CN104574399A - Image quality evaluation method based on multi-scale vision significance and gradient magnitude - Google Patents

Abstract

The invention relates to the technical field of digital image quality evaluation, and provides a full-reference image quality evaluation method, so that an objective image quality evaluation result better meets the subjective feeling of a human visual system. According to the technical scheme, the image quality evaluation method based on multi-scale vision significance and the gradient magnitude includes the following steps that Laplacian pyramid transformation is adopted; a phase spectrum based on Fourier transform is calculated and converted into an airspace to obtain a visual saliency map; the value of the gradient magnitude of an image is calculated through a Scharr gradient operator; the similarity value VSGMj (x,y) of a reference image and a distortion image at the position (x,y) in the same pyramid layer is obtained; the image quality evaluation value MS-VSGM based on multi-scale vision significance and the gradient magnitude is obtained. The image quality evaluation method based on multi-scale vision significance and the gradient magnitude is mainly used for quality evaluation of digital images.

Description

Based on the image quality evaluating method of multiple scale vision conspicuousness and gradient magnitude

Technical field

The present invention relates to digital picture quality assessment technique field, is a kind of full reference type based on the image quality evaluating method of multiple scale vision conspicuousness and gradient magnitude.

Technical background

Tele-medicine is study hotspot new in biomedical sector and computer communication field in recent years.The transmission of the data such as tele-medicine image is mainly carried out in tele-medicine by means such as networks, thus realizes remote medical diagnosis.The transmission of current tele-medicine image (comprising other data files) mainly through following steps, the collection of view data and compression, the transmission of view data and reception, the decompress(ion) of view data and reproduction.In the process transmitted, above-mentioned each step can cause damage to the image quality of image, and high-quality medical image can help doctor to make judges accurately, avoid causing false positive and false-negative diagnosis due to image fault, the quality assessment being therefore applied to tele-medicine image becomes particularly important.

Image quality evaluation with the subjective assessment of people for standard, but this evaluation method is not suitable for real-time process, therefore needs to develop objective quality evaluating method.Conventional structural similarity (Structural Similarity, SSIM) method is evaluated the mass loss of distorted image.Picture breakdown is different yardsticks by Multi-scale model similarity (Multi-scale SSIM, MS-SSIM) method, is weighted according to the importance difference of each yardstick to image, its evaluation effect than structural similarity method closer to subjective assessment.Visual information fidelity (VisualInformation Fidelity, VIF) picture breakdown is the frequency band with different space frequency and direction by pyramid transform by method, each frequency band is set up the gauss hybrid models of image, the shortcoming of this method is that computation complexity is high, and arithmetic speed is slow.

Summary of the invention

For overcoming the deficiencies in the prior art, a kind of full reference image quality appraisement method being provided, making Objective image quality evaluation result more meet the subjective feeling of human visual system.For this reason, the technical scheme that the present invention takes is, based on the image quality evaluating method of multiple scale vision conspicuousness and gradient magnitude, comprises the steps:

Step 1 adopts Laplacian Pyramid Transform, is decomposed into the pyramidal layer of 5 different resolutions respectively with reference to image and distorted image;

Step 2 is for each pyramidal layer of image, calculate the phase spectrum based on Fourier transform, then phase spectrum is transformed into spatial domain and can obtains visual saliency map (phase spectrum visual saliency map, PSVS), if piece image is f (x, y), Fourier transform is done to image, image is changed to frequency domain from transform of spatial domain, in other words, the distributed function of image is transformed to the frequency distribution function of image;

F(u,v)＝F(f(x,y)) (1)

F is the Fourier transform asking image, according to real part and the imaginary part of function after Fourier transform, calculates the phase spectrum of image with trigonometric function, i.e. A (u, v);

A(u,v)＝angle(F(u,v)) (2)

The phase spectrum of image can be tried to achieve by angle (), inverse Fourier transform is carried out to the phase spectrum of image, by acquired results square with 2-d gaussian filters function do convolution algorithm, obtain the visual saliency map PSVS (x, y) based on phase spectrum:

PSVS(x,y)＝g(x,y)*(F ^-1[exp(i·A(u,v))]) ²(3)

X, y are volume coordinate, and u, v are frequencies different in frequency domain, F ^-1for inverse Fourier transform, g (x, y) is 2-d gaussian filters device, σ=8, and sigma is standard variance, and * represents convolution algorithm; Reference picture and the distorted image similarity P between the visual saliency map of same pyramidal layer _j(x, y) is:

