CN102006497A - No-reference blurred image evaluation method based on local statistical characteristics of images - Google Patents

Abstract

The invention discloses a no-reference blurred image evaluation method based on local statistical characteristics of images, mainly solving the problem of no-reference objective evaluation of the blurred images. The method comprises the following steps: (1) firstly generating a blurred image on the image to be tested by filtering; (2) detecting the edge of the original image to be tested by the Sobel operator and selecting the local area set surrounding the edge points; (3) carrying out statistics on variation of the original image and the generated blurred image according to the selected local areas; (4) properly adjusting the variation statistics of the local areas; and (5) constructing an evaluation metric of the blurred image according to the variation statistics. The method has the advantages of simple structure, low computational complexity, easy hardware implementation and consistence with subjective evaluation and can be used for detecting effectiveness of the image and video processing method.

Description

Blurred picture based on the image local statistical nature does not have with reference to evaluating method

Technical field

The present invention relates to a kind of blurred picture based on the image local statistical nature does not have with reference to evaluating method, belongs to technical field of image processing.

Background technology

Along with computer network and development of Communication Technique, current digital image has become a kind of important media that people obtain information and exchange and interdynamic.Digital picture is being obtained, and compression is handled, and is easy to generate various distortions in transmission and the reconstruction processes, and picture quality is to weigh the leading indicator of these distortions.The method of evaluation and test picture quality has subjective method and objective method, because the final observer of image is the people, so subjective method reliability height, but subjective method wastes time and energy and is not easy to and be embedded in the automated system, so objective method is the emphasis of current quality assessment.Whether objective evaluating method according to having reference picture to be divided into 3 classes again, full reference mass evaluation and test, partial reference quality assessment and do not have the reference mass evaluation and test.Full reference and partial reference quality assessment method need the Partial Feature of reference picture or reference picture, and in many practical applications, reference picture or its Partial Feature are difficult to maybe can't obtain, and this moment, non-reference picture quality assessment method just seemed particularly important.

Fuzzy is again the common phenomenon of image fault.There are many reasons can cause the fuzzy of image, for example denoising, filtering, compression or motion or the like.In recent years, many scholars begin to be devoted to not have the research with reference to the blurred picture quality assessment, introduce several exemplary process below.

P.Marziliano etc. have proposed a kind of perceived blur and ring alteration scale of measurement " P.Marziliano; F.Dufaux; S.Winkler and T.Ebrahimi.Perceptual blur and ringing metrics:Applications to JPEG2000; Signal Processing:Image Communication.; vol.19; no.2, pp.163-172,2004 " at the JPEG2000 image.This method at first detects the vertical edge of image by the SOBEL operator, line by line scan then and find the beginning and the end position at edge, the distance of beginning and end position is taken as the local fuzzy mearue when leading edge, and last average all local fuzzy mearues are as final fuzzy tolerance.R.Ferzli and L.J.Karam proposes a kind of nothing with reference to objective image sharpening yardstick (JNBM) " R.Ferzli and L.J.Karam; A No-Reference Objective Image Sharpness Metric Based on the Notion of Just Noticeable Blur (JNB); IEEE Transactions on Image Processing; vol.18; no.4; pp.717-728 based on just noticeable fuzzy concept, 2009. " this method at first becomes image division 64 * 64 image block, the edge pixel that detects according to the SOBEL operator is counted and according to certain threshold value piece is divided into smooth and non-smooth.Width and local contrast according to the edge derive non-smooth fuzzy mearue, are calculated the fuzzy mearue of entire image by non-smooth fuzzy alteration situation.Propositions such as Rania Hassen a kind of based on local phase the consistent non-reference picture sharpening of measuring estimate (LPCM) " R.Hassen; Z.Wang and M.Salama; No-reference image sharpness assessment based on local phase coherence measurement; in Proc.IEEE Int.Conf.Acoustics, Speech ﹠amp; Signal Processing, Mar.2010. " pyramid wavelet decomposition image that this method utilization can be controlled, obtain the local phase coherent image, estimate as final image sharpening by the consistent figure of weighted average local phase at last.This method relates to complicated wavelet transformation and phase calculation, the algorithm complex height, and it is low to carry out efficient.Therefore how it is extremely important that high efficiency is measured fuzzy alteration in high quality.

