CN104899893A - Image quality detection method based on vision attention - Google Patents

Abstract

The invention relates to an image quality detection method based on vision attention. A vision attention distribution model for simulating a human being vision system on an image is established by virtue of the image structural information analysis adopting variance as a vision decision factor on the basis of a vision information fidelity (VIF) method. The method comprises the following steps: firstly carrying out the variance calculation for the structural information complexity in different areas of the image, carrying out the weighted calculation for the VIF method according to the structural information complexity variance of different image areas, and finally evaluating the image quality by comparing with a reference image, so that the subjective quality evaluation of the simulation human being vision system on the image is realized. By adopting the image quality detection method, a result consistent with the subjective evaluation of the human beings can be obtained.

Description

The picture quality detection method of view-based access control model notice

Technical field

The present invention relates to a kind of picture quality detection method, especially a kind of picture quality detection method of view-based access control model notice, its can use computation model as much as possible simulating human subjective consciousness carry out the assessment of picture quality.

Background technology

At present, the method for picture quality being carried out to objective evaluation has multiple, comprises PSNR, SSIM, VIF, SEDLAI and IGM etc.But all there is the defect do not conformed to the subjective evaluation result of human visual system's (i.e. Human Visual System is called for short HVS) in these existing methods.

With VIF (i.e. visual information fidelity, Visual Information Fidelity are called for short VIF), method is example.

VIF is applied to natural sense image quality evaluation at first, depends on natural scene statistical model, picture signal distortion passage and human eye vision distortion model.VIF evaluation model schematic diagram as shown in Figure 1, VIF evaluation model need meet following hypothesis: the statistical distribution of (1) image source meets GSM (i.e. Gaussian Scale Mixture is called for short GSM) model; (2) image wavelet domain coefficient is uncorrelated, and each wavelet sub-band is separate; (3) human eye vision distorted interpretation is the noisiness of intrinsic nerve unit; (4) input signal (reference picture coefficient) and output signal (distorted image coefficient) all known.

Namely suppose in VIF evaluation model that input picture, image fault passage, distorted image model are all accurate, utilize I (C; And I (C E|z); F|z) information extracted from input picture and distorted image particular sub-band that human eye can be desirable is represented respectively.I (C; E|z) input image information content is interpreted as, I (C; F|z) association relationship of input picture and distorted image is interpreted as.This value is visually extracted from distorted image relative in the information content occurred input picture, and because each wavelet sub-band is separate, then the ratio of two kinds of information measurements can expand to multiple subband, with be that the corresponding mutual information of a kth subband is measured respectively, wherein k is sub band number, and VIF index can be expressed as:

$VIF=\frac{\underset{k=1}{\overset{K}{\Σ}}\[I(C_r^k;F^k|z_r^k)\]}{\underset{k=1}{\overset{K}{\Σ}}\[I(C_r^k;E^k|z_r^k)\]}.$

Image information quality testing object uses computation model to simulate the assessment that subjective consciousness carries out picture quality as much as possible.And existing image information quality determining method still can not obtain and evaluates and tests result the most consistent with human subject.

Summary of the invention

In order to overcome the above-mentioned defect of prior art, the object of the invention is the picture quality detection method providing a kind of view-based access control model notice, and it can obtain evaluates and tests more consistent result with human subject.

In order to achieve the above object, main technical schemes provided by the invention comprises:

A kind of picture quality detection method of view-based access control model notice, it is mainly on the basis of visual information fidelity (VIF) method, in conjunction with the image structure information analysis taking variance as vision resolution factor, set up the visual attention distributed model of simulating human vision system to image.

The picture quality detection method of one embodiment of the invention, its key step comprises:

S1, variance calculating is carried out to the structural information complexity in image zones of different;

The structural information complexity variance in S3, foundation different images region is to VIF method weighted calculation;

S4, final contrast reference picture evaluate image quality, to realize the subjective quality assessment of simulating human vision system to image.

The picture quality detection method of one embodiment of the invention, wherein, step S1 specifically: to fragmental image processing, and calculate the pixel pixel value variance of every block image respectively.

