM between original image and distorted image luminance channel pixel_L, wherein, x, y generation respectively Table original picture block and distorted image block, μ_x, μ_yRepresent original picture block and the average of distorted image block, C respectively₁For constant；

(3b) compare between original image and distorted image pixel red to dark green color distortion, derive a color of the two Degree Edge difference DM_a；

(3c) compare yellow between original image and distorted image pixel and, to blue color distortion, derive the b color of the two Degree Edge difference DM_b；

(4) luminance edges difference and chroma edge difference are carried out linear fusion, obtain final quality evaluation value CGBM:

$CGBM=\underset{i=1}{\overset{N}{\Σ}}\frac{1}{N}({\mathrm{\ω}}_1\·{\mathrm{DM}}_L(i)+{\mathrm{\ω}}_2\·{\mathrm{DM}}_a(i)+{\mathrm{\ω}}_3\·{\mathrm{DM}}_b(i\left)\right)$

Wherein, i=1,2,3N, N are the quantity of all pixels, ω in image₁, ω₂, ω₃Represent luminance channel, a color respectively Degree passage, the weight parameter that distortion-aware is affected by b chrominance channel, CGBM end value in the range of [0,1], result closer to The quality of 1 representative image is the best.

Present invention have the advantage that

1) present invention is compared with other method, has significantly improvement, it is adaptable to more in color image quality evaluation accuracy Eurypalynous cross-color, and achieve the result more consistent with subjective evaluation result.

2) due to the fact that and image is transformed to suitable color space by colored quantum noise, with other method first by coloured image R, G, B color channel is utilized respectively a certain quality evaluation algorithm of gray scale area image and calculates, then R, G, B The result of passage carries out the mode of linear superposition and compares, and describes brightness and the chrominance properties of image more accurately, has more Reasonability.

3) present invention utilizes chroma edge difference that coloured image is carried out quality evaluation, well overcomes coloured silk in traditional method Color image transforms to the problem that gray scale territory carries out losing number of colors information during quality evaluation.Test result indicate that, the present invention couple Calculating in color part difference is effective.

Accompanying drawing explanation

Fig. 1 is the flowchart of the present invention；

Fig. 2 is 8 kinds of different types of cross-color images that the present invention chooses on TID2013 data base；

Fig. 3 is the fitted figure of picture quality Q and the distorted image MOS value utilizing the present invention to try to achieve；

Detailed description of the invention

With reference to Fig. 1, the present invention to realize step as follows:

Step 1. utilizes S-CIELAB colored quantum noise that image is carried out color-aware conversion.

Human visual system can be along with the difference of the environmental factorss such as lighting condition, observed range, scope to the perception of color And change.If the color that two identical be placed under different observation conditions, human visual system is to the two face The perception of color is different.Therefore, before coloured image is processed, need original image image and distortion map The impact for color-aware of the different observation conditions is eliminated as carrying out perception conversion.

This example have chosen S-CIELAB colored quantum noise, enters to transform to LAB2000HL space by original image and distorted image, The result that this space colour difference formula calculates is higher than other color spaces with the color distortion matching degree of human eye perception, can be preferably Analog vision perception characteristic, describe the attributes such as the colourity of color, form and aspect more accurately, be the uniform color space of perception.

Step 2. utilizes linear convolution filtering to extract original image and the luminance edges of distorted image and chroma edge.

For coloured image, traditional method carrying out edge extracting based on brightness of image is insufficient, and this example utilizes face Color sensation know transformation model original image and distorted image are separately disassembled into luminance channel L and two chrominance channels (a, b), Wherein, L represents that the brightness of color, the color of a passage are to dark green from redness, and the color of b passage is from yellow to blueness；

Utilize linear convolution filtering that each passage carries out gradient calculation:

$g_R^L(x,y)=\sqrt{{(L_R\⊗f_x)}^2(x,y)+{(L_R\⊗f_y)}^2(x,y)}$

$g_D^L(x,y)=\sqrt{{(L_D\⊗f_x)}^2(x,y)+{(L_D\⊗f_y)}^2(x,y)}$

$g_R^a(x,y)=\sqrt{{(a_R\⊗f_x)}^2(x,y)+{(a_R\⊗f_y)}^2(x,y)}$

$g_D^a(x,y)=\sqrt{{(a_D\⊗f_x)}^2(x,y)+{(a_D\⊗f_y)}^2(x,y)}$

$g_R^b(x,y)=\sqrt{{(b_R\⊗f_x)}^2(x,y)+{(b_R\⊗f_y)}^2(x,y)}$

$g_D^b(x,y)=\sqrt{{(b_D\⊗f_x)}^2(x,y)+{(b_D\⊗f_y)}^2(x,y)}$

Wherein, x represents the horizontal direction of wave filter, and y represents wave filter vertical direction, L_RRepresent the luminance channel of original image R, L_DRepresent the luminance channel of distorted image D,Represent the gradient amplitude of original image luminance channel,Represent and lose The gradient amplitude of true brightness of image passage；a_RRepresent red to dark green color, a in original image R_DRepresent distorted image D Middle redness to dark green color,Represent the gradient of a chrominance channel in original image and distorted image respectively Amplitude；b_RRepresent in original image R that yellow is to the color of blueness, b_DRepresent that in distorted image D, yellow is to blue color,Represent the gradient amplitude of b chrominance channel in original image and distorted image respectively；

Original image and the luminance edges of distorted image and chroma edge is obtained by each passage being carried out gradient calculation.

Step 3. calculates the Edge difference parameter between original image and distorted image.

3.1) luminance edges difference DM between original image and distorted image luminance channel pixel is calculated_L:

Utilize brightness similarity evaluation index in structural similarity algorithm SSIM

Derive luminance edges difference DM between original image and distorted image luminance channel pixel_L:

${\mathrm{DM}}_L=\frac{1}{1+n_1*{\left(g_R^L\right(x,y)-g_D^L(x,y\left)\right)}^2}$

Wherein, x, y represent original picture block and distorted image block, μ respectively_x, μ_yRepresent original picture block and distortion map respectively As the average of block, C₁For constant, n₁=150 are used for adjusting DM_LSize；

3.2) original image and a chroma edge difference DM of distorted image are calculated_aWith b chroma edge difference DM_b:

In SSIM algorithm, brightness similarity evaluation index is calculating difference in linear brightness space, and the most colored in perception In space, this example takes Guha T et al. at document " Learning sparse models for image quality Assessmen " in the difference measurement mode that proposes calculate a chroma edge difference DM of beginning image and distorted image_aWith b colourity Edge difference DM_b:

${\mathrm{DM}}_a=1-\frac{\left|\right|g_R^a(x,y)-g_D^a(x,y)|{|}_2+m_1}{\left|\right|g_R^a(x,y)|{|}_2+|\left|g_D^a(x,y)\right|{|}_2+m_1}$

${\mathrm{DM}}_b=1-\frac{\left|\right|g_R^b(x,y)-g_D^b(x,y)|{|}_2+m_2}{\left|\right|g_D^b(x,y)|{|}_2+|\left|g_D^b(x,y)\right|{|}_2+m_2}$

Wherein m₁,m₂For the constant that value is the least, with the singularity avoiding denominator to be zero or to cause during close to zero, in the present invention Take m₁=0.5, m₂=0.5.

Step 4. quality evaluation.

Luminance edges difference and chroma edge difference are carried out linear fusion, obtain final quality evaluation value CGBM:

$CGBM=\underset{i=1}{\overset{N}{\Σ}}\frac{1}{N}({\mathrm{\ω}}_1\·{\mathrm{DM}}_L+{\mathrm{\ω}}_2\·{\mathrm{DM}}_a+{\mathrm{\ω}}_3\·{\mathrm{DM}}_b)$

Wherein, i=1,2,3N, N are the quantity of all pixels, ω in image₁, ω₂, ω₃Represent luminance channel respectively, A chrominance channel, the weight parameter that distortion-aware is affected by b chrominance channel, CGBM end value is in the range of [0,1], result Quality closer to 1 representative image is the best.

Advantages of the present invention can be further illustrated by following experiment:

1. evaluation and test condition:

Employ the scholar Nikolay Ponomarenko of Aero-Space university of Ukraine and the TID2013 figure of its seminar foundation As storehouse.This data base is to comprise the public data storehouse that distortion classification is most, amount of images is maximum at present, and it is having been obtained for 7 kinds of new type of distortion are added again on the basis of wide variety of TID2008 data base, and to all type of distortion Grade adds 1 grade, and therefore this data base contains 24 kinds of type of distortion, 5 specified distortion level, altogether 3000 width distortion map Picture and 25 width original images.

