Summary of the invention
Technical problem to be solved by this invention provides a kind of adaptive image quality method for objectively evaluating, and it can improve the consistency between picture quality objective evaluation result and the subjective perception effectively.
The present invention solves the problems of the technologies described above the technical scheme that adopts: a kind of adaptive image quality method for objectively evaluating is characterized in that its processing procedure is:
At first, determine the type of distortion of distorted image to be evaluated, and original undistorted image and distorted image to be evaluated are divided into a plurality of equitant sizes respectively is 8 * 8 image block;
Secondly, by calculating brightness average and the standard deviation of all pixels in each image block in original undistorted image and the distorted image to be evaluated, and the covariance between all pixels in two image blocks that all coordinate position is identical in original undistorted image and the distorted image to be evaluated, and in conjunction with the type of distortion of distorted image to be evaluated, obtain the structural similarity between two identical image blocks of coordinate positions all in original undistorted image and the distorted image to be evaluated;
At last, according to the structural similarity between two identical image blocks of coordinate positions all in original undistorted image and the distorted image to be evaluated, obtain the objective quality score value of distorted image to be evaluated.
A kind of adaptive image quality method for objectively evaluating of the present invention is characterized in that it specifically may further comprise the steps:
1. make X represent original undistorted image, make Y represent distorted image to be evaluated, determine the type of distortion of Y then by the type of distortion method of discrimination, the type of distortion of Y is wherein a kind of in white Gaussian noise distortion, JPEG distortion, Gaussian Blur distortion, the class JPEG2000 distortion, and wherein class JPEG2000 distortion comprises JPEG2000 distortion and rapid fading distortion;
2. adopting size is that 8 * 8 sliding window moves by pixel in X, and it is 8 * 8 image block that X is divided into M * N equitant and size, is that (i, image block j) is designated as x with coordinate position among the X
I, jEqually, adopting size is that 8 * 8 sliding window moves by pixel in Y, and it is 8 * 8 image block that Y is divided into M * N equitant and size, is that (i, image block j) is designated as y with coordinate position among the Y
I, jWherein,
H represents the height of X and Y, and W represents the width of X and Y, symbol
For rounding symbol downwards, 1≤i≤M, 1≤j≤N;
3. calculate brightness average and the standard deviation of all pixels in each image block among the X, and brightness average and the standard deviation of all pixels in each image block among the calculating Y, calculate the covariance between all pixels in two identical image blocks of coordinate positions all among X and the Y then, be (i, image block x j) with coordinate position among the X
I, jIn brightness average and the standard deviation correspondence of all pixels be designated as
With
Be (i, image block y j) with coordinate position among the Y
I, jIn brightness average and the standard deviation correspondence of all pixels be designated as
With
Be (i, image block x j) with coordinate position among the X
I, jIn all pixels and Y in coordinate position be (i, image block y j)
I, jIn all pixels between covariance be designated as
Wherein, x
I, j(u v) represents x
I, jMiddle coordinate position is (u, the brightness value of pixel v), y
I, j(u v) represents y
I, jMiddle coordinate position is (u, the brightness value of pixel v), 1≤u≤8,1≤v≤8;
4. calculating luminance function, contrast function and degree of structuration function between two identical image blocks of coordinate positions all among X and the Y, is (i, image block x j) with coordinate position among the X
I, jBe (i, image block y j) with coordinate position among the Y
I, jBetween luminance function, contrast function and degree of structuration function correspondence be designated as l (x respectively
I, j, y
I, j), c (x
I, j, y
I, j) and s (x
I, j, y
I, j),
Wherein, C
1, C
2, C
3For avoiding denominator the zero little numerical constant that arranges to occur;
5. according to luminance function, contrast function and degree of structuration function between two identical image blocks of coordinate positions all among X and the Y, calculate the structural similarity between two identical image blocks of coordinate positions all among X and the Y, be (i, image block x j) with coordinate position among the X
I, jBe (i, image block y j) with coordinate position among the Y
I, jBetween structural similarity be designated as SSIM (x
I, j, y
I, j),
Wherein, t is used for the type of distortion of expression Y, gets t=1 when the type of distortion of Y is the white Gaussian noise distortion, gets t=2 when the type of distortion of Y is the JPEG distortion, when the type of distortion of Y is the Gaussian Blur distortion, get t=3, when the type of distortion of Y is class JPEG2000 distortion, get t=4; α
tThe weight factor of brightness, β are regulated in expression
tThe weight factor of contrast, γ are regulated in expression
tThe weight factor of expression adjustment structure degree, and satisfy α
t+ β
t+ γ
t=3.
