CN104361593A - Color image quality evaluation method based on HVSs and quaternions - Google Patents

Abstract

The invention discloses a color image quality evaluation method based on HVSs and quaternions and belongs to the technical field of image processing and computer vision. The method includes the steps that firstly, mathematic evaluation models of an original reference image and a distorted image to be evaluated are established through analysis of human vision features, wherein the mathematic evaluation models comprise spatial location functions Q<L>, local variances Q<V>, texture edge complexity functions Q<TE> and color functions Q<C> of the images; secondly, quaternion arrays of the original reference image and the distorted image to be evaluated are established and singular value decomposition is conducted on the quaternion arrays, so that singular value feature vectors of the images can be acquired; thirdly, the image distortion degree is measured through the Euclidean distance of the singular value feature vectors of the original reference image and the distorted image to be evaluated. According to the method, the human vision features and the quaternions are combined, brightness and chromaticity information of the images is extracted, the spatial location functions, the texture edge complexity functions and the local variances are established through the human vision features, and an evaluation result accords with a result generated in the mode that the images are sensed by the human eyes.

Description

A kind of color image quality evaluation method based on HVS and hypercomplex number

Technical field

The present invention relates to Image processing and compute machine vision technique field, more particularly, relate to a kind of characteristic of human visual system that utilizes and build the mathematical model consistent with eye-observation image, combine with hypercomplex number svd, carry out the method for color image quality evaluation.

Background technology

Picture quality is one of parameter important in image procossing and computer vision field, along with the development of computer science and technology, the requirements of aspect to picture quality such as printing, ceramic tile, image, image retrieval are more and more higher, but the image fault that can produce in the process such as collection, process, compression, transmission, display of image in various degree and image deterioration problem.

The mankind, as the final recipient of image, make its subjective quality assessment to image (Difference Mean OpinionScore, DMOS) be considered to the most reliable.Subjective quality assessment allows observer according to oneself subjective perception experience or the good evaluation criterion of some prior Uniform provisions, the visual perception treating evaluation objective image is made quality assessment and gives a mark, and then the mark of all observers is weighted on average, the result of gained is the subjective quality scores of image.But subjective picture quality evaluation is wasted time and energy, the impact by observer, image type and surrounding environment is comparatively large, and real-time is more weak.Thus people are devoted to study the Objective image quality evaluation method that correctly can reflect the perception of people's subjective vision timely and effectively always.Objective image quality evaluation utilizes algorithm, mathematical model etc. to carry out in time picture quality, feeds back fast to obtain the evaluation result consistent with the subjective feeling of people.The method is varied, and due to the difference of point of penetration, basic thought, sorting technique is also different.According to the reference to original image, method for evaluating objective quality is divided into full reference type, partial reference type and without reference type 3 kinds.Full reference type is applicable to the Performance comparision of encoder design and different coding device, partial reference type and be applicable to band-limited multimedia application without reference type.Because full reference type can utilize the full detail of original image, it more meets human subject to the evaluation result of image and evaluates.

Y-PSNR (the Peak Signal Noise Ratio proposed in " Image QualityAssessment Based on Gradient Similarity " that the people such as Liu A delivered on " IEEE Transactions on image processing " in 2012, PSNR) and square error (Mean Square Error, MSE) be the most classical full reference type Objective image quality evaluation method.PSNR reflects the fidelity (Fidelity) of image to be evaluated, and MSE reflects the otherness (Diversity) of image to be evaluated and original image.The theory of above-mentioned two kinds of methods is simple and clear, easy understand, calculate also very convenient, but they only considered the comparison of each pixel of image, not considering the structural relation etc. that may exist between each pixel of image, there is deviation in that truly sees with human eye.

The people such as Z Wang propose the more original undistorted image of SSIM algorithm synthesis and image to be evaluated difference between the information that brightness, contrast are different with structural similarity three class in " the Image qualityassessment from error measurement to structural similarity " within 2004, to deliver on " IEEE Transactions on Image Processing ", consider the structural relation between pixel, but details is held bad under there is serious ambiguity, and index parameters determines the problems such as difficulty.

The similarity deviation algorithm GMSD based on gradient magnitude proposed in " Gradientmagnitude similarity deviation:A highly efficient perceptual image quality index " that the people such as W.Xue delivered on " IEEE Transactions on Image Processing " in 2013 considers that gradient is extremely sensitive to image fault, but must be transformed on gray scale territory the process of coloured image.Above method must be converted into gray level image for the evaluation of coloured image, and evaluation result and the actual situation about seeing of human eye exist deviation.

