CN106709958A - Gray scale gradient and color histogram-based image quality evaluation method - Google Patents

Abstract

The invention discloses a gray scale gradient and color histogram-based image quality evaluation method. The method comprises the following steps of: 1, constructing a group of two-dimensional Sobel detection operators, and carrying out convolutional processing on an input reference image and a distorted image so as to obtain gradient feature information of the reference image and the to-be-detected distorted image; 2, switching the reference image and the distorted image from an RGB space to an HSV space, and solving color histogram feature information of the images; 3, respectively calculating a gray scale gradient similarity and a color histogram similarity between the reference image and the distorted image; and 4, inputting the gray scale gradient similarity and the color histogram similarity, and carrying out quality mapping and measurement by utilizing a machine learning method, so as to obtain objective evaluation values about the image quality. According to the method, the gray scale gradient feature information and the color histogram feature information are efficiently extracted, so that the calculation complexity is low and the operation is fast; and the objective image quality evaluation method on the basis of the two types of feature information is consistent with that of subjective evaluation.

Description

A kind of image quality evaluating method based on shade of gray and color histogram

Technical field

The invention belongs to technical field of image processing, more particularly to a kind of image based on shade of gray and color histogram Quality evaluating method.

Background technology

Image quality evaluation is for evaluating during obtaining, compressing, storing, transmitting and rebuilding because distortion is introduced Caused image quality loss.Picture quality is to weigh an important indicator of image processing system performance, therefore image matter Amount is evaluated most important for evaluating and optimizing video image processing system, it has also become the basis of digital image processing techniques research And key, with great theory and construction value, receive increasing attention.

Image quality evaluation is divided into subjective assessment and objective evaluation.Subjective assessment is that picture quality is commented by observer Point.The method is the most reliable, but due to its intrinsic defect, such as wastes time and energy, poor operability, it is difficult to as engineering survey hand Section directly applies to the measurement of picture quality, is more not suitable for the application of real time processing system.And picture quality objective evaluation side Method is automatically scored picture quality by design mathematic model according to certain yardstick, with simple, repeatable and meter The advantages of calculating efficiency high, the study hotspot as image quality evaluation.

Due at this stage to the cognitive insufficient of human visual system (HVS), based on existing human vision Physiological Psychology Etc. achievement in research there is no method to build evaluation and prediction of the perfect computation model realization to picture quality.Therefore, it is near several The research of year image quality evaluating method focuses more on the physical significance of image, the i.e. various perceptual properties of image, by scheming The intellectual analysis of picture, extract the characteristic information related to quality, and measure and compare testing image with original image in statistics Difference in meaning, realizes the quality mapping and evaluation to testing image.

The content of the invention

The purpose of the present invention is directed to during the full reference image quality appraisement of existing natural image statistical property, due to To image characteristics extraction not enough efficiently caused by the defect such as quality evaluation performance is relatively low, computation complexity is bigger than normal, propose a kind of Image quality evaluating method based on shade of gray and color histogram.

The present invention is adopted the technical scheme that：

Extract the shade of gray characteristic information and color histogram characteristic information of reference picture and distorted image respectively first, It is then similar between similarity and color histogram feature respectively between acquisition reference picture and distorted image Gradient Features Degree, as the benchmark of image quality evaluation, and then, the quality evaluation of distorted image to be measured is obtained using the method for machine learning Value.

The technical solution adopted for the present invention to solve the technical problems is as follows：

Step (1) input reference pictures I_RWith distorted image I to be measured_D；

Step (2) sets up one group of two dimension Sobel detective operators S_xAnd S_y：

The reference picture I that step (3) is input into using the two-dimentional Sobel detective operators that step (2) is set up to step (1)_R With distorted image I_DConvolutional calculation is carried out, reference picture I is respectively obtained_RWith distorted image I_DIn the gradient information of two-dimensional space；Note G_RAnd G_DRespectively reference picture I_RWith distorted image I_DGradient information：

Wherein, i represents the position of pixel,Represent convolution algorithm.

Step (4) is using gradient information G obtained by step (3)_RAnd G_D, calculate reference picture I_RIn each pixel with it is to be measured Distorted image I_DThe similarity (being designated as GS) of gradient between middle correspondence position pixel：

Wherein, i represents the position of pixel, and c represents the constant value of setting.

Further, view picture reference picture I is calculated_RWith distorted image I to be measured_DBetween gradient similarity：

Wherein, N represents the sum of pixel in image,

The reference picture I that be input into for step (1) respectively by step (5)_RWith distorted image I_DHSV is transformed into by rgb space empty Between, conversion formula is：

Wherein, R represents the red component of image, and G represents the green component of image, and B represents the blue component of image, H tables Value of the diagram picture on HSV space H passages, S represents value of the image in HSV space channel S, and V represents image in HSV space V Value on passage.