6. according to the structural similarity between two identical image blocks of coordinate positions all among X and the Y, calculate the objective quality score value of Y, be designated as Q,
Described step determines that by the type of distortion method of discrimination detailed process of the type of distortion of Y is in 1.:
1.-and a, X is carried out 2-d wavelet decompose, obtain the approximate component X of X
A, horizontal component
And vertical component
And calculate
With
The small echo gross energy, be designated as
Then to the approximate component X of X
ACarry out a 2-d wavelet again and decompose, obtain the approximate component X of X
AThe subband horizontal component
With the subband vertical component
And calculate
With
The small echo gross energy, be designated as
At this, 1≤m'≤M'
1, 1≤n'≤N'
1, M'
1Expression
With
Height, N'
1Expression
With
Width,
Expression
In coordinate position be (m', the coefficient value of n') locating,
Expression
Middle coordinate position is (m', the coefficient value of n') locating, 1≤m''≤M'
2, 1≤n''≤N'
2, M'
2Expression
With
Height, N'
2Expression
With
Width,
Expression
In coordinate position be (m'', the coefficient value of n'') locating,
Expression
Middle coordinate position is (m'', the coefficient value of n'') locating;
Equally, Y is carried out a 2-d wavelet decompose, obtain the approximate component Y of Y
A, horizontal component
And vertical component
And calculate
With
The small echo gross energy, be designated as
Then to the approximate component Y of Y
ACarry out a 2-d wavelet again and decompose, obtain the approximate component Y of Y
AThe subband horizontal component
With the subband vertical component
And calculate
With
The small echo gross energy, be designated as
At this, 1≤m'≤M'
1, 1≤n'≤N'
1, M'
1Expression
With
Height, N'
1Expression
With
Width,
Expression
In coordinate position be (m', the coefficient value of n') locating,
Expression
Middle coordinate position is (m', the coefficient value of n') locating, 1≤m''≤M'
2, 1≤n''≤N'
2, M'
2Expression
With
Height, N'
2Expression
With
Width,
Expression
In coordinate position be (m'', the coefficient value of n'') locating,
Expression
Middle coordinate position is (m'', the coefficient value of n'') locating;
Calculate the capacity volume variance of the small echo gross energy of the X correspondence small echo gross energy corresponding with Y again, be designated as Δ W,
1.-b, judgement Δ W<Th
WnWhether set up, if set up, determine that then the type of distortion of Y is the white Gaussian noise distortion, finish then; Otherwise, execution in step 1.-c; Wherein, Th
WnBe white Gaussian noise distortion discrimination threshold;
1.-and c, calculate the luminance difference figure of X, be designated as X
h, with X
hMiddle coordinate position is that (x, the brightness value of pixel y) is designated as X
h(x, y), X
h(x, y)=| X (x, y)-X (x, y+1) |, wherein, 1≤x≤H herein, 1≤y≤W-1, X (x, y) coordinate position is (x, the brightness value of pixel y) among the expression X, X (x, y+1) coordinate position is that (symbol " || " is the symbol that takes absolute value for x, the brightness value of pixel y+1) among the expression X;
Equally, calculate the luminance difference figure of Y, be designated as Y
h, with Y
hMiddle coordinate position is that (x, the brightness value of pixel y) is designated as Y
h(x, y), Y
h(x, y)=| Y (x, y)-Y (x, y+1) |, wherein, and Y (x, y) coordinate position is that ((x y+1) represents that coordinate position is (x, the brightness value of pixel y+1) among the Y to Y for x, the brightness value of pixel y) among the expression Y;
1.-d, to X
hIn every row in all pixels carry out N
hLeaf transformation in the point discrete Fourier obtains X
hIn Fourier's energy spectrum of every row, with X
hIn the capable Fourier's energy spectrum of x be designated as
Wherein,
Symbol
Be the symbol that rounds up, l ∈ [1,
N
h2+1], X
f x(l) expression X
hIn x all pixels in capable carry out N
hL the DFT coefficient value that obtains behind the leaf transformation in the point discrete Fourier, symbol " || " is the symbol that takes absolute value;
Equally, to Y
hIn every row in all pixels carry out N
hLeaf transformation in the point discrete Fourier obtains Y
hIn Fourier's energy spectrum of every row, with Y
hIn the capable Fourier's energy spectrum of x be designated as
Wherein,
Expression Y
hIn x all pixels in capable carry out N
hL the DFT coefficient value that obtains behind the leaf transformation in the point discrete Fourier; 1.-e, according to X
hIn Fourier's energy spectrum of every row, calculate X
hWhole Fourier's energy spectrum, be designated as P
X,
And according to Y
hIn Fourier's energy spectrum of every row, calculate Y
hWhole Fourier's energy spectrum, be designated as P
Y,
Wherein, l ∈ [1, N
h/ 2+1]; Calculate P then
XWith P
YFourier's energy difference, be designated as Δ P,
1.-f, according to Δ P, the JPEG distortion first judgment threshold Th
Jpeg1With the JPEG distortion second judgment threshold Th
Jpeg2, determine the type of distortion of Y, as Δ P 〉=Th
Jpeg1The time, the type of distortion of determining Y is the JPEG distortion, finishes then; Work as Th
Jpeg2≤ Δ P<Th
Jpeg1The time, the type of distortion of uncertain Y, then execution in step 1.-g; As Δ P<Th
Jpeg2The time, the type of distortion of determining Y is Gaussian Blur distortion or class JPEG2000 distortion, then execution in step 1.-h;
1.-and g, calculate Y and compare the percentage that X small echo gross energy reduces, be designated as Δ W',
Judge Δ W'<Th then
Jpeg3Whether set up, if set up, determine that then the type of distortion of Y is the JPEG distortion, finish then, otherwise, execution in step 1.-h, wherein, Th
Jpeg3Be JPEG distortion the 3rd judgment threshold;