Through retrieval, Chinese Patent Application No. 200610027433.1, the applying date is on June 8th, 2006, and invention and created name is: a kind of image quality measure method based on supercomplex svd; This application case utilizes supercomplex (hypercomplex number) directly to coloured image modeling, coloured image self-energy feature is extracted by supercomplex svd, utilize the distance between original image and distorted image singular value to construct distortion map matrix, and use this distortion map matrix to assess color image quality.Chinese Patent Application No. 201210438606.4, the applying date is on November 6th, 2012, invention and created name is: color image quality evaluation algorithm, this application case is respectively by the colourity of image, brightness and saturation degree are as the imaginary part of hypercomplex number, the Quaternion Matrix of structure reference picture and image to be evaluated, and respectively svd is carried out to them, obtain singular value feature vector, finally apply the degree of association between the singular value feature vector of grey relational grade computing reference image and the singular value feature vector of each image to be evaluated, the degree of association is larger, show that the quality of image to be evaluated is better.But still there is relatively large deviation in the evaluation result that above-mentioned application case obtains and the actual situation about seeing of human eye, the evaluation method of color image quality still needs further optimization.

Summary of the invention

1. invent the technical matters that will solve

The present invention overcomes traditional evaluation method when building evaluation model, coloured image need be converted into gray level image to process, cause evaluation result and the larger problem of the actual situation deviation seen of human eye, provide a kind of color image quality evaluation method based on HVS and hypercomplex number; The present invention proposes human-eye visual characteristic and hypercomplex number to combine, extract brightness and the chrominance information of image, human-eye visual characteristic is utilized to construct function of spatial position, texture fringe complexity function and local variance, the three-channel method of R, G, B is isolated for improving tradition, utilize hypercomplex number svd to extract the energy feature of image, evaluation result is more conformed to the effect of Human Perception image.

2. technical scheme

For achieving the above object, technical scheme provided by the invention is:

A kind of color image quality evaluation method based on HVS and hypercomplex number of the present invention, the steps include:

Step one, build the mathematics appraisal of original reference image and distorted image to be evaluated by analyzing human-eye visual characteristic, described mathematics appraisal comprises the function of spatial position Q of image _l, local variance Q _v, texture fringe complexity function Q _tEwith color function Q _c;

Step 2, by Q _l, Q _v, Q _tEas the imaginary part of hypercomplex number, Q _cas the real part of hypercomplex number, construct the Quaternion Matrix of original reference image and distorted image to be evaluated respectively, and the singular value feature vector that svd obtains image is carried out to Quaternion Matrix;

Step 3, utilize the euclidean distance metric image fault degree of the singular value feature vector of original reference image and distorted image to be evaluated.

Further, step one builds the detailed process of mathematics appraisal and is:

(1) the RGB tristimulus values of original reference image and distorted image to be evaluated is obtained;

(2) extract the spatial positional information of original reference image and distorted image to be evaluated, build function of spatial position Q _lwith texture fringe complexity function Q _tE;

(3) original reference image and distorted image to be evaluated are converted to YUV color space by rgb space, extract image luminance information and build local variance Q _v, extract brightness of image and chrominance information structure color function Q _c.

Further, step one utilizes the middle concave characteristic of human visual system to build function of spatial position Q _l, described function of spatial position

$Q_L(i,j)=\frac{e_c}{e_c+e_L}$

In formula, e _lfor human eye Visual Observations Observations pixel (i, j) to picture centre pixel (M/2, N/2) distance with business; e _cfor constant.

Further, step one utilizes the shielding effect of human visual system to build texture fringe complexity function Q _tE, described texture fringe complexity function

Q _TE＝Q _T×Q _E

In formula, Q _tfor the Texture complication function of pixel (i, j), Q _zfor the fringe complexity function of pixel (i, j).

Further, step one utilizes the hyperchannel characteristic of human visual system to build local variance Q _v, described local variance

$Q_V\left(I_{i,j}\right)=\frac{1}{L}\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{\&eta;}}_p-\stackrel{\&OverBar;}{I_{i,j}})}^2$

Wherein, the piecemeal carrying out non-overlapping copies according to the luminance component of image obtains I _i,j, L is image block I _i,jin the pixel η that comprises _pnumber, $\stackrel{\&OverBar;}{I_{i,j}}=\frac{1}{L}\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&eta;}}_p.$

Further, described color function

Q _C＝αQ _L+βQ _U

In formula, Q _lfor the monochrome information of image, Q _ufor the chrominance information of image, α, β are respectively the proportion shared by brightness and colourity.