$P_j(x,y)={\left(\frac{{2\mathrm{PSVS}}_1(x,y)\·{\mathrm{PSVS}}_2(x,y)}{{\mathrm{PSVS}}_1^2(x,y)+{\mathrm{PSVS}}_2^2(x,y)}\right)}_j---\left(4\right)$

1≤j≤5, j represents different pyramidal layer, PSVS ₁(x, y) and PSVS ₂(x, y) is respectively reference picture and the distorted image visual saliency map in same pyramidal layer;

Step 3, for each pyramidal layer of image, adopts the value of the gradient amplitude of Scharr gradient operator computed image, the first partial derivative g of computed image f (x, y) _x(x, y) and g _y(x, y):

$g_x(x,y)=\frac{1}{16}\left[\begin{array}{ccc}3& 0& -3\\ 10& 0& -10\\ 3& 0& -3\end{array}\right]*f(x,y)---\left(5\right)$

$g_y(x,y)=\frac{1}{16}\left[\begin{array}{ccc}3& 10& 3\\ 10& 0& 0\\ -3& -10& -3\end{array}\right]*f(x,y)---\left(6\right)$

Then pass through obtain the gradient magnitude of (x, y) position in image, reference picture and the distorted image similarity G between the gradient magnitude of same pyramidal layer _j(x, y) is:

$G_j(x,y)={\left(\frac{{2\mathrm{GA}}_1(x,y)\·{\mathrm{GA}}_2(x,y)}{{\mathrm{GA}}_1^2(x,y)+{\mathrm{GA}}_2^2(x,y)}\right)}_j---\left(7\right)$

GA ₁(x, y) and GA ₂(x, y) is respectively reference picture and the distorted image value in the gradient amplitude of same pyramidal layer;

Step 4 by P (x, y) and G (x, y) according to certain multiplied by weight and summation can obtain reference picture and distorted image at the Similarity value VSGMj (x, y) of same pyramidal layer in position (x, y) place:

VSGM(x,y) _j＝∑ _x,y∈Ω[P _j(x,y)] ^α·[G _j(x,y)] ^β(8)

Ω represents the spatial domain of image, α and β is the weighted value of balance vision significance and gradient magnitude, selects α=1 here, β=0.5;

Select value larger between P1 (x, y) and P2 (x, y), the weighted value as quality assessment value between computing reference image and distorted image:

P _max(x,y)＝max(P ₁(x,y),P ₂(x,y)) (9)

By the Similarity value between the computed image by pixel, just obtain the value SS-VSGM of the image quality evaluation based on single scale (Single-Scale, SS) vision significance and gradient characteristics _j:

$\mathrm{SS}-{\mathrm{VSGM}}_j=\frac{{\mathrm{\Σ}}_{x,y\∈\mathrm{\Ω}}\mathrm{VSGM}(x,y)\·P_{\max }(x,y)}{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{P}_{\max }(x,y)}---\left(10\right)$

Step 5 is weighted the value MS-VSGM of the image quality evaluation obtained based on multi-scale image vision significance and gradient magnitude that is multiplied with reference to image and the quality assessment value of distorted image on five different scales:

$\mathrm{MS}-\mathrm{VSGM}=\underset{j=1}{\overset{5}{\mathrm{\Π}}}{(\mathrm{SS}-{\mathrm{VSGM}}_j)}^{\mathrm{\γ}}---\left(11\right)$

Here γ value represents different weighted values, is respectively 0.0448,0.2856,0.3001,0.2363,0.1333.

Compared with the prior art, technical characterstic of the present invention and effect:

The invention provides the quality evaluating method of a kind of full reference type based on multiple scale vision conspicuousness and gradient magnitude.The method is applicable to the quality assessment of tele-medicine image.Because full reference type evaluation method can utilize the full detail of original image, therefore opposite segments reference type and without reference type evaluation method, full reference type more meets human subject to the evaluation result of image and evaluates, and has a wide range of applications.

Accompanying drawing explanation

Fig. 1 method block diagram of the present invention.

Fig. 2 (a) reference picture; (b) JPEG 2000 compression artefacts image.

Fig. 3 (a) carries out the ground floor image after the decomposition of Laplce's gold tower to reference picture; B () carries out the ground floor image after the decomposition of Laplce's gold tower to distorted image.