1.-h, adopt the Sobel operator respectively R, G, the B Color Channel subgraph of X to be carried out edge extracting, find out the pixel of brightness value maximum in the edge of each Color Channel subgraph of X then, the brightness value of the pixel of brightness value maximum in the edge of the c Color Channel subgraph of X is designated as
Wherein, c=R, G, B;
Equally, adopt the Sobel operator respectively R, G, the B Color Channel subgraph of Y to be carried out edge extracting, find out the pixel of brightness value maximum in the edge of each Color Channel subgraph of Y then, the brightness value of the pixel of brightness value maximum in the edge of the c Color Channel subgraph of Y is designated as
Judge again
With
Whether all set up, if all set up, then execution in step 1.-i; Otherwise the type of distortion of determining Y is class JPEG2000 distortion, finishes then;
1.-i, with X from the RGB color space conversion to the HSI color space, extract the chromatic component of X then, be designated as X
H, then calculate X
HThe average chrominance value, be designated as
Wherein, X
H(m, n) expression X
HMiddle coordinate position is at (m, the chromatic value of n) locating;
Equally, with Y from the RGB color space conversion to the HIS color space, extract the chromatic component of Y then, be designated as Y
H, then calculate Y
HThe average chrominance value, be designated as
Wherein, Y
H(m, n) expression Y
HMiddle coordinate position is at (m, the chromatic value of n) locating;
Afterwards, calculate the colourity relative different of X and Y, be designated as C,
Judge again whether C<0.3 sets up, if set up, then execution in step 1.-j; Otherwise the type of distortion of determining Y is class JPEG2000 distortion, finishes then;
1.-and j, the R to X, G, B Color Channel subgraph carry out two dimensional discrete Fourier transform respectively, obtains the amplitude spectrum of the R Color Channel subgraph of X, the amplitude spectrum of G Color Channel subgraph and the amplitude spectrum of B Color Channel subgraph, and correspondence is designated as respectively
With
Equally, R, G, the B Color Channel subgraph of Y carried out two dimensional discrete Fourier transform respectively, obtain the amplitude spectrum of the R Color Channel subgraph of Y, the amplitude spectrum of G Color Channel subgraph and the amplitude spectrum of B Color Channel subgraph, correspondence is designated as respectively
Then, calculate the frequency response of each Color Channel subgraph, the frequency response of c Color Channel subgraph is designated as
Will
In coordinate position in that (m, the DFT coefficient value of n) locating is designated as
(m, n),
Wherein,
(m, n) expression
In coordinate position (m, the DFT coefficient value of n) locating,
(m, n) expression
Middle coordinate position is at (m, the DFT coefficient value of n) locating;
Then, extract the vertical zero-frequency component of the frequency response of each Color Channel subgraph, with the frequency response of c Color Channel subgraph
Vertical zero-frequency component be designated as
Will
In m DFT coefficient value be designated as
Wherein,
Expression
The DFT coefficient value that middle coordinate position is located in (m, 1);
Afterwards, in the middle of the center zero-frequency of the vertical zero-frequency component of the frequency response of each Color Channel subgraph moved on to, obtain new vertical zero-frequency component, the new vertical zero-frequency component of the frequency response of c Color Channel subgraph is designated as
Will
In m DFT coefficient value be designated as
Wherein, symbol
For rounding symbol downwards;
Extract again decentre point in the new vertical zero-frequency component of frequency response of each Color Channel subgraph ±
The Frequency point of distance, to use window width be 3 filter carries out medium filtering to the Frequency point of each Color Channel subgraph correspondence of extracting, with filtered Frequency point and the match of one dimension Gaussian function, the Frequency point after the calculation of filtered and the Pearson correlation coefficient of the data after the match; For
Extract
The decentre point ±
The Frequency point of distance is designated as
Using window width is that 3 filter is right
Carry out medium filtering, obtain filtered Frequency point, be designated as
Will
With the match of one dimension Gaussian function, obtain the data after the match, be designated as
Calculate
With
Pearson correlation coefficient, be designated as Pe
c,
Wherein,
For
In the DFT coefficient value of d,
For
In the DFT coefficient value after match of d;
At last, judge Pe
R≤ 0.9, Pe
G≤ 0.9 and Pe
BWhether≤0.9 all set up, if all set up, determines that then the type of distortion of Y is class JPEG2000 distortion, finishes then; Otherwise the type of distortion of determining Y is the Gaussian Blur distortion, finishes then.
Described step 1.-b in white Gaussian noise distortion discrimination threshold Th
WnValue be-0.1; Described step 1.-f in the JPEG distortion first judgment threshold Th
Jpeg1Value be 0.1, JPEG distortion, the second judgment threshold Th
Jpeg2Value be-0.2; Described step 1.-g in JPEG distortion the 3rd judgment threshold Th
Jpeg3Value be 0.075.
Described step is got C in 4.
1=0.01, C
2=0.02, C
3=0.01.
Described step is middle α 5.
t, β
tAnd γ
tValue determined by the type of distortion of Y, when the type of distortion of Y is the white Gaussian noise distortion, get α
t=2.5, β
t=0.1, γ
t=0.4; When the type of distortion of Y is the JPEG distortion, get
α
t=1.3, β
t=1.6, γ
t=0.1; When the type of distortion of Y is the Gaussian Blur distortion, get α
t=2.9, β
t=0, γ
t=0.1; When the type of distortion of Y is class JPEG2000 distortion, get α
t=0.8, β
t=2, γ
t=0.2.