The picture quality detection method of one embodiment of the invention, wherein, the variance of a pictures represents with following formula:

wherein, x _irepresent the pixel value of i point in the picture, represent the mean pixel of a pictures.

Any one picture quality detection method above-mentioned, wherein, before step S3, also comprises step S2, carries out garbage rejecting operation according to the threshold value of setting.

The picture quality detection method of one embodiment of the invention, wherein, in step S3, weighting factor represents with following formula:

$w_{i,j}=\left\{\begin{array}{cc}{\mathrm{\&sigma;}}_{i,j}& {\mathrm{\&sigma;}}_{i,j}>T\\ 0& {\mathrm{\&sigma;}}_{i,j}<T\end{array}\right.,$ Wherein, threshold T determines according to initial image variance mean value, namely and λ=0.1.

The picture quality detection method of one embodiment of the invention, wherein, for the region of the jth under i scale, its weight is defined as:

$W_{i,j}=\frac{w_{i,j}}{\underset{j=1}{\overset{n}{\&Sigma;}}{w}_{i,j}}.$

The picture quality detection method of one embodiment of the invention, wherein, the visual information fidelity through variance weighted represents with following formula:

$VVIF(x,y)=\frac{\underset{k=1}{\overset{s}{\&Sigma;}}\underset{i=1}{\overset{M_j}{\&Sigma;}}{W}_{i,j}I(c_{i,j},f_{i,j})}{\underset{j=1}{\overset{s}{\mathrm{\&Sigma;}}}\underset{i=1}{\overset{M_j}{\&Sigma;}}{W}_{i,j}I(c_{i,j},e_{i,j})}.$

Wherein, reference picture (i.e. original image) and blurred picture (namely through the image of loss) is only had can standardized weight to be obtained at same ratio.Therefore, when calculating weight, reference picture and blurred picture are at same ratio.

Based on preceding method, the advantage that the present invention compares prior art is: picture quality detection method of the present invention, can obtain and evaluate and test more consistent result with human subject.

Accompanying drawing explanation

Fig. 1 is VIF evaluation model schematic diagram of the prior art;

Fig. 2 is the overall flow figure of the method for one embodiment of the invention;

Fig. 3 adopts method of the present invention to choose the zones of different in piece image and the close-up schematic view (wherein, the variance of region a is 4.33, and the variance of region b is 66.39) of chosen area;

Fig. 4 be adopt method of the present invention and the relatively objective IQA algorithm of DMOS adopted in the same width blurred picture that obtains respectively of method of the prior art scatter diagram (wherein, figure (a) is VVIF of the present invention, figure (b) is existing VIF, figure (c) is existing SSIM, figure (d) is existing PSNR, figure (e) is existing SEDLAI, and figure (f) is existing IGM).

Embodiment

In order to better explain the present invention, so that understand, below in conjunction with accompanying drawing, by embodiment, the present invention is elaborated.

The picture quality detection method of a kind of view-based access control model notice of the present invention, on the basis using visual information fidelity (VIF) method, adding with variance is the image structure information analysis that vision makes a decision factor, and set up the visual attention distributed model of simulating human vision system to image with this, structural information complexity variance by fragmental image processing and by calculating every block image respectively, the structural information complexity variance in foundation different images region, to VIF method weighted calculation, finally contrasts reference picture evaluate image quality.Method of the present invention can realize the subjective quality assessment of simulating human vision system to image, thus efficiently solves the low practicality of traditional objective graphical quality assessment and the low consistency problem to subjective picture quality assessment.

Wherein, the present invention is based on gauss hybrid models (GMM) and visual information fidelity (VIF) method, carried out the method for simulating human visual attention by " carrying out variance calculating to pixel pixel value in image zones of different; and carry out garbage according to the threshold value of priori setting and reject operation; the variance result of calculation afterwards according to image zones of different carries out the weighting of VIF method; simulating human vision system distributes the visual attention of image zones of different ", be particularly suitable for the image quality measure of realization to 2D plane picture.

In picture quality detects, visual attention plays a very important role, in a given picture, human visual system will give more visual attention to interested region, in the method for the invention, owing to using the structural information representated by variance to carry out computation vision notice, making method of the present invention compare other Objective image quality appraisal procedures and human subject, to evaluate and test mark consistance the highest.