The visual saliency map of Fig. 4 (a) reference picture; The visual saliency map of (b) distorted image.

Embodiment

The present invention proposes a kind of image quality evaluating method based on multiple scale vision conspicuousness and gradient magnitude MVSGM (multi-scale visual saliencyand gradient magnitude).The step of MVSGM algorithm as shown in Figure 1.Concrete steps are as follows:

Step 1 adopts Laplacian Pyramid Transform, is decomposed into the pyramidal layer of 5 different resolutions respectively with reference to image and distorted image.

Step 2, for each pyramidal layer of image, calculates the phase spectrum based on Fourier transform, then phase spectrum is transformed into spatial domain and can obtains visual saliency map (phase spectrum visual saliency map, PSVS).If piece image is f (x, y), Fourier transform is done to image, image is changed to frequency domain from transform of spatial domain, in other words, the distributed function of image is transformed to the frequency distribution function of image.

F(u,v)＝F(f(x,y)) (1)

F is the Fourier transform asking image.According to real part and the imaginary part of function after Fourier transform, calculate the phase spectrum of image with trigonometric function, i.e. A (u, v).

A(u,v)＝angle(F(u,v)) (2)

The phase spectrum of image can be tried to achieve by angle ().Inverse Fourier transform is carried out to the phase spectrum of image, by acquired results square with 2-d gaussian filters function do convolution algorithm, the visual saliency map PSVS (x, y) based on phase spectrum can be obtained:

PSVS(x,y)＝g(x,y)*(F ^-1[exp(i·A(u,v))]) ²(3)

X, y are volume coordinate, and u, v are the frequency of harmonic component, are frequencies different in frequency domain, F ^-1for inverse Fourier transform, g (x, y) is 2-d gaussian filters device (σ=8, sigma is standard variance), and * represents convolution algorithm.Reference picture and the distorted image similarity Pj (x, y) between the visual saliency map of same pyramidal layer (1≤j≤5, j represents different pyramidal layer) is:

$P_j(x,y)={\left(\frac{{2\mathrm{PSVS}}_1(x,y)\·{\mathrm{PSVS}}_2(x,y)}{{\mathrm{PSVS}}_1^2(x,y)+{\mathrm{PSVS}}_2^2(x,y)}\right)}_j---\left(4\right)$

PSVS1 (x, y) and PSVS2 (x, y) are respectively reference picture and the distorted image visual saliency map in same pyramidal layer.

Step 3, for each pyramidal layer of image, adopts the value of the gradient amplitude of Scharr gradient operator computed image.First the partial derivative gx (x, y) of computed image f (x, y) and gy (x, y):

$g_x(x,y)=\frac{1}{16}\left[\begin{array}{ccc}3& 0& -3\\ 10& 0& -10\\ 3& 0& -3\end{array}\right]*f(x,y)---\left(5\right)$

$g_y(x,y)=\frac{1}{16}\left[\begin{array}{ccc}3& 10& 3\\ 10& 0& 0\\ -3& -10& -3\end{array}\right]*f(x,y)---\left(6\right)$

Then pass through obtain the gradient magnitude of (x, y) position in image.Reference picture and the distorted image similarity Gj (x, y) between the gradient magnitude of same pyramidal layer (1≤j≤5, j represents different pyramidal layer) is:

$G_j(x,y)={\left(\frac{{2\mathrm{GA}}_1(x,y)\·{\mathrm{GA}}_2(x,y)}{{\mathrm{GA}}_1^2(x,y)+{\mathrm{GA}}_2^2(x,y)}\right)}_j---\left(7\right)$

GA1 (x, y) and GA2 (x, y) are respectively reference picture and the distorted image value in the gradient amplitude of same pyramidal layer.

Step 4 is by P (x, y) with G (x, y) according to certain multiplied by weight and summation can obtain reference picture and distorted image in same pyramidal layer in position (x, y) the Similarity value VSGMj (x at place, y) (1≤j≤5, j represents different pyramidal layer):

VSGM(x,y) _j＝∑ _x,y∈Ω[P _j(x,y)] ^α·[G _j(x,y)] ^β(8)

Ω represents the spatial domain of image, α and β is the weighted value of balance vision significance and gradient magnitude, selects α=1 here, β=0.5.