Compared with prior art, the invention has the advantages that:
1) during the structural similarity between the inventive method two image blocks that coordinate position is identical in obtaining original undistorted image and distorted image to be evaluated, brightness average and the standard deviation of all pixels in each image block in original undistorted image and the distorted image to be evaluated have not only been utilized, and the covariance between all pixels in two image blocks that all coordinate position is identical in original undistorted image and the distorted image to be evaluated, but also combine the type of distortion of distorted image to be evaluated, make that the inventive method can be from the angle of self adaptation evaluation, determine two structural similarity between the image block according to the type of distortion of distorted image to be evaluated, thereby improved the consistency between picture quality objective evaluation result and the subjective perception.
2) the inventive method is at the weights ratio of brightness, contrast and degree of structuration three in the structural similarity between two image blocks of the adaptive adjustment of the distorted image of different type of distortion, come distorted image is carried out optimum evaluation, not only computation complexity is low, and utilized the evaluating ability of structural similarity to the distorted image of different type of distortion, therefore improved the consistency of objective evaluation result and subjective perception effectively.
The distortion characteristics that show when 3) the inventive method is subjected to the white Gaussian noise distortion by combining image in the process of the type of distortion of differentiating distorted image, the distortion characteristics that show when being subjected to the JPEG distortion, and the distortion characteristics that show when being subjected to the Gaussian Blur distortion, under the situation that the original reference image is arranged, realized the judgement of the type of distortion of distorted image, the differentiation process of this type of distortion is portable high, and any situation that needs to judge whether image is subjected to the wherein a kind of distortion in these three kinds of type of distortion can be used.
Embodiment
Describe in further detail below in conjunction with the present invention of accompanying drawing embodiment.
A kind of adaptive image quality method for objectively evaluating that the present invention proposes, its processing procedure is:
At first, determine the type of distortion of distorted image to be evaluated, and original undistorted image and distorted image to be evaluated are divided into a plurality of equitant sizes respectively is 8 * 8 image block;
Secondly, by calculating brightness average and the standard deviation of all pixels in each image block in original undistorted image and the distorted image to be evaluated, and the covariance between all pixels in two image blocks that all coordinate position is identical in original undistorted image and the distorted image to be evaluated, and in conjunction with the type of distortion of distorted image to be evaluated, obtain the structural similarity between two identical image blocks of coordinate positions all in original undistorted image and the distorted image to be evaluated;
At last, according to the structural similarity between two identical image blocks of coordinate positions all in original undistorted image and the distorted image to be evaluated, obtain the objective quality score value of distorted image to be evaluated.
Adaptive image quality method for objectively evaluating of the present invention, it totally realizes block diagram as shown in Figure 1, it specifically may further comprise the steps:
1. make X represent original undistorted image, make Y represent distorted image to be evaluated, determine the type of distortion of Y then by the type of distortion method of discrimination, the type of distortion of Y is wherein a kind of in white Gaussian noise distortion, JPEG distortion, Gaussian Blur distortion, the class JPEG2000 distortion.
At present, the type of distortion of image generally has white Gaussian noise distortion (WN, white noise), JPEG distortion (JPEG), Gaussian Blur distortion (Gblur, Gaussian blur) and class JPEG2000 distortion four classes, wherein, class JPEG2000 distortion comprises two kinds of JPEG2000 distortion and rapid fading distortions (FF, fast fading).At this, by the type of distortion method of discrimination, determine which kind of type of distortion Y is.
In this specific embodiment, as shown in Figure 2, step determines that by the type of distortion method of discrimination detailed process of the type of distortion of Y is in 1.:
1.-and a, X is carried out 2-d wavelet decompose (bior4.4 small echo), obtain the approximate component X of X
A, horizontal component
And vertical component
And calculate
With
The small echo gross energy, be designated as
Then to the approximate component X of X
ACarry out a 2-d wavelet again and decompose (bior4.4 small echo), obtain the approximate component X of X
AThe subband horizontal component
With the subband vertical component
And calculate
With
The small echo gross energy, be designated as
At this, 1≤m'≤M'
1, 1≤n'≤N'
1, M'
1Expression
With
Height, N'
1Expression
With
Width,
Expression
In coordinate position be (m', the coefficient value of n') locating,
Expression
Middle coordinate position is (m', the coefficient value of n') locating, 1≤m''≤M'
2, 1≤n''≤N'
2, M'
2Expression
With
Height, N'
2Expression
With
Width,
Expression
In coordinate position be (m'', the coefficient value of n'') locating,
Expression
Middle coordinate position is (m'', the coefficient value of n'') locating.
Equally, Y is carried out a 2-d wavelet decompose (bior4.4 small echo), obtain the approximate component Y of Y
A, horizontal component
And vertical component
And calculate
With
The small echo gross energy, be designated as
Then to the approximate component Y of Y
ACarry out a 2-d wavelet again and decompose (bior4.4 small echo), obtain the approximate component Y of Y
AThe subband horizontal component
With the subband vertical component
And calculate
With
The small echo gross energy, be designated as
At this, 1≤m'≤M'
1, 1≤n'≤N'
1, M'
1Expression
With
Height, namely
With
Height with
With
The height unanimity, N'
1Expression
With
Width, namely
With
Width with
With
The width unanimity,
Expression
In coordinate position be (m', the coefficient value of n') locating,
Expression
Middle coordinate position is (m', the coefficient value of n') locating, 1≤m''≤M'
2, 1≤n''≤N'
2, M'
2Expression
With
Height, namely
With
Height with
With
The height unanimity, N'
2Expression
With
Width, namely
With
Width with
With
The width unanimity,
Expression
In coordinate position be (m'', the coefficient value of n'') locating,
Expression
Middle coordinate position is (m'', the coefficient value of n'') locating.