Below in conjunction with a specific embodiment, method of the present invention is elaborated, see Fig. 2, the picture quality detection method of a kind of view-based access control model notice of the present invention, it comprises:

S1, variance calculating is carried out to the structural information complexity in image zones of different;

S2, according to setting threshold value carry out garbage reject operation;

The structural information complexity variance in S3, foundation different images region is to VIF method weighted calculation;

S4, final contrast reference picture evaluate image quality, to realize the subjective quality assessment of simulating human vision system to image.

Concrete, for the image shown in Fig. 3, its variance can be expressed as:

$\mathrm{\&sigma;}=\sqrt{\frac{\underset{i=1}{\overset{n}{\&Sigma;}}{(x_i-\stackrel{\&OverBar;}{x})}^2}{n-1}},$

Wherein Xi represents the pixel value of i point in the picture, and x represents the mean pixel of a pictures.

Such as, when picture region being divided into several regions, wherein, the variance of region a is 4.33, and the variance of region b is 66.39, so in this pictures, we can think that human visual system will give the more visual attention of region b.

In addition, due to human visual system can not give comprise less information image-region with visual attention, therefore, in order to focus is placed on comprise more image informations region on, can directly ignore the image-region only comprising a small amount of information.

Wherein, threshold function table (threshold value T) can be passed through and select useful region.

Such as, weighting factor is expressed as: $w_{i,j}=\left\{\begin{array}{cc}{\mathrm{\&sigma;}}_{i,j}& {\mathrm{\&sigma;}}_{i,j}>T\\ 0& {\mathrm{\&sigma;}}_{i,j}<T\end{array}\right..$

Wherein, threshold value T depends on initial image variance mean value,

$\begin{array}{cc}T=\mathrm{\&lambda;}\stackrel{\&OverBar;}{\mathrm{\&sigma;}}& \mathrm{\&lambda;}=0.1\end{array}.$

Thus, for the region of the jth under i scale, its weight is defined as:

$W_{i,j}=\frac{w_{i,j}}{\underset{j=1}{\overset{n}{\&Sigma;}}{w}_{i,j}}.$

Due to, reference picture (i.e. original image) and blurred picture (namely through the image of loss) is only had can standardized weight to be obtained at same ratio, therefore, can be expressed as according to the variance weighted visual information fidelity of existing VIF algorithm in conjunction with weight process gained of the present invention:

$VVIF(x,y)=\frac{\underset{j=1}{\overset{s}{\&Sigma;}}\underset{i=1}{\overset{M_j}{\&Sigma;}}{W}_{i,j}I(c_{i,j},f_{i,j})}{\underset{j=1}{\overset{s}{\&Sigma;}}\underset{i=1}{\overset{M_j}{\&Sigma;}}{W}_{i,j}I(c_{i,j},e_{i,j})}.$

Under normal circumstances (except blurred picture is excessively strengthened), the value of VVIF between 0-1, when reference picture and blurred picture similarity high time, VVIF=1.

In order to further illustrate beneficial effect of the present invention, contrast below in conjunction with the method for specific embodiment to method of the present invention and prior art, concrete, be that itself and these existing methods of PSNR, SSIM, VIF, SEDLAI and IGM are carried out Experimental comparison, and contrast the performance of measuring and calculating different I QA algorithm by three evaluation indexes (the highest CC value and minimum MAE and RMSE value).Specific experiment result is see table 1, table 2, table 3:

The comparison of CC performance under each type of distortion of table 1.

Algorithm	JP2K	JPEG	Blur	WN	FF	All
							PSNR	0.874	0.865	0.777	0.939	0.907	0.858
SSIM	0.892	0.928	0.889	0.932	0.904	0.829
							SEDLAI	0.884	0.922	0.944	0.948	0.888	0.847
IGM	0.936	0.858	0.883	0.865	0.831	0.768
							VIF	0.959	0.904	0.934	0.97	0.921	0.938
VVIF	0.963	0.92	0.953	0.982	0.911	0.941

The comparison of MAE performance under each type of distortion of table 2.