Step (6) is in HSV space, the reference picture I that statistical computation step (5) is obtained respectively_RWith distorted image I_DIt is straight Fang Tu.It is as follows in the calculation of the color histogram of HSV space image：

First, according to color different range and subjective color perceives the reference picture I obtained to step (5)_RWith it is to be measured Distorted image I_DValue on H passages, channel S and V passages is quantified：

Then, structuring one-dimensional characteristic vector, according to the quantized level that formula (9), (10), (11) obtain, each color component Synthesize one-dimensional characteristic vector：

HS=HQ_SQ_V+SQ_V+V (12)

Wherein, HS is the one-dimensional characteristic vector of synthesis, Q_SAnd Q_VIt is respectively the quantization series of component S and V.

Finally, calculating is normalized to the one-dimensional characteristic vector HS for obtaining, obtains histogram of the image in HSV space HSH:

Wherein, HS (i) represents values of the one-dimensional characteristic vector HS at i, and M represents the length of characteristic vector HS.

Step (7) obtains reference picture I according to the histogram that step (6) is obtained_RHistogram HSH_RWith distorted image I_D Histogram HSH_DBetween similarity, note HSHD be reference picture I_RWith distorted image I_DSimilarity between histogram：

Step (8) in known image quality evaluation database, using method (such as SVMs, the god of machine learning Through methods such as networks) the histogram similarity HSHD synthesis that obtains of the gradient similarity GSD that obtains step (4) and step (7) And image fault measurement is mapped to, obtain the quality evaluation value of image.

Q=F_ML(GSD,HSHD|DMOS) (15)

Wherein, Q is the objective evaluation mass fraction of distorted image, and Q is bigger, illustrates that picture quality is higher.F_MLIt is engineering Learning method, DMOS is the corresponding subjective assessment fraction of each image in image quality evaluation database.

Beneficial effects of the present invention：

The present invention realizes the extraction of reference picture and distorted image characteristic information using gradient and color histogram, and passes through The method of machine learning carries out synthesis and quality mapping to the image feature information for being extracted, so as to obtain distorted image to be measured Quality evaluation.Test result indicate that, the picture quality objective evaluation based on method proposed by the invention has very with subjective assessment Good uniformity.

Brief description of the drawings

Fig. 1 is the structured flowchart of image quality evaluating method of the present invention based on gradient and color histogram.

Specific embodiment

The inventive method is described further below in conjunction with the accompanying drawings.

As shown in figure 1, the image quality evaluating method based on gradient and color histogram, its specific implementation step is as follows：

Step (1) is programmed under Matlab environment, circulation read in well known data storehouse (LIVE, CSIQ, TID2008 and TID2013 etc.) in reference picture I_RWith distorted image I_D；

Step (2) sets up one group of two dimension Sobel detective operators S_xAnd S_y：

The reference picture I that step (3) is input into using the two-dimentional Sobel detective operators that step (2) is set up to step (1)_R With distorted image I_DConvolutional calculation is carried out, reference picture I is respectively obtained_RWith distorted image I_DIn the gradient information of two-dimensional space；Note G_RAnd G_DRespectively reference picture I_RWith distorted image I_DGradient information：

Wherein, i represents the position of pixel,Represent convolution algorithm.

Step (4) is using gradient information G obtained by step (3)_RAnd G_D, calculate reference picture I_RIn each pixel with it is to be measured Distorted image I_DThe similarity (being designated as GS) of gradient between middle correspondence position pixel：

Wherein, i represents the position of pixel, and c represents the constant value of setting, and in the present embodiment, the constant value that c takes is 150。

Further, view picture reference picture I is calculated_RWith distorted image I to be measured_DBetween gradient similarity：

Wherein, N represents the sum of pixel in image,

The reference picture I that be input into for step (1) respectively by step (5)_RWith distorted image I_DHSV is transformed into by rgb space empty Between, conversion formula is：

Wherein, R represents the red component of image, and G represents the green component of image, and B represents the blue component of image, H tables Value of the diagram picture on HSV space H passages, S represents value of the image in HSV space channel S, and V represents image in HSV space V Value on passage.