Calculate the capacity volume variance of the small echo gross energy of the X correspondence small echo gross energy corresponding with Y again, be designated as Δ W,
1.-b, judgement Δ W<Th
WnWhether set up, if set up, determine that then the type of distortion of Y is the white Gaussian noise distortion, finish then; Otherwise, execution in step 1.-c; Wherein, Th
WnBe white Gaussian noise distortion discrimination threshold, in the present embodiment, white Gaussian noise distortion discrimination threshold Th
WnValue be-0.1.
1.-and c, calculate the luminance difference figure of X, be designated as X
h, with X
hMiddle coordinate position is that (x, the brightness value of pixel y) is designated as X
h(x, y), X
h(x, y)=| X (x, y)-X (x, y+1) |, 1≤x≤H herein, 1≤y≤W-1, wherein, X (x, y) coordinate position is (x, the brightness value of pixel y) among the expression X, (x, y+1) coordinate position is that (symbol " || " is the symbol that takes absolute value for x, the brightness value of pixel y+1) to X among the expression X.
Equally, calculate the luminance difference figure of Y, be designated as Y
h, with Y
hMiddle coordinate position is that (x, the brightness value of pixel y) is designated as Y
h(x, y), Y
h(x, y)=| Y (x, y)-Y (x, y+1) |, wherein, (x, y) coordinate position is that ((x, y+1) coordinate position is (x, the brightness value of pixel y+1) to Y among the expression Y for x, the brightness value of pixel y) to Y among the expression Y.
1.-d, to X
hIn every row in all pixels carry out N
hLeaf transformation in the point discrete Fourier (DFT, Discrete Fourier Transform) obtains X
hIn Fourier's energy spectrum of every row, with X
hIn the capable Fourier's energy spectrum of x be designated as
Wherein,
Symbol
Be the symbol that rounds up, l ∈ [1, N
h/ 2+1], X
f x(l) expression X
hIn x all pixels in capable carry out N
hL the DFT coefficient value that obtains behind the leaf transformation in the point discrete Fourier, symbol " || " is the symbol that takes absolute value.
Equally, to Y
hIn every row in all pixels carry out N
hLeaf transformation in the point discrete Fourier (DFT, Discrete Fourier Transform) obtains Y
hIn Fourier's energy spectrum of every row, with Y
hIn the capable Fourier's energy spectrum of x be designated as
Wherein, Y
f x(l) expression Y
hIn x all pixels in capable carry out N
hL the DFT coefficient value that obtains behind the leaf transformation in the point discrete Fourier.
1.-e, according to X
hIn Fourier's energy spectrum of every row, calculate X
hWhole Fourier's energy spectrum, be designated as P
X,
And according to Y
hIn Fourier's energy spectrum of every row, calculate Y
hWhole Fourier's energy spectrum, be designated as P
Y,
Wherein, l ∈ [1, N
h/ 2+1]; Calculate P then
XWith P
YFourier's energy difference, be designated as Δ P,
1.-f, according to Δ P, the JPEG distortion first judgment threshold Th
Jpeg1With the JPEG distortion second judgment threshold Th
Jpeg2, determine the type of distortion of Y, as Δ P 〉=Th
Jpeg1The time, the type of distortion of determining Y is the JPEG distortion, finishes then; Work as Th
Jpeg2≤ Δ P<Th
Jpeg1The time, the type of distortion of uncertain Y, then execution in step 1.-g; As Δ P<Th
Jpeg2The time, the type of distortion of determining Y is Gaussian Blur distortion or class JPEG2000 distortion, then execution in step 1.-h.
In the present embodiment, the JPEG distortion first judgment threshold Th
Jpeg1Value be 0.1, JPEG distortion, the second judgment threshold Th
Jpeg2Value be-0.2.
1.-and g, calculate Y and compare the percentage that X small echo gross energy reduces, be designated as Δ W',
Judge Δ W'<Th then
Jpeg3Whether set up, if set up, determine that then the type of distortion of Y is the JPEG distortion, finish then, otherwise, execution in step 1.-h, wherein, Th
Jpeg3Be JPEG distortion the 3rd judgment threshold, in the present embodiment, JPEG distortion the 3rd judgment threshold Th
Jpeg3Value be 0.075.
1.-h, adopt the Sobel operator respectively R, G, the B Color Channel subgraph of X to be carried out edge extracting, find out the pixel of brightness value maximum in the edge of each Color Channel subgraph of X then, the brightness value of the pixel of brightness value maximum in the edge of the c Color Channel subgraph of X is designated as
, wherein, c=R, G, B.
Equally, adopt the Sobel operator respectively R, G, the B Color Channel subgraph of Y to be carried out edge extracting, find out the pixel of brightness value maximum in the edge of each Color Channel subgraph of Y then, the brightness value of the pixel of brightness value maximum in the edge of the c Color Channel subgraph of Y is designated as
Judge again
With
Whether all set up, if all set up, then execution in step 1.-i; Otherwise the type of distortion of determining Y is class JPEG2000 distortion, finishes then.