Algorithm	JP2K	JPEG	Blur	WN	FF	All
							PSNR	12.22	15.98	11.68	5.13	13.77	13.87
SSIM	11.38	11.87	8.49	7.99	14.44	15.27
							SEDLAI	7.36	7.69	6.24	8.86	8.81	9.55
IGM	8.41	8.15	7.37	11.01	9.14	12.78
							VIF	6.75	6.79	5.61	5.34	6.40	6.94
VVIF	6.45	6.23	4.75	4.18	6.79	6.79

The comparison of RMS performance under each type of distortion of table 3.

Algorithm	JP2K	JPEG	Blur	WN	FF	All
							PSNR	12.22	15.98	11.68	5.13	13.77	13.87
SSIM	11.38	11.87	8.49	7.99	14.44	15.27
							SEDLAI	7.36	7.69	6.24	8.86	7.81	8.55
IGM	8.41	8.15	7.37	11.01	9.14	12.78
							VIF	6.75	6.79	5.61	5.34	6.40	6.94
VVIF	6.45	6.23	4.75	4.18	6.79	6.79

Contrast the above results is known, the visual attention algorithm adopted due to VVIF algorithm of the present invention can carry out different process for the region that image is different, therefore, it has minimum standard error in all algorithms, therefore, VVIF algorithm of the present invention compares existing algorithm has and more presses close to the result of human visual system to image evaluation.

Again see the scatter diagram illustrating VVIF and contrast algorithm in Fig. 4, figure, can significantly find out by figure, VVIF algorithm of the present invention is compared existing algorithm and is more pressed close to the result of human visual system to image evaluation.

Claims (9)

1. the picture quality detection method of a view-based access control model notice, it is characterized in that, mainly on the basis of visual information fidelity method, in conjunction with the image structure information analysis taking variance as vision resolution factor, set up the visual attention distributed model of simulating human vision system to image.

2. picture quality detection method as claimed in claim 1, is characterized in that, comprise step:

S1, variance calculating is carried out to the structural information complexity in image zones of different;

The structural information complexity variance in S3, foundation different images region is to visual information fidelity method weighted calculation;

S4, final contrast reference picture evaluate image quality, to realize the subjective quality assessment of simulating human vision system to image.

3. picture quality detection method as claimed in claim 2, is characterized in that, step S1 specifically: to fragmental image processing, and calculate the pixel pixel value variance of every block image respectively.

4. picture quality detection method as claimed in claim 3, it is characterized in that, the variance of a pictures represents with following formula:

wherein, x _irepresent the pixel value of i point in the picture, represent the mean pixel of a pictures.

5. the picture quality detection method as described in claim 2,3 or 4, is characterized in that: before step S3, also comprises step S2, carries out garbage rejecting operation according to the threshold value of setting.

6. picture quality detection method as claimed in claim 5, it is characterized in that, in step S3, weighting factor represents with following formula:

$w_{i,j}=\left\{\begin{array}{cc}{\mathrm{\&sigma;}}_{i,j}& {\mathrm{\&sigma;}}_{i,j}>T\\ 0& {\mathrm{\&sigma;}}_{i,j}<T\end{array}\right.,$ Wherein, threshold T determines according to initial image variance mean value, namely and λ=0.1.

7. picture quality detection method as claimed in claim 6, it is characterized in that, for the region of the jth under i scale, its weight is defined as:

$W_{i,j}=\frac{w_{i,j}}{\underset{j=1}{\overset{n}{\&Sigma;}}{w}_{i,j}}.$

8. picture quality detection method as claimed in claim 7, is characterized in that: the visual information fidelity through variance weighted represents with following formula:

$VVIF(x,y)=\frac{\underset{j=1}{\overset{s}{\&Sigma;}}\underset{i=1}{\overset{M_j}{\&Sigma;}}{W}_{i,j}I(c_{i,j},f_{i,j})}{\underset{j=1}{\overset{s}{\&Sigma;}}\underset{i=1}{\overset{M_j}{\&Sigma;}}{W}_{i,j}I(c_{i,j},e_{i,j})}.$

9. picture quality detection method as claimed in claim 8, is characterized in that: when calculating weight, reference picture and blurred picture are at same ratio.