Further, the Euclidean distance described in step 3

$D=\sqrt{\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}({\mathrm{\&lambda;}}_i-\widehat{{\mathrm{\&lambda;}}_i})}$

In formula, λ _ifor the singular value feature vector of original reference image, for the singular value feature vector of distorted image to be evaluated, K is the minimum value of two singular value feature vector eigenwert numbers, i.e. the minimum value of two Quaternion Matrix orders:

$K=\min (\mathrm{rank}\left(I\right),\mathrm{rank}\left(\hat{I}\right)).$

3. beneficial effect

Adopt technical scheme provided by the invention, compared with existing known technology, there is following remarkable result:

(1) a kind of color image quality evaluation method based on HVS and hypercomplex number of the present invention, it constructs the function of spatial position Q of original reference image and distorted image to be evaluated by analyzing human-eye visual characteristic _l, local variance Q _v, texture fringe complexity function Q _tEwith color function Q _cand by hypercomplex number by above-mentioned four parts of images information integration, the energy feature of image is obtained through svd, improve tradition and isolate the three-channel method of R, G, B, ensure that the integrality of colouring information well, the image information extracted comprises the overall situation and local message, makes the full detail of the token image that evaluation result can be more complete;

(2) a kind of color image quality evaluation method based on HVS and hypercomplex number of the present invention, human-eye visual characteristic and hypercomplex number are combined, evaluation result more conforms to the effect of Human Perception image, is better than traditional SSIM and other several typical image quality evaluation algorithms.

Accompanying drawing explanation

Fig. 1 is the algorithm flow chart of a kind of color image quality evaluation method based on HVS and hypercomplex number of the present invention;

Fig. 2 is the middle concave characteristic isoboles of human visual system in the present invention;

Fig. 3 is the fitting result figure of quality evaluating method of the present invention, classic method and image subjective quality; Wherein, the nonlinear fitting curve map that (a) in Fig. 3 is worth for PSNR and DMOS, the nonlinear fitting curve map that (b) in Fig. 3 is worth for SSIM and DMOS, the nonlinear fitting curve map that (c) in Fig. 3 is worth for MS-SSIM and DMOS, the nonlinear fitting curve map that (d) in Fig. 3 is worth for SVD and DMOS, the nonlinear fitting curve map that (e) in Fig. 3 is worth for GMSD and DMOS, the nonlinear fitting curve map that (f) in Fig. 3 is quality evaluating method of the present invention and DMOS value;

(a) in Fig. 4 ~ (e) is five groups of different HVS-QSVD of type of distortion image, nonlinear fitting curve comparison figure of GMSD, SSIM and DMOS value.

Embodiment

For understanding content of the present invention further, the present invention is described in detail in conjunction with the accompanying drawings and embodiments.

Embodiment 1

Composition graphs 1, the present embodiment for traditional evaluation method build evaluation model time, coloured image need be converted into gray level image to process, cause the problems such as evaluation result and Human Perception are not inconsistent, provide a kind of color image quality evaluation method based on HVS and hypercomplex number.The present embodiment, by human-eye visual characteristic and hypercomplex number being combined, obtains the image energy feature meeting Human Perception.Experiment shows, the evaluation of the present embodiment to coloured image is better than additive method, and evaluation result is more consistent with Human Perception image.Described in detail below in conjunction with the image quality evaluating method of experimental result to the present embodiment:

Step one, build the mathematics appraisal of original reference image and distorted image to be evaluated by analyzing human-eye visual characteristic:

How the human-eye visual characteristic that the basis understood human visual system's physiological structure proposes, observe external environment with human eye, image be closely bound up.The final recipient of image because behaving, thus evaluation result will with human eye actual see conform to, the present embodiment constructs corresponding mathematical model by analysis human-eye visual characteristic.

Human eye convex lens variable to focal length are similar, but its affect by the labyrinth of human brain, different again with general convex lens.Usually, human-eye visual characteristic comprises middle concave characteristic, vision hyperchannel characteristic, visual properties, CSF and shielding effect.The present embodiment is from middle concave characteristic, vision hyperchannel characteristic, shielding effect structure mathematics appraisal.

(1) function of spatial position

When the middle concave characteristic of human visual system refers to that image occurs, first the center position information of image can be arrived by eye-observation, especially the change in location information at the immediate vicinity texture edge of image, because human eye is than being easier to the marginal information experiencing image.

First human eye can see the centre of image, and then to surrounding diffusion, and surrounding distance center distance point human eye equally should be put on an equal footing.As shown in Figure 2, suppose that center of circle O is the center of image, the point on circle is equal to the distance in the center of circle, and the probability that A, B, C, D 4 is arrived by eye-observation is identical, and E and F is also identical.

The present embodiment is according to document (CHEN T, WU H R.Space variant median filters for the restoration ofimpulse noise corrupted images [J] .IEEE Transactions on Circuits and Systems II:Analog andDigital Signal Processing, 2001,48 (8): 784-789.) mention, according to the central fovea characteristic of human visual system, be how to affect eye-observation image by definite formula representation space resolution.Detailed process is, first the RGB tristimulus values of original reference image and distorted image to be evaluated is obtained, extract the spatial positional information of original reference image and distorted image to be evaluated again, build the function of spatial position Q of original reference image and distorted image to be evaluated respectively _l:

$Q_L(i,j)=\frac{e_c}{e_c+e_L}---\left(1\right)$

In formula, e _lfor the pixel (i, j) of human eye Visual Observations Observations to the distance of picture centre pixel (M/2, N/2) divided by the distance of this width image first pixel (0,0) first pixel to picture centre pixel e _cthe constant determined according to test result, the present embodiment setting e after tested _cbe 0.6.