1.-i, with X from the RGB color space conversion to HSI(Hue Saturation Intensity) color space, extract the chromatic component of X then, be designated as X
H, then calculate X
HThe average chrominance value, be designated as
Wherein, X
H(m, n) expression X
HMiddle coordinate position is at (m, the chromatic value of n) locating;
Equally, with Y from the RGB color space conversion to the HIS color space, extract the chromatic component of Y then, be designated as Y
H, then calculate Y
HThe average chrominance value, be designated as
Wherein, Y
H(m, n) expression Y
HMiddle coordinate position is at (m, the chromatic value of n) locating.
Afterwards, calculate the colourity relative different of X and Y, be designated as C,
Judge again whether C<0.3 sets up, if set up, then execution in step 1.-j; Otherwise the type of distortion of determining Y is class JPEG2000 distortion, finishes then.
1.-and j, the R to X, G, B Color Channel subgraph carry out two dimensional discrete Fourier transform respectively, obtains the amplitude spectrum of the R Color Channel subgraph of X, the amplitude spectrum of G Color Channel subgraph and the amplitude spectrum of B Color Channel subgraph, and correspondence is designated as respectively
With
Equally, R, G, the B Color Channel subgraph of Y carried out two dimensional discrete Fourier transform respectively, obtain the amplitude spectrum of the R Color Channel subgraph of Y, the amplitude spectrum of G Color Channel subgraph and the amplitude spectrum of B Color Channel subgraph, correspondence is designated as respectively
Then, calculate the frequency response of each Color Channel subgraph, the frequency response of c Color Channel subgraph is designated as
Will
In coordinate position in that (m, the DFT coefficient value of n) locating is designated as
Wherein,
Expression
In coordinate position (m, the DFT coefficient value of n) locating,
Expression
In coordinate position (m, the DFT coefficient value of n) locating, 1≤m≤H, 1≤n≤W, H is
With
Height, namely
With
Height consistent with the height of X and Y, W is
With
Width, namely
With
Width consistent with the width of X and Y.
Then, extract the vertical zero-frequency component of the frequency response of each Color Channel subgraph, with the frequency response of c Color Channel subgraph
Vertical zero-frequency component be designated as
Will
In m DFT coefficient value be designated as
Wherein,
Expression
The DFT coefficient value that middle coordinate position is located in (m, 1).
Afterwards, in the middle of the center zero-frequency of the vertical zero-frequency component of the frequency response of each Color Channel subgraph moved on to, obtain new vertical zero-frequency component, the new vertical zero-frequency component of the frequency response of c Color Channel subgraph is designated as
Will
In m DFT coefficient value be designated as
Wherein, symbol
For rounding symbol downwards.
Extract again decentre point in the new vertical zero-frequency component of frequency response of each Color Channel subgraph ±
The Frequency point of distance, to use window width be 3 filter carries out medium filtering to the Frequency point of each Color Channel subgraph correspondence of extracting, with filtered Frequency point and the match of one dimension Gaussian function, the Frequency point after the calculation of filtered and the Pearson correlation coefficient of the data after the match; For
Extract
The decentre point ±
The Frequency point of distance is designated as
Using window width is that 3 filter is right
Carry out medium filtering, obtain filtered Frequency point, be designated as
Will
With the match of one dimension Gaussian function, obtain the data after the match, be designated as
Calculate
With
Pearson correlation coefficient, be designated as
Wherein,
For
In the DFT coefficient value of d,
For
In the DFT coefficient value after match of d.
At last, judge Pe
R≤ 0.9, Pe
G≤ 0.9 and Pe
BWhether≤0.9 all set up, if all set up, determines that then the type of distortion of Y is class JPEG2000 distortion, finishes then; Otherwise the type of distortion of determining Y is the Gaussian Blur distortion, finishes then.
In the present embodiment, 808 width of cloth images that the view data of using provides as U.S.'s Texas university image and the disclosed picture quality estimation database in video engineering experiment chamber (LIVE) are comprising undistorted reference picture 29 width of cloth, distorted image 779 width of cloth.In addition, this 779 width of cloth distorted image is assigned in the 5 number of sub images storehouses by type of distortion, that is: white Gaussian noise (WN, white noise) distorted image storehouse (comprising 145 width of cloth images), Gaussian Blur (Gblur, Gaussian blurring) distorted image storehouse (comprising 145 width of cloth images), JPEG distorted image storehouse (comprising 175 width of cloth images), JPEG2000 distorted image storehouse (comprising 169 width of cloth images) and rapid fading (FF, fast fading) distorted image storehouse (comprising 145 width of cloth images).Simultaneously, the type of distortion of these distorted images is single.
Type of distortion according to distorted image shown in Figure 2 is differentiated flow process, and the first step is isolated the distorted image of white Gaussian noise distortion from all distorted images of database.Whether the type of distortion at definite distorted image Y is the white Gaussian noise distortion discrimination threshold Th that relates in the process of white Gaussian noise distortion
WnThe time, in interval [0.5,0.5], get a point as Th every 0.08
WnProbe value, to each probe value, carry out 200 groups of discriminating step 1.-a and step 1.-the data training of b, and ask the accuracy rate of judgement respectively, wherein every group of data are from the image of half quantity of selecting at random in each distortion subimage storehouse, Th
WnThe relation curve of size and differentiation accuracy is worked as Th as can be seen from Fig. 3 a shown in Fig. 3 a
Wn=-0.1 o'clock can be that the white Gaussian noise distorted image is isolated on 100% ground with accuracy.