Step (6) is in HSV space, the reference picture I that statistical computation step (5) is obtained respectively_RWith distorted image I_DIt is straight Fang Tu.It is as follows in the calculation of the color histogram of HSV space image：

First, according to color different range and subjective color perceives the reference picture I obtained to step (5)_RWith it is to be measured Distorted image I_DValue on H passages, channel S and V passages is quantified：

Then, structuring one-dimensional characteristic vector, according to the quantized level that formula (9), (10), (11) obtain, each color component Synthesize one-dimensional characteristic vector：

HS=HQ_SQ_V+SQ_V+V (12)

Wherein, HS is the one-dimensional characteristic vector of synthesis, Q_SAnd Q_VIt is respectively the quantization series of component S and V.In the present embodiment In, Q_SValue be taken as 4, Q_VValue be taken as 4.

Finally, calculating is normalized to the one-dimensional characteristic vector HS for obtaining, obtains histogram of the image in HSV space HSH:

Wherein, HS (i) represents values of the one-dimensional characteristic vector HS at i, and M represents the length of characteristic vector HS.

Step (7) obtains reference picture I according to the histogram that step (6) is obtained_RHistogram HSH_RWith distorted image I_D Histogram HSH_DBetween similarity, note HSHD be reference picture I_RWith distorted image I_DSimilarity between histogram：

Step (8) in known image quality evaluation database, using method (such as SVMs, the god of machine learning Through methods such as networks) the histogram similarity HSHD synthesis that obtains of the gradient similarity GSD that obtains step (4) and step (7) And image fault measurement is mapped to, obtain the quality evaluation value of image.

Q=F_ML(GSD,HSHD|DMOS) (15)

Wherein, Q is the objective evaluation mass fraction of distorted image, and Q is bigger, illustrates that picture quality is higher.F_MLIt is engineering Learning method, in the present embodiment, calls the machine learning function of lib-SVM, and data work will be obtained in step (4) and step (7) It is the input of the function.DMOS is the corresponding subjective assessment fraction of each image in image quality evaluation database.

Claims (4)

1. a kind of image quality evaluating method based on shade of gray and color histogram, it is characterised in that comprise the following steps：

1) one group of two dimension Sobel detective operators is built, and the reference picture and distorted image that are input into are carried out at convolution using it Reason, obtains the Gradient Features information of reference picture and distorted image to be measured respectively；Calculate between reference picture and distorted image Shade of gray similarity；

2) reference picture and distorted image are transformed into HSV space by rgb space, the color histogram of image is asked in HSV space Figure characteristic information；Calculate the color histogram similarity between reference picture and distorted image；

3) it is input with shade of gray similarity and color histogram similarity, realizes that quality maps using the method for machine learning And measurement, obtain the objective evaluation value of picture quality.

2. the image quality evaluating method based on shade of gray and color histogram according to claim 1, its feature exists In described step 1) specifically include following steps：

Step (1) input reference pictures I_RWith distorted image I to be measured_D；

Step (2) sets up one group of two dimension Sobel detective operators S_xAnd S_y：

$S_x=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right];S_y=\left[\begin{array}{ccc}1& 2& 1\\ 0& 0& 0\\ -1& -2& -1\end{array}\right]---\left(1\right)$

The reference picture I that step (3) is input into using the two-dimentional Sobel detective operators that step (2) is set up to step (1)_RAnd mistake True image I_DConvolutional calculation is carried out, reference picture I is respectively obtained_RWith distorted image I_DIn the gradient information of two-dimensional space；Note G_RWith G_DRespectively reference picture I_RWith distorted image I_DGradient information：

$G_R\left(i\right)=\sqrt{{\left(I_R\right(i)\⊗S_x(i\left)\right)}^2+{\left(I_R\right(i)\⊗S_y(i\left)\right)}^2}---\left(2\right)$

$G_D\left(i\right)=\sqrt{{\left(I_D\right(i)\⊗S_x(i\left)\right)}^2+{\left(I_D\right(i)\⊗S_y(i\left)\right)}^2}---\left(3\right)$

Wherein, i represents the position of pixel,Represent convolution algorithm；

Step (4) is using gradient information G obtained by step (3)_RAnd G_D, calculate reference picture I_RIn each pixel and distortion to be measured Image I_DThe similarity of gradient, is designated as GS between middle correspondence position pixel：

$GS\left(i\right)=\frac{2G_R\left(i\right)G_D\left(i\right)+c}{G_R^2\left(i\right)+G_D^2\left(i\right)+c}---\left(4\right)$

Wherein, i represents the position of pixel, and c represents the constant value of setting.