(2) texture fringe complexity function

The shielding effect of human visual system refers to that the phenomenon that originally can be noted has been left in the basket due to the existence of other phenomenons.In zones of different, the shielding effect of human visual system can be reflected by respective weight proportion, and this is more consistent with the characteristic of eye-observation image.

The present embodiment extracts texture feature information and the edge feature information of image, obtains the texture fringe complexity function Q of image _tE, Q _tElarger expression texture is simpler, and human eye is all the more paid close attention to, and it is larger on the impact of picture quality quality; Contrary, QTE less expression texture is more complicated, and human eye is more easily ignored.Concrete computation process is as follows:

First compute gradient direction:

Wherein, θ (i, j) represents the gradient direction of pixel (i, j), with represent level, the vertical gradient value of pixel (i, j).Utilize Sobel edge detection operator, calculate corresponding picture edge characteristic information, by marginal information normalization, be denoted as E (i, j).[0,360 °) gradient direction is divided into following several region by scope, is shown below:

θ(i,j)∈{0°,180°,45°,225°,90°,270°,135°,315°} (3)

Wherein 0 ° and 180 °, 45 ° and 225 °, 90 ° and 270 °, 135 ° and 315 °, respectively about origin symmetry, namely have the region, direction that 4 different.

Calculate Texture complication:

If a ₁for direction kind number, i.e. the kind number of θ (i, j), a ₂for marginal point number, namely the pixel number of E (i, j)=1, calculates their value respectively.Work as a ₂time less than the threshold value of setting, a ₂=0, otherwise a ₂=1, this threshold value is set to 40 through experiment test.Therefore, we are by the Texture complication function Q of pixel (i, j) a certain in image _tbe defined as following form:

Fringe complexity:

First three vectorial P=(-1,0,2,0,1) are defined, L=(Isosorbide-5-Nitrae, 6,4,1), E=(-1 ,-2,0,2,1).Wherein, P represents " point " Feature Descriptor, L represents " line " Feature Descriptor, E represents " limit " Feature Descriptor, utilizes these 3 operators can obtain 6 mask: L ^t× E, L ^t× P, E ^t× L, E ^t× P, P ^t× L, P ^t× E.If the response at these 6 masks certain pixel (i, j) place is in the drawings respectively f _i,j(L ^t× E), f _i,j(L ^t× P), f _i,j(E ^t× L), f _i,j(E ^t× P), f _i,j(P ^t× L), f _i,j(P ^t× E), then the fringe complexity of this pixel (i, j) is:

$Q_E=\sqrt{{\left(f_{i,j}(L^T\&times;E)\right)}^2+{\left(f_{i,j}(L^T\&times;P)\right)}^2+{\left(f_{i,j}(E^T\&times;L)\right)}^2+{\left(f_{i,j}(E^T\&times;P)\right)}^2+{\left(f_{i,j}(P^T\&times;L)\right)}^2+{\left(f_{i,j}(P^T\&times;E)\right)}^2}---\left(5\right)$

The texture fringe complexity function of pixel (i, j):

Q _TE＝Q _T×Q _E(6)

The larger shielding effect of result of calculation is more weak, and shielding effect is more weak, and texture is about simple, and human eye is just seen clearly than being easier to, namely stronger to the visual impact of human eye.

(3) local variance

The hyperchannel characteristic of human visual system refers to that eye-observation image carries out in different passages, can only tell general outline, and resolution height just can tell detailed information when resolution is low.The detailed information of image can be represented by the local variance of image, so the local variance of image is used as description, analyzes a kind of means of image content information, some important feature information of image also can be distributed by the local variance of image and summarize simultaneously.

Use Q _vrepresent the variance of the regional area of image I centered by pixel (i, j), i.e. local variance.First the present embodiment needs original reference image and distorted image to be evaluated to be converted to YUV color space by rgb space, utilizes Y (sign brightness) to classify and calculates local variance.Adopt moving window the Y-component of image to be carried out to the piecemeal of non-overlapping copies, obtain the variance of each piecemeal, this is the local variance of image.For each image block I _i,j, comprise L pixel, use η _prepresent its each pixel inner, then local variance can be expressed as:

$Q_V\left(I_{i,j}\right)=\frac{1}{L}\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{\&eta;}}_p-\stackrel{\&OverBar;}{I_{i,j}})}^2---\left(7\right)$

Wherein, for piecemeal I _i,javerage.