Second step was isolated the JPEG distorted image from the image of non-white Gaussian noise distortion.In between interval [1,1], get a bit as Th every 0.08
JpegProbe value, equally to each probe value do 200 groups of discriminating step 1.-c to step 1.-the data training of f, experimental image is from the subimage storehouse except the white Gaussian noise database, quantity is half of the total picture number of each image word bank, selection mode is selection at random.Obtain Th
JpegThe relation curve of size and correct judgment rate as can be seen, does not have the Th that is fit to from Fig. 3 b shown in Fig. 3 b
JpegCan accomplish the right-on JPEG of isolating distorted image, therefore, according to discriminating step 1.-f is described, gets Th
Jpeg1=0.1, Th
Jpeg2=-0.2, for Δ P ∈ [0.2,0.1) image, carry out in the discriminating step 1.-the data training of g, the experimental image source is identical with the former.Experiment obtains power difference that reference picture is original undistorted image X and distorted image Y and wavelet sub-band energy difference relation curve shown in Fig. 3 c, as can be seen, gets Th from Fig. 3 c
Jpeg3Whether=0.075 o'clock, can 100% correctly judging distorted image Y is the JPEG distortion.
2. adopting size is that 8 * 8 the sliding window order by Row Column in X moves by pixel, and it is 8 * 8 image block that X is divided into M * N equitant and size, is that (i, image block j) is designated as x with coordinate position among the X
I, jEqually, adopting size is that sliding window order by Row Column in Y of 8 * 8 moves by pixel, and it is 8 * 8 image block that Y is divided into M * N equitant and size, is that (i, image block j) is designated as y with coordinate position among the Y
i,
jWherein,
H represents the height of X and Y, and W represents the width of X and Y, symbol
For rounding symbol downwards, 1≤i≤M, 1≤j≤N.
In the image block cutting procedure, if boundary less than 8 * 8 sizes of X and Y, then this partial pixel point is not cut apart.
3. calculate brightness average and the standard deviation of all pixels in each image block among the X, and brightness average and the standard deviation of all pixels in each image block among the calculating Y, calculate the covariance between all pixels in two identical image blocks of coordinate positions all among X and the Y then, be (i, image block x j) with coordinate position among the X
I, jIn brightness average and the standard deviation correspondence of all pixels be designated as
With
Be (i, image block y j) with coordinate position among the Y
I, jIn brightness average and the standard deviation correspondence of all pixels be designated as
With
Be (i, image block x j) with coordinate position among the X
I, jIn all pixels and Y in coordinate position be (i, image block y j)
I, jIn all pixels between covariance be designated as
Wherein, x
I, j(u v) represents x
I, jMiddle coordinate position is (u, the brightness value of pixel v), y
I, j(u v) represents y
I, jMiddle coordinate position is (u, the brightness value of pixel v), 1≤u≤8,1≤v≤8.
4. calculating luminance function, contrast function and degree of structuration function between two identical image blocks of coordinate positions all among X and the Y, is (i, image block x j) with coordinate position among the X
I, jBe (i, image block y j) with coordinate position among the Y
I, jBetween luminance function, contrast function and degree of structuration function correspondence be designated as l (x respectively
I, j, y
I, j), c (x
I, j, y
I, j) and s (x
I, j, y
I, j),
Wherein, C
1, C
2, C
3For avoiding denominator the zero little numerical constant that arranges to occur, get C in the present embodiment
1=0.01, C
2=0.02, C
3=0.01.
5. according to luminance function, contrast function and degree of structuration function between two identical image blocks of coordinate positions all among X and the Y, calculate the structural similarity between two identical image blocks of coordinate positions all among X and the Y, be (i, image block x j) with coordinate position among the X
I, jBe (i, image block y j) with coordinate position among the Y
I, jBetween structural similarity be designated as SSIM (x
I, j,y
I, j),
Wherein, t is used for the type of distortion of expression Y, gets t=1 when the type of distortion of Y is the white Gaussian noise distortion, gets t=2 when the type of distortion of Y is the JPEG distortion, when the type of distortion of Y is the Gaussian Blur distortion, get t=3, when the type of distortion of Y is class JPEG2000 distortion, get t=4; α
tThe weight factor of brightness, β are regulated in expression
tThe weight factor of contrast, γ are regulated in expression
tThe weight factor of expression adjustment structure degree, and satisfy α
t+ β
t+ γ
t=3, its concrete size is distributed relevant with the type of distortion of distorted image.
6. according to the structural similarity between two identical image blocks of coordinate positions all among X and the Y, calculate the objective quality score value of Y, be designated as Q,
In the present embodiment, in order to determine α
t, β
t, γ
t(t ∈ { 1,2,3,4}, corresponding different type of distortion) 29 width of cloth undistorted images and four kinds of type of distortion (white Gaussian noise distortion, corresponding white Gaussian noise distorted image storehouses of providing in the LIVE storehouse are provided the best value when the distorted image of different type of distortion is estimated; The Gaussian Blur distortion, corresponding Gaussian Blur distorted image storehouse; The JPEG distortion, corresponding JPEG distorted image storehouse; Class JPEG2000 distortion, corresponding JPEG2000 distortion and rapid fading distorted image storehouse) corresponding subimage storehouse view data, and the corresponding average subjective scoring difference (DMOS of each width of cloth distorted image, difference mean opinion scores) (wherein the quality of the more big expression distorted image of DMOS value is more poor, the quality of the more little expression distorted image of DMOS value is more good, and the span of DMOS is [0,100]), carry out parameter optimization.