Further, view picture reference picture I is calculated_RWith distorted image I to be measured_DBetween gradient similarity：

$GSD=\sqrt{\frac{1}{N}{\mathrm{\Σ}}_{i=1}^N{\left(GS\right(i)-GSM)}^2}---\left(5\right)$

Wherein, N represents the sum of pixel in image,

3. the image quality evaluating method based on shade of gray and color histogram according to claim 1, its feature exists In described step 2) specifically include following steps：

Step (5) is respectively by reference picture I_RWith distorted image I_DHSV space is transformed into by rgb space, conversion formula is：

$H={\cos }^{-1}\left\{\frac{1/2\[(R-G)+(R-B)\]}{\sqrt{{(R-G)}^2+(R-B)(G-B)}}\right\}---\left(6\right)$

$S=1-\frac{3}{R+G+B}\[min(R,G,B)\]---\left(7\right)$

$V=\frac{1}{3}(R+G+B)---\left(8\right)$

Wherein, R represents the red component of image, and G represents the green component of image, and B represents the blue component of image, and H represents figure As the value on HSV space H passages, S represent value of the image in HSV space channel S, V represents image in HSV space V passages On value.

Step (6) is in HSV space, the reference picture I that statistical computation step (5) is obtained respectively_RWith distorted image I_DHistogram； It is as follows in the calculation of the color histogram of HSV space image：

First, according to color different range and subjective color perceives the reference picture I obtained to step (5)_RWith distortion map to be measured As I_DValue on H passages, channel S and V passages is quantified：

$H=\left\{\begin{array}{cc}0,& H\∈(345,15\]\\ 1,& H\∈(15,25\]\\ 2,& H\∈(25,45\]\\ 3,& H\∈(45,55\]\\ 4,& H\∈(55,80\]\\ 5,& H\∈(80,108\]\\ 6,& H\∈(108,140\]\\ 7,& H\∈(140,165\]\\ 8,& H\∈(165,190\]\\ 9,& H\∈(190,220\]\\ 10,& H\∈(220,255\]\\ 11,& H\∈(255,275\]\\ 12,& H\∈(275,290\]\\ 13,& H\∈(290,316\]\\ 14,& H\∈(316,330\]\\ 15,& H\∈(330,345\]\end{array}\right.---\left(9\right)$

$S=\left\{\begin{array}{cc}0,& S\∈(0,0.15\]\\ 1,& S\∈(0.15,0.4\]\\ 2,& S\∈(0.4,0.75\]\\ 3,& S\∈(0.75,1\]\end{array}\right.---\left(10\right)$

$V=\left\{\begin{array}{cc}0,& V\∈(0,0.15\]\\ 1,& V\∈(0.15,0.4\]\\ 2,& V\∈(0.4,0.75\]\\ 3,& V\∈(0.75,1\]\end{array}\right.---\left(11\right)$

Then, structuring one-dimensional characteristic vector, according to the quantized level that formula (9), (10), (11) obtain, the synthesis of each color component It is one-dimensional characteristic vector：

HS=HQ_SQ_V+SQ_V+V (12)

Wherein, HS is the one-dimensional characteristic vector of synthesis, Q_SAnd Q_VIt is respectively the quantization series of component S and V；

Finally, calculating is normalized to the one-dimensional characteristic vector HS for obtaining, obtains histogram HSH of the image in HSV space:

$HSH=\frac{1}{M}{\mathrm{\Σ}}_{i=1}^{M}HS\left(i\right)---\left(13\right)$

Wherein, HS (i) represents values of the one-dimensional characteristic vector HS at i, and M represents the length of characteristic vector HS.

Step (7) obtains reference picture I according to the histogram that step (6) is obtained_RHistogram HSH_RWith distorted image I_DIt is straight Scheme HSH in side_DBetween similarity, note HSHD be reference picture I_RWith distorted image I_DSimilarity between histogram：

$HSHD=\frac{\underset{i=1}{\overset{M}{\Σ}}min\[{\mathrm{HSH}}_R\left(i\right),{\mathrm{HSH}}_D\left(i\right)\]}{\underset{i=1}{\overset{M}{\Σ}}{\mathrm{HSH}}_R\left(i\right)}---\left(14\right).$

4. the image quality evaluating method based on shade of gray and color histogram according to claim 1, its feature exists In described step 3) it is specially：

Step (8) in known image quality evaluation database, the gradient similarity that will be obtained using the method for machine learning GSD and histogram similarity HSHD synthesis are simultaneously mapped to image fault measurement, obtain the quality evaluation value of image：

Q=F_ML(GSD,HSHD|DMOS) (15)

Wherein, Q is the objective evaluation mass fraction of distorted image, and Q is bigger, illustrates that picture quality is higher；F_MLFor machine learning side Method, DMOS is the corresponding subjective assessment fraction of each image in image quality evaluation database.