Because the size of each sub-block, mode can have influence on the structure of image, for being included in I _i,jpixel η in piecemeal _padopt document (Z Wang, Z Bovik, et al.Image quality assessment from error measurement to structuralsimilarity [J] .IEEE Transactions on Image Processing.2004,13 (4): 600-612.) the Gauss's weighted method computation of mean values mentioned and variance, as follows:

Piecemeal average: $\stackrel{\&OverBar;}{I_{i,j}}=\frac{\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{X}_p{\mathrm{\&eta;}}_p}{\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&eta;}}_p}---\left(8\right)$

Piecemeal local variance: $Q_V\left(I_{i,j}\right)=\frac{\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{X}_p{({\mathrm{\&eta;}}_p-\stackrel{\&OverBar;}{I_{i,j}})}^2}{\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{X}_p}---\left(9\right).$

X in formula _pfor pixel η _pnumber.

(4) colouring information

Hue, saturation, intensity is three attributes of color, also referred to as color three elements.They are inherent characteristics of color, and different.Tone and saturation degree can be embodied by colourity.The unique features of gray level image is brightness, and coloured image also has chromaticity.

Brightness is a physical quantity, is the impression of people to the intensity of light, reflects the power of luminophor (refractive body) surface light emitting (reflective).Tone refers to the general inclination of a width picture-in-picture complexion coloured silk, is large color effect.Saturation degree, also referred to as the purity of color, refers to the bright-coloured degree of color, represents the ratio of contained color composition in color.The saturation degree of color increases along with the increase of color ratio, contacts directly with the surface structure of light radiation situation and subject.Because tone and saturation degree can be unified to represent by colourity, so the present embodiment utilizes brightness and colourity to represent the essential attribute of color.

Human visual system is to the susceptibility of brightness higher than the susceptibility to colourity, and the present embodiment method of weighting represents the chromatic information of image, and namely to different coloured images, brightness is different with the weight proportion shared by colourity, and concrete calculated relationship is as follows:

Q _C＝αQ _L+βQ _U(10)

Wherein, Q _lfor the monochrome information of image, Q _ufor the chrominance information of image, α, β are respectively the proportion shared by brightness and colourity, and through experimental result test, α, β get 1.063 and 0.937 the best respectively.

The Quaternion Matrix of step 2, respectively structure original reference image and distorted image to be evaluated, and the singular value feature vector that svd obtains image is carried out to Quaternion Matrix:

(a) hypercomplex number

1843, British mathematician Hamiltonian (W.R.Hamilton) created hypercomplex number.A hypercomplex number comprise 4 parts, 1 real part adds 3 imaginary parts, and its citation form is:

$\stackrel{\&RightArrow;}{q}=q_r+q_i\stackrel{\&RightArrow;}{i}+q_j\stackrel{\&RightArrow;}{j}+q_k\stackrel{\&RightArrow;}{k}$

Wherein, q _r, q _i, q _j, q _kbe four real numbers, primitive meet:

${\stackrel{\&RightArrow;}{i}}^2={\stackrel{\&RightArrow;}{j}}^2={\stackrel{\&RightArrow;}{k}}^2=\stackrel{\&RightArrow;}{i}\stackrel{\&RightArrow;}{j}\stackrel{\&RightArrow;}{k}=-1,$

$\stackrel{\&RightArrow;}{i}\stackrel{\&RightArrow;}{j}=-\stackrel{\&RightArrow;}{j}\stackrel{\&RightArrow;}{i}=\stackrel{\&RightArrow;}{k},\stackrel{\&RightArrow;}{j}\stackrel{\&RightArrow;}{k}=-\stackrel{\&RightArrow;}{k}\stackrel{\&RightArrow;}{j}=\stackrel{\&RightArrow;}{i},\stackrel{\&RightArrow;}{k}\stackrel{\&RightArrow;}{i}=-\stackrel{\&RightArrow;}{i}\stackrel{\&RightArrow;}{k}=\stackrel{\&RightArrow;}{j}$

Quaternion Matrix real number field can be analyzed to following form:

$Q=\left(\begin{array}{cccc}{Q}_r& Q_i& Q_j& Q_k\\ -Q_i& Q_r& -Q_k& Q_j\\ -Q_j& Q_k& Q_r& -Q_i\\ -Q_k& -Q_j& Q_i& Q_r\end{array}\right)$

Singular Value of Quaternion Matrices decomposition theorem can be expressed as: for any Quaternion Matrix Q _{e (q)}=U _(q)Λ ' V _(q), if rank (A)=r, then there is unitary quaternion matrix U in x _(q)and V _(q), make

Q _(q)＝U _(q)ΛV _(q)λ

Wherein,

$U_{\left(q\right)}{U}_{\left(q\right)}^H=V_{\left(q\right)}{V}_{\left(q\right)}^H=I$

$\mathrm{\&Lambda;}={\left(\begin{array}{cccccc}{\mathrm{\&lambda;}}_1& & & & & \\ & ...& & & & \\ & & {\mathrm{\&lambda;}}_r& & & \\ & & & 0& & \\ & & & & ...& \\ & & & & & 0\end{array}\right)}_{n\&times;n}={\left(\begin{array}{cccc}{\mathrm{\&Lambda;}}_r& & & \\ & 0& & \\ & & ...& \\ & & & 0\end{array}\right)}_{n\&times;n}$

And meet λ _i∈ R, | λ ₁|>=| λ ₂|>=...>=| λ _r| > 0, λ _ifor non-zero singular value.