Optimizing process is: to each class distortion, the first step is set α
Tem∈ [0,3], and 1 being step-length, β
Temp∈ [0,3-α
Temp], and 0.1 to be step-length, γ simultaneously
Temp=3-α
Temp-β
Temp, 6. 1. the above-mentioned four kinds of corresponding subimage of type of distortion storehouse view data calculated the corresponding quality evaluation score value of every width of cloth distorted image Q to step by the step of the inventive method; Adopt four parameter L ogistic functions to carry out nonlinear fitting then, obtain quality evaluation score value after the match and the Pearson correlation coefficient (CC between the DMOS value, Correlation Coefficient), Spearman coefficient correlation (SROCC, Spearman Rank-Order Correlation Coefficient), with calculating quality evaluation score value and the CC coefficient between the DMOS value and the SROCC coefficient of t class distorted image, be designated as CC respectively
tAnd SROCC
t, with T
t=max (CC
t+ SROCC
t) be target function, wherein { max () function is for getting max function for 1,2,3,4}, corresponding different type of distortion for t ∈; Obtain the maximum of every class distortion, be designated as
, then note and work as T
tObtain maximum
The time
,
Value, as work as T
1Get maximum
The time
,
Value.
And 0.2 being step-length, β
Temp∈ [0,3-α
Temp], and 0.1 being step-length, γ
Temp=3-α
Temp-β
Temp6. 1. the above-mentioned four kinds of corresponding subimage of type of distortion storehouse view data calculated the corresponding quality evaluation score value of every width of cloth distorted image Q to step by the inventive method step, adopt four parameter L ogistic functions to carry out nonlinear fitting then, quality evaluation score value after the calculating match and coefficient correlation CC, the SROCC between the DMOS value, with calculating quality evaluation score value and the CC coefficient between the DMOS value and the SROCC coefficient of t class distorted image, be designated as CC respectively
tAnd SROCC
t, with T
t=max (CC
t+ SROCC
t) be target function; Obtain the maximum of every class distortion, be designated as
, then note and work as T
tObtain maximum
The time
,
Value.
β
Temp∈ [0,3-α
Temp], be step-length with 0.1 all, γ
Temp=3-α
Temp-β
Temp6. 1. the above-mentioned four kinds of corresponding subimage of type of distortion storehouse view data calculated the corresponding quality evaluation score value of every width of cloth distorted image Q to step by the inventive method step, adopt four parameter L ogistic functions to carry out nonlinear fitting then, quality evaluation score value after the calculating match and coefficient correlation CC, the SROCC between the DMOS value, with calculating quality evaluation score value and the CC coefficient between the DMOS value and the SROCC coefficient of t class distorted image, be designated as CC respectively
tAnd SROCC
t, with T
t=max (CC
t+ SROCC
t) be target function; Obtain the maximum of every class distortion, be designated as
, then note and work as T
tObtain maximum
The time
,
Value.
In the 4th step, determine that finally the weight factor under the different type of distortion is assigned as:
,
, γ
t=3-α
t-β
t, t ∈ { 1,2,3,4}, corresponding different type of distortion wherein.
In the present embodiment, obtain α by experiment
t, β
tAnd γ
tValue such as table 1 listed.
α under the different type of distortion of table 1
t, β
tAnd γ
tValue
At this, 29 width of cloth undistorted images that provide in the LIVE storehouse and 779 single distorted images and the corresponding DMOS value of each width of cloth distorted image are provided, 1. 6. calculate the quality evaluation score value Q of each width of cloth distorted image to step according to step, quality evaluation score value Q and the DMOS value of 779 width of cloth distorted images are carried out four parameter L ogistic function nonlinear fittings; Utilize 4 objective parameters commonly used of evaluate image quality evaluating method as evaluation index, be coefficient correlation (CC, SROCC), dispersion ratio (OR, out ratio) and mean square error coefficient (RMSE, rooted mean squared error), wherein CC coefficient and SROCC coefficient calculations method are shown in the parameter optimisation procedure first step, and the OR value calculating method is:
Z wherein
0Satisfy expression formula in the expression sample to be tested
Element number, wherein S represents reference sample, { S
z, z=1,2 ..., Z}; O represents sample to be tested, { O
z, z=1,2 ..., Z}, Z are the total number S of element in the sample.The RMSE value calculating method is:
S wherein
z, O
z, the Z meaning is with last identical.The more high key diagram of CC and SROCC value is more good as method for objectively evaluating and DMOS correlation, and the more low key diagram of OR value and RMSE value is more good as method for objectively evaluating and DMOS correlation.
Table 2 has been listed the value of CC, SROCC, OR and the RMSE coefficient of assess performance under the various type of distortion, from the listed data of table 2 as seen, objective quality score value Q and the correlation between the subjective scores DMOS of the distorted image that present embodiment obtains are very high, CC value and SROCC value all surpass 0.91, the OR value is lower than 0.5, the RMSE value is lower than 5.5, and this result who has shown the objective evaluation result of the inventive method and human eye subjective perception is more consistent, has proved absolutely the validity of the inventive method.
Correlation between the objective evaluation branch of the distorted image that this enforcement of table 2 obtains and subjective assessment divide