Above-mentioned abundant picture material, comprehensive distortion kind and substantial amounts of image number all make TID2013 be more suitable for carrying out matter The checking of amount evaluation algorithms, in addition, selects another major reason of this data base to be that it contains multiple to color-aware Produce the type of distortion of impact.The present invention have chosen in this data base that 8 classes are about the type of distortion of color, as shown in table 1.

8 kinds of chromatic distortions in table 1 TID2013 data base

The image of this 8 class type of distortion shown in table 1 as in figure 2 it is shown, wherein Fig. 2 a, 2b, 2c, 2d, 2e, 2f, 2g, #2 distortion in 2h, I corresponding table 1 respectively adds makes an uproar, #7 distortion quantizing noise, #10 distortion JPEG compression, #16 distortion Average drifting, #17 distortion contrast changes, and #18 distortion color saturation changes, and #22 distortion color quantizing and #23 lose Euchroic difference and original image.

2. emulation experiment

Experiment 1, consistency checking.

In order to test color image quality objective evaluation result and the concordance of subjective quality assessment that the present invention proposes, choose Pearson linearly dependent coefficient PLCC, the accuracy of reflection method for objectively evaluating prediction, PLCC value, closer to 1, represents and calculates Method accuracy is the highest；

By the present invention and existing several up-to-date color image quality evaluation method iCID, CID, QSSIM, S-SSIM and FSIM is directed to chromatic distortion image and is contrast experiment.Result such as table 2.

Table 2 present invention and other algorithm CC value on TID2013 data base

From Table 2, it can be seen that the result that the present invention is in all colours type of distortion is better than CID, QSSIM, and S-SSIM algorithm.Compared with iCID algorithm, the present invention all achieves higher PLCC value in 6 class distortions, especially exists In this kind of distortion of Contrast change with the obvious advantage.In addition, the present invention compensate for FSIM algorithm at Change of Deficiency in color saturation type of distortion.In sum, the present invention is evaluating in accuracy compared with control methods There is significantly improvement, achieve the result consistent with subjective evaluation result.

Experiment 2, soundness verification.

In order to verify the reasonability of the present invention, for four kinds of chromatic distortions, quantizing noise (#7), JPEG compression (#10), Color quantizing (#22) and aberration (#23) contrived experiment of image are analyzed and prove.In the experimental data chosen, bag altogether Including 4 kinds of cross-color types of 25 width original images and its correspondence, wherein each type has 5 distortion levels to increase successively Distorted image.

First, each width distorted image is calculated picture quality, then the institute of the same distortion level of same type of distortion will be belonged to The Q-value having distorted image corresponding converges and takes its average.Meanwhile, all of the same distortion level of same type of distortion will be belonged to The MOS value of image takes average, and result is as shown in Figure 3.

As seen from Figure 3, objective quality assessment result Q increases, the prediction of the present invention along with the reduction of image fault degree Trend and image fault degree have good concordance, can effectively detect the change of picture quality, have reasonability.

Experiment 3, color validation verification.

In order to verify color part effect during the overall evaluation in the present invention, by Color Channel a and b in the present invention Weight parameter ω₂And ω₃Being set to 0, the most only utilize the luminance part in the present invention to carry out quality evaluation, experimental result is designated as CGBM*.Obtained experimental data is compared with experimental data CGBM utilizing the present invention to obtain.Result such as table 3.

The MOS average of the different distortion level collection of table 3 and evaluation result (CGBM) average

From table 3 it can be seen that after the color part in the present invention is removed, its experimental result is at type of distortion #2, #10, #22 Almost keep consistent with on #23 with result of the present invention, and the PLCC value in #16, #17 and #18 these three distortion significantly under Fall, the result of especially #17 is only 0.3, it can be seen that, in algorithm, the calculating for color part difference is effective.

Claims (4)

1. a color image quality evaluation method based on gradient, including:

(1) utilize the S-CIELAB colored quantum noise that International Commission on Illumination CIE proposes respectively to original image R and distorted image D carries out color-aware conversion, both images are separately disassembled into luminance channel L and two chrominance channels (a, b), Wherein, L represents that the brightness of color, the color of a passage are to dark green from redness, and the color of b passage is from yellow to blueness；

(2) utilize linear convolution filtering that each passage is carried out gradient calculation, it is thus achieved that the brightness limit of original image and distorted image Edge and chroma edge:

$g_R^L(x,y)=\sqrt{{(L_R\⊗f_x)}^2(x,y)+{(L_R\⊗f_y)}^2(x,y)}$

$g_D^L(x,y)=\sqrt{{(L_D\⊗f_x)}^2(x,y)+{(L_D\⊗f_y)}^2(x,y)}$

$g_R^a(x,y)=\sqrt{{(a_R\⊗f_x)}^2(x,y)+{(a_R\⊗f_y)}^2(x,y)}$

$g_D^a(x,y)=\sqrt{{(a_D\⊗f_x)}^2(x,y)+{(a_D\⊗f_y)}^2(x,y)}$

$g_R^b(x,y)=\sqrt{{(b_R\⊗f_x)}^2(x,y)+{(b_R\⊗f_y)}^2(x,y)}$

$g_D^b(x,y)=\sqrt{{(b_D\⊗f_x)}^2(x,y)+{(b_D\⊗f_y)}^2(x,y)}$

Wherein, x represents the horizontal direction of wave filter, and y represents wave filter vertical direction, L_RRepresent that the brightness of original image R is led to Road, L_DRepresent the luminance channel of distorted image D,Represent the gradient amplitude of original image luminance channel,Represent The gradient amplitude of distorted image luminance channel；a_RRepresent red to dark green color, a in original image R_DRepresent distorted image D Middle redness to dark green color,Represent the gradient width of a chrominance channel in original image and distorted image respectively Degree；b_RRepresent in original image R that yellow is to the color of blueness, b_DRepresent that in distorted image D, yellow is to blue color,Represent the gradient amplitude of b chrominance channel in original image and distorted image respectively；

(3) the Edge difference parameter between original image and distorted image is calculated:

(3a) brightness similarity evaluation index in structural similarity algorithm SSIM is utilized Derive luminance edges difference DM between original image and distorted image luminance channel pixel_L, wherein, x, y represent respectively Original picture block and distorted image block, μ_x, μ_yRepresent original picture block and the average of distorted image block, C respectively₁For constant；

(3b) compare between original image and distorted image pixel red to dark green color distortion, derive a colourity of the two Edge difference DM_a；

(3c) compare yellow between original image and distorted image pixel and, to blue color distortion, derive the b colourity of the two Edge difference DM_b；

(4) luminance edges difference and chroma edge difference are carried out linear fusion, obtain final quality evaluation value CGBM:

$CGBM=\underset{i=1}{\overset{N}{\Σ}}\frac{1}{N}({\mathrm{\ω}}_1\·{\mathrm{DM}}_L(i)+{\mathrm{\ω}}_2\·{\mathrm{DM}}_a(i)+{\mathrm{\ω}}_3\·{\mathrm{DM}}_b(i\left)\right)$

Wherein, i=1,2,3 N, N are the quantity of all pixels, ω in image₁, ω₂, ω₃Represent luminance channel, a respectively Chrominance channel, the weight parameter that distortion-aware is affected by b chrominance channel, CGBM end value in the range of [0,1], result closer to The quality of 1 representative image is the best.

Method the most according to claim 1, the original image wherein derived in step (3a) and distorted image luminance channel Luminance edges difference DM between pixel_L, it is expressed as follows:

${\mathrm{DM}}_L=\frac{1}{1+n_1*{\left(g_R^L\right(x,y)-g_D^L(x,y\left)\right)}^2}$

Wherein n₁=150, it is used for adjusting DM_LSize.

Method the most according to claim 1, the original image wherein derived in step (3b) and distorted image pixel it Between a chroma edge difference DM_a, it is expressed as follows:

${\mathrm{DM}}_a=1-\frac{\left|\right|g_R^a(x,y)-g_D^a(x,y)|{|}_2+m_1}{\left|\right|g_R^a(x,y)|{|}_2+|\left|g_D^a(x,y)\right|{|}_2+m_1}$

Wherein m₁=0.5, with the singularity avoiding denominator to be zero or to cause during close to zero.

Method the most according to claim 1, the original image wherein derived in step (3c) and distorted image pixel it Between b chroma edge difference DM_b, it is expressed as follows:

${\mathrm{DM}}_b=1-\frac{\left|\right|g_R^b(x,y)-g_D^b(x,y)|{|}_2+m_2}{\left|\right|g_D^b(x,y)|{|}_2+|\left|g_D^b(x,y)\right|{|}_2+m_2}$

Wherein m₂=0.5, with the singularity avoiding denominator to be zero or to cause during close to zero.