Select value larger between P1 (x, y) and P2 (x, y), the weighted value as quality assessment value between computing reference image and distorted image:

P _max(x,y)＝max(P ₁(x,y),P ₂(x,y)) (9)

By the Similarity value between the computed image by pixel, just obtain the value SS-VSGMj (1≤j≤5, j represents different yardsticks) of the image quality evaluation based on single scale (Single-Scale, SS) vision significance and gradient characteristics:

$\mathrm{SS}-{\mathrm{VSGM}}_j=\frac{{\mathrm{\Σ}}_{x,y\∈\mathrm{\Ω}}\mathrm{VSGM}(x,y)\·P_{\max }(x,y)}{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{P}_{\max }(x,y)}---\left(10\right)$

Ω represents the spatial domain part of image.

Step 5 is weighted the value MS-VSGM of the image quality evaluation that can obtain based on multi-scale image vision significance and gradient magnitude that is multiplied with reference to image and the quality assessment value of distorted image on five different scales:

$\mathrm{MS}-\mathrm{VSGM}=\underset{j=1}{\overset{5}{\mathrm{\Π}}}{(\mathrm{SS}-{\mathrm{VSGM}}_j)}^{\mathrm{\γ}}---\left(11\right)$

Here γ value represents different weighted values, is respectively 0.0448,0.2856,0.3001,0.2363,0.1333.

By above five steps, just can calculate the quality assessment value of image quickly and efficiently, picture quality is evaluated accurately, thus obtain the evaluation result higher with human eye subjective assessment consistance.

The concrete steps of the image quality evaluating method based on multiple scale vision conspicuousness and gradient magnitude that the present invention realizes as shown in Figure 1.

As shown in Figure 2, (a) is undistorted reference picture, the distorted image that (b) obtains for carrying out JPEG2000 compression on reference picture basis.

According to evaluation method proposed by the invention, first Laplce's gold tower is carried out respectively to two width images and decompose.Respectively two width images are respectively decomposed into the pyramidal layer of 5 different resolutions, wherein reference picture and distorted image decompose the image of the 1st layer that obtains respectively as shown in Figure 3.

Respectively Fourier transform is carried out to each pyramidal layer of reference picture and distorted image, the image of every one deck is changed to frequency domain from transform of spatial domain one by one.Then according to real part and the imaginary part of the function obtained after Fourier transform, the phase spectrum of each tomographic image is calculated with trigonometric function.On this basis, inverse Fourier transform is carried out to the phase spectrum of image, by acquired results square with 2-d gaussian filters function carry out convolution algorithm, just can obtain the visual saliency map of each tomographic image based on phase spectrum, as shown in Figure 4.

For each pyramidal layer of image, Scharr gradient operator is adopted to calculate the value of the gradient amplitude of each tomographic image.

Respectively the visual saliency map of each obtained pyramidal layer and gradient magnitude are sued for peace according to certain multiplied by weight, just can obtain the Similarity value of same pyramidal layer between reference picture and distorted image.Select value larger in reference picture and the Similarity value of distorted image between same pyramidal layer visual saliency map, as the weighted value of quality assessment value between computing reference image and distorted image.By the Similarity value between the computed image by pixel, just obtain the value of the image quality evaluation based on single scale vision significance and gradient characteristics.The value of the image quality evaluation of five different scales being weighted is multiplied just can obtain the value of the image quality evaluation based on multiple scale vision conspicuousness and gradient magnitude.In Fig. 2, the image quality evaluation values of reference picture is 1, and the image quality evaluation values of JPEG 2000 compression artefacts image is 0.9673.Can see, the method utilizing the present invention to propose can make evaluation to picture quality fast and effectively, consistent with the result height of human eye subjective assessment.

Claims (1)

1., based on an image quality evaluating method for multiple scale vision conspicuousness and gradient magnitude, it is characterized in that, comprise the steps:

Step 1 adopts Laplacian Pyramid Transform, is decomposed into the pyramidal layer of 5 different resolutions respectively with reference to image and distorted image;

Step 2 is for each pyramidal layer of image, calculate the phase spectrum based on Fourier transform, then phase spectrum is transformed into spatial domain and can obtains visual saliency map (phase spectrum visual saliency map, PSVS), if piece image is f (x, y), Fourier transform is done to image, image is changed to frequency domain from transform of spatial domain, in other words, the distributed function of image is transformed to the frequency distribution function of image;

F(u,v)＝F(f(x,y)) (1)

F is the Fourier transform asking image, according to real part and the imaginary part of function after Fourier transform, calculates the phase spectrum of image with trigonometric function, i.e. A (u, v);