B () hypercomplex number represents

Four characteristic informations of the coloured image of above-mentioned analysis gained are integrated into a hypercomplex number form by the present embodiment, as follows:

Q＝Q _C+Q _Li+Q _TEj+Q _Vk (11)

Wherein, Q _cfor the colouring information of image, Q _lfor the spatial positional information of image, Q _tEfor the texture marginal information of image, Q _vfor the local variance of image.

So, the coloured image of one width M × N can regard a Quaternion Matrix as, the singular value feature vector of Quaternion Matrix characterizes the energy feature of hypercomplex number, so also can be used for representing the energy feature of corresponding coloured image by the Quaternion Matrix of coloured image gained.

Because a Quaternion Matrix q=q _r+ q _ii+q _jj+q _kk, can represent by its real matrix, so Q is converted into its corresponding wave function by the present embodiment carry out svd (SVD) to it.Each Quaternion Matrix can obtain a singular value feature vector by svd, and each element of proper vector be greater than 0 real number.What deserves to be explained is, in view of theoretical research Quaternion Matrix being carried out to svd is more ripe, the present embodiment is considered from the angle reducing length, repeats no more herein.

Step 3, utilize the euclidean distance metric image fault degree of the singular value feature vector of original reference image and distorted image to be evaluated:

The present embodiment utilizes the Euclidean distance of the singular value feature vector of original reference image and distorted image to be evaluated (EuclideanDistance) to measure corresponding image fault, namely

$D=\sqrt{\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}({\mathrm{\&lambda;}}_i-\widehat{{\mathrm{\&lambda;}}_i})}---\left(12\right)$

Wherein, λ _iwith for the singular value feature vector that the original reference image and distorted image to be evaluated that calculate acquisition are corresponding, K is the minimum value of two singular value feature vector eigenwert numbers, i.e. the minimum value of two Quaternion Matrix orders:

$K=\min (\mathrm{rank}\left(I\right),\mathrm{rank}\left(\hat{I}\right))$

The color image quality evaluation method based on HVS and hypercomplex number of the present embodiment, human-eye visual characteristic and hypercomplex number are combined, evaluation result more conforms to the effect of Human Perception image, improve tradition and isolate the three-channel method of R, G, B, ensure that the integrality of colouring information well, the image information extracted comprises the overall situation and local message, makes the full detail of the token image that evaluation result can be more complete.Its evaluation result is better than traditional SSIM and other several typical image quality evaluation algorithms, the experimental result from two aspects to the present embodiment is analyzed below:

1) nonlinear fitting of quality evaluating method of the present invention, PSNR, SSIM, MS-SSIM, Y-SVD, GMSD and DMOS value; 2) quality evaluating method of the present invention compares with PSNR, SSIM, MS-SSIM, Y-SVD, GMSD performance evaluation.

Quality assessment picture is Texas ,Usa (TEXAS) state university's Jane Austen branch school image and video engineering experiment room (Library for Image and Video Engineering, LIVE) image quality evaluation database D atabase Release 2 image library provided, 982 width, have JPEG2000, JPEG, white Gaussian noise, Gaussian Blur, Fast Fading Rayleigh channel distortion five kinds of type of distortion altogether.When carrying out method comparison, the difference in each algorithm dimension and unit can be produced.Therefore, non-linear regression is carried out in the Objective image quality scoring obtained by algorithm to be evaluated.Logistic function is utilized to carry out non-linear regression as nonlinear mapping function to the original scoring of Objective image quality that the algorithm to be evaluated that the present invention proposes draws:

$\mathrm{Qq}\left(x\right)=\frac{{\mathrm{\&alpha;}}_1-{\mathrm{\&alpha;}}_2}{1+\exp [-(x-{\mathrm{\&alpha;}}_3)/{\mathrm{\&alpha;}}_4]}+{\mathrm{\&alpha;}}_2---\left(13\right)$

Wherein, x is the original quality scoring that algorithm to be evaluated that the present invention proposes draws, α ₁, α ₂, α ₃, α ₄for the parameter of self-adaptative adjustment in non-linear regression process.The index of quantitative test evaluation result is generally acknowledged degree and quotes the more MAE/RMSE/CC/SROCC/OR of number of times.

1) mean absolute error (the Mean Absolute Error after subjective and objective non-linear regression between mark, MAE), reflect the average error level of evaluating objective quality result and subjective evaluation result, littlely show that image quality evaluation result accuracy is higher, defined formula is as follows:

$\mathrm{MAE}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}|x_i-y_i|---\left(14\right)$

2) root-mean-square error (Root Mean Square Error, RMSE) after subjective and objective non-linear regression between mark, reflects the accuracy of objective evaluation result, littlely shows that image quality evaluation result accuracy is higher, and defined formula is as follows:

$\mathrm{RMSE}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(x_i-y_i)}^2}---\left(15\right)$

3) Pearson linearly dependent coefficient (the Correlation Coe cient after subjective and objective non-linear regression between mark, CC), the consistance of reflection objective evaluation result and accuracy, span is [-1,1], the absolute value of result is more close to 1, and the correlativity of subjective evaluation method is better, and defined formula is as follows:

$\mathrm{CC}=\frac{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}(x_i-\stackrel{\&OverBar;}{x})(y_i-\stackrel{\&OverBar;}{y})}{\sqrt{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(x_i-\stackrel{\&OverBar;}{x})}^2\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(y_i-\stackrel{\&OverBar;}{y})}^2}}---\left(16\right)$

4) Spearman coefficient of rank correlation (the Spearman Rank OrderCorrelation Coe cient after subjective and objective non-linear regression between mark, SROCC), apply nonparametric discrimination method method more widely, reflect the monotonicity of evaluating objective quality result and subjective evaluation result, span is in [-1,1], and the absolute value of result is more close to 1, the consistance of subjective evaluation method is better, and defined formula is as follows:

$\mathrm{SROCC}=\frac{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}(u_i-\stackrel{\&OverBar;}{u})(v_i-\stackrel{\&OverBar;}{v})}{\sqrt{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(u_i-\stackrel{\&OverBar;}{u})}^2}\&CenterDot;\sqrt{\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(v_i-\stackrel{\&OverBar;}{v})}^2}}---\left(17\right)$

5) after subjective and objective non-linear regression between mark from going out rate (Outlier Rate, OR), stability, the predictability of reflection objective evaluation model, span is [0,1], numerical value is less, and the consistance of subjective evaluation is better, the predictability of evaluation model is also better, and defined formula is as follows:

$\mathrm{OR}=\frac{N_{\mathrm{out}}}{N}---\left(18\right)$

Wherein N is the sum of image data base, namely 982, x _i, y _irepresent the subjective evaluation value of the i-th width image after non-linear regression respectively, u _i, v _irepresent the rank of the subjective evaluation value of the i-th width image in all evaluations of estimate of whole image data base respectively, N _outrepresent that objective evaluation value is greater than the image number of subjective assessment value standard deviation twice.

Fig. 3 respectively illustrates the scatter diagram of various algorithm and subjective assessment value DMOS, in figure, each point represents piece image, the horizontal ordinate of point is that algorithm is marked to the evaluating objective quality of this image, and ordinate is the subjective assessment DMOS value of this image, and solid line represents matched curve.Loose point is more tightly distributed near matched curve, and represent that the consistance of algorithm and subjective evaluation result is better, this algorithm is also better.Can find out that method provided by the invention 982 scatter diagrams are near matched curve, illustrate that method effect after nonlinear fitting that the present invention proposes is better than other four kinds.

Table 1 LIVE image data base image quality evaluating method Performance comparision

By table 1 experimental data we find, quality evaluating method of the present invention is best in 6 kinds of evaluation indexes, mean absolute error and root-mean-square error is minimum, the highest, minimum from going out rate with the correlativity of subjective vision perception.Because PSNR algorithm does not consider the correlativity between each pixel, each pixel put on an equal footing, in 6 kinds of comparison algorithms, overall performance is the poorest.SSIM algorithm utilizes the structural information of image to carry out assess image quality, relevant to human eye visual perception pattern.MS-SSIM algorithm is on the basis of SSIM algorithm, and utilize multiresolution analysis technology to carry out multi-scale image quality assessment, therefore performance is better than PSNR and SSIM.Only obviously PSNR is better than to the svd algorithm performance that luminance component carries out svd, illustrates that the algorithm to image carries out svd has certain superiority.Visually GMSD algorithm and DMOS value nonlinear fitting curve are out closest to straight line, but can find out that some loose some distributions comparatively disperse, away from curve.Can find out that quality evaluating method of the present invention is obviously better than traditional PSNR algorithm, structural similarity SSIM algorithm, Multi-scale model similarity MS-SSIM algorithm, singular value decomposition algorithm SVD, gradient magnitude similarity deviation GMSD algorithm by last column of table 1, illustrate that the image quality evaluation algorithm based on hypercomplex number and human-eye visual characteristic of the present invention can reflect that human eye is experienced the subjective vision of image better.

Because the image word bank of 982 width images by 5 kinds of different type of distortion forms, for proving quality evaluating method superiority of the present invention further, the present embodiment, for 5 kinds of image word banks, carries out the Performance comparision of HVS-QSVD algorithm and GMSD algorithm, SSIM algorithm respectively.As shown in Figure 4, in figure, every three is one group, is respectively the nonlinear fitting curve map of HVS-QSVD algorithm, GMSD algorithm, SSIM algorithm, totally five groups.(a) in first picture group 4 is respectively JPEG2000, JPEG, white Gaussian noise, Gaussian Blur, Fast Fading Rayleigh channel distortion five kinds of type of distortion to (e) in the 5th picture group 4.Can find out, HVS-QSVD algorithm provided by the invention is all good than GMSD, SSIM algorithm with the fitting effect of subjective assessment value in different type of distortion situation.