A(u,v)＝angle(F(u,v)) (2)

The phase spectrum of image can be tried to achieve by angle (), inverse Fourier transform is carried out to the phase spectrum of image, by acquired results square with 2-d gaussian filters function do convolution algorithm, obtain the visual saliency map PSVS (x, y) based on phase spectrum:

PSVS(x,y)＝g(x,y)*(F ^-1[exp(i·A(u,v))]) ²(3)

X, y are volume coordinate, and u, v are frequencies different in frequency domain, F ^-1for inverse Fourier transform, g (x, y) is 2-d gaussian filters device, σ=8, and sigma is standard variance, and * represents convolution algorithm; Reference picture and the distorted image similarity P between the visual saliency map of same pyramidal layer _j(x, y) is:

$P_j(x,y)={\left(\frac{{2\mathrm{PSVS}}_1(x,y)\·{\mathrm{PSVS}}_2(x,y)}{{\mathrm{PSVS}}_1^2(x,y)+{\mathrm{PSVS}}_2^2(x,y)}\right)}_j---\left(4\right)$

1≤j≤5, j represents different pyramidal layer, PSVS ₁(x, y) and PSVS ₂(x, y) is respectively reference picture and the distorted image visual saliency map in same pyramidal layer;

Step 3, for each pyramidal layer of image, adopts the value of the gradient amplitude of Scharr gradient operator computed image, the first partial derivative g of computed image f (x, y) _x(x, y) and g _y(x, y):

$g_x(x,y)=\frac{1}{16}\left[\begin{array}{ccc}3& 0& -3\\ 10& 0& -10\\ 3& 0& -3\end{array}\right]*f(x,y)---\left(5\right)$

$g_y(x,y)=\frac{1}{16}\left[\begin{array}{ccc}3& 10& 3\\ 10& 0& 0\\ -3& -10& -3\end{array}\right]*f(x,y)---\left(6\right)$

Then pass through obtain the gradient magnitude of (x, y) position in image, reference picture and the distorted image similarity G between the gradient magnitude of same pyramidal layer _j(x, y) is:

$G_j(x,y)={\left(\frac{{2\mathrm{GA}}_1(x,y)\·{\mathrm{GA}}_2(x,y)}{{\mathrm{GA}}_1^2(x,y)+{\mathrm{GA}}_2^2(x,y)}\right)}_j---\left(7\right)$

GA ₁(x, y) and GA ₂(x, y) is respectively reference picture and the distorted image value in the gradient amplitude of same pyramidal layer;

Step 4 by P (x, y) and G (x, y) according to certain multiplied by weight and summation obtains reference picture and distorted image at the Similarity value VSGMj (x, y) of same pyramidal layer in position (x, y) place:

VSGM(x,y) _j＝Σ _x,y∈Ω[P _j(x,y)] ^α·[G _j(x,y)] ^β(8)

Ω represents the spatial domain of image, α and β is the weighted value of balance vision significance and gradient magnitude, selects α=1 here, β=0.5;

Select value larger between P1 (x, y) and P2 (x, y), the weighted value as quality assessment value between computing reference image and distorted image:

P _max(x,y)＝max(P ₁(x,y),P ₂(x,y)) (9)

By the Similarity value between the computed image by pixel, just obtain the value SS-VSGM of the image quality evaluation based on single scale (Single-Scale, SS) vision significance and gradient characteristics _j:

$\mathrm{SS}-{\mathrm{VSGM}}_j=\frac{{\mathrm{\Σ}}_{x,y\∈\mathrm{\Ω}}\mathrm{VSGM}(x,y)\·P_{\max }(x,y)}{{\mathrm{\Σ}}_{x\∈\mathrm{\Ω}}{P}_{\max }(x,y)}---\left(10\right)$

Step 5 is weighted the value MS-VSGM of the image quality evaluation obtained based on multi-scale image vision significance and gradient magnitude that is multiplied with reference to image and the quality assessment value of distorted image on five different scales:

$\mathrm{MS}-\mathrm{VSGM}=\underset{j=1}{\overset{5}{\mathrm{\Π}}}{(\mathrm{SS}-{\mathrm{VSGM}}_j)}^{\mathrm{\γ}}---\left(11\right)$

Here γ value represents different weighted values, is respectively 0.0448,0.2856,0.3001,0.2363,0.1333.