A kind of color image quality evaluation method based on HVS and hypercomplex number described in embodiment 1, more conform to Human Perception for making evaluation result, utilize human-eye visual characteristic construct mathematical model, isolating the three-channel method of R, G, B for improving tradition, utilizing hypercomplex number svd to extract the characteristic information of image.Experimental result shows, evaluation result more conforms to the effect of Human Perception image.

Schematically above be described the present invention and embodiment thereof, this description does not have restricted, yet just one of the embodiments of the present invention shown in accompanying drawing, is actually not limited thereto.So, if those of ordinary skill in the art enlightens by it, when not departing from the invention aim, designing the frame mode similar to this technical scheme and embodiment without creationary, all should protection scope of the present invention be belonged to.

Claims (7)

1., based on a color image quality evaluation method for HVS and hypercomplex number, the steps include:

Step one, build the mathematics appraisal of original reference image and distorted image to be evaluated by analyzing human-eye visual characteristic, described mathematics appraisal comprises the function of spatial position Q of image _l, local variance Q _v, texture fringe complexity function Q _tEwith color function Q _c;

Step 2, by Q _l, Q _v, Q _tEas the imaginary part of hypercomplex number, Q _cas the real part of hypercomplex number, construct the Quaternion Matrix of original reference image and distorted image to be evaluated respectively, and the singular value feature vector that svd obtains image is carried out to Quaternion Matrix;

Step 3, utilize the euclidean distance metric image fault degree of the singular value feature vector of original reference image and distorted image to be evaluated.

2. a kind of color image quality evaluation method based on HVS and hypercomplex number according to claim 1, is characterized in that: the detailed process that step one builds mathematics appraisal is:

(1) the RGB tristimulus values of original reference image and distorted image to be evaluated is obtained;

(2) extract the spatial positional information of original reference image and distorted image to be evaluated, build function of spatial position Q _lwith texture fringe complexity function Q _tE;

(3) original reference image and distorted image to be evaluated are converted to YUV color space by rgb space, extract image luminance information and build local variance Q _v, extract brightness of image and chrominance information structure color function Q _c.

3. a kind of color image quality evaluation method based on HVS and hypercomplex number according to claim 2, is characterized in that: step one utilizes the middle concave characteristic of human visual system to build function of spatial position Q _l, described function of spatial position

$Q_L(i,j)=\frac{e_c}{e_c+e_L}$

In formula, e _lfor human eye Visual Observations Observations pixel (i, j) to picture centre pixel (M/2, N/2) distance with business; e _cfor constant.

4. a kind of color image quality evaluation method based on HVS and hypercomplex number according to Claims 2 or 3, is characterized in that: step one utilizes the shielding effect of human visual system to build texture fringe complexity function Q _tE, described texture fringe complexity function

Q _TE＝Q _T×Q _E

In formula, Q _tfor the Texture complication function of pixel (i, j), Q _efor the fringe complexity function of pixel (i, j).

5. a kind of color image quality evaluation method based on HVS and hypercomplex number according to claim 4, is characterized in that: step one utilizes the hyperchannel characteristic of human visual system to build local variance Q _v, described local variance

$Q_V\left(I_{i,j}\right)=\frac{1}{L}\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{({\mathrm{\&eta;}}_p-\stackrel{\&OverBar;}{I_{i,j}})}^2$

Wherein, the piecemeal carrying out non-overlapping copies according to the luminance component of image obtains I _i,j, L is image block I _i,jin the pixel η that comprises _pnumber, $\stackrel{\&OverBar;}{I_{i,j}}=\frac{1}{L}\underset{p=1}{\overset{L}{\mathrm{\&Sigma;}}}{\mathrm{\&eta;}}_p.$

6. a kind of color image quality evaluation method based on HVS and hypercomplex number according to claim 5, is characterized in that: described color function

Q _C＝αQ _L+βQ _U

In formula, Q _lfor the monochrome information of image, Q _ufor the chrominance information of image, α, β are respectively the proportion shared by brightness and colourity.

7. a kind of color image quality evaluation method based on HVS and hypercomplex number according to claim 6, is characterized in that: the Euclidean distance described in step 3

$D=\sqrt{\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}({\mathrm{\&lambda;}}_i-\widehat{{\mathrm{\&lambda;}}_i})}$

In formula, λ _ifor the singular value feature vector of original reference image, for the singular value feature vector of distorted image to be evaluated, K is the minimum value of two singular value feature vector eigenwert numbers, i.e. the minimum value of two Quaternion Matrix orders:

$K=\min (\mathrm{rank}\left(I\right),\mathrm{rank}\left(\hat{I}\right)).$