Summary of the invention

It is higher that existing blurred picture does not have the common complexity of reference mass evaluating method, and it is limited with subjective evaluating method consistency, the objective of the invention is in order to overcome this limitation, propose a kind ofly simple and effectively to put the fuzzy mearue of regional partial statistics characteristic based on the image border, realize simple effectively, be easy to hard-wired blurred picture and do not have with reference to evaluation and test.

According to technical scheme provided by the invention, described blurred picture based on the image local statistical nature do not have with reference to evaluating method by treat test pattern and again blurred picture fringe region partial statistics characteristic make up the fuzzy scale of measurement, process is as follows:

(1) selects the regional area set: use the Sobel operator to detect original edge of image pixel to be tested, distribute the center of each edge pixel to one block of pixels then;

(2) produce blurred picture again: the filter that uses 3*3

Do convolution algorithm with original image, regeneration one width of cloth blurred picture on original testing image;

(3) calculate the regional area variation:

If the set of the regional area pixel of selecting according to step (1) is L={l _i1≤i≤M},

M is the number of regional area pixel in the formula, l _iThe gray value of representing i pixel;

At image to be tested with calculate the variation of regional area pixel again in the blurred picture respectively

$\mathrm{SV}={\left(\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{(l_i-\mathrm{\μ})}^2\right)}^{\mathrm{\α}},$

In the formula

Represent this area grayscale average; Parameter alpha is a constant, preferred 0.3＜α＜10, and α reached best effect at 4.3 o'clock.

(4) will be again the blurred picture variation statistics of comparing the zone that regional area variation statistics raises with original image to be tested be changed to 0;

(5) make up fuzzy evaluation and test yardstick: after finishing all regional area variation statistics, the partial statistics in two zones of adding up respectively promptly obtains

With

Ambiguity in definition image evaluation and test metric is

$\mathrm{SVBM}=\frac{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{ori}}-\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{re}-\mathrm{blur}}}{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{ori}}+\mathrm{\ϵ}},$

N is a regional area number in the image in the formula,

With

Be respectively the original testing image and the variation statistics of j regional area of blurred picture again; ε is a very little normal amount, in order to avoid denominator is removed by zero, can get 0＜ε≤0001.

Advantage of the present invention is: the present invention utilizes the human vision feature, promptly human relatively difficulty is distinguished the image of the identical content of different fog-levels, utilizes the partial statistics characteristic of two width of cloth blurred picture fringe regions to define the blurred picture quality dexterously and does not have with reference to estimating; The evaluation and test performance has a distinct increment, and is better with human subjective vision perception consistency; Method is simple, and amount of calculation is little; Being easy to hardware realizes.

Description of drawings

Fig. 1 is a flow chart of the present invention.

Fig. 2 is the subregion schematic diagram of 3 * 3 block of pixels.

Fig. 3 is the subregion schematic diagram of 5 * 5 block of pixels.

Fig. 4 is the coefficient of relationship schematic diagram between parameter alpha and the inventive method and the subjective score.

Fig. 5 is a diffusing some schematic diagram of the inventive method and subjective perception score.

Embodiment

The invention will be further described below in conjunction with accompanying drawing and example.Relate to the evaluating method of nothing, can be used in the image/video related application evaluation process image blurring degree with reference to the blurred picture quality.Blurred picture based on the image local statistical nature of the present invention does not have with reference to evaluating method, by to test pattern and again blurred picture fringe region partial statistics characteristic make up the scale of measurement, detailed process is as shown in Figure 1.

(1) selects the regional area set

Use the Sobel operator to detect the edge pixel of original testing image, distribute the center of each edge pixel to one block of pixels then.

The minimum pixel piece that surrounds marginal point is 3 * 3.We are divided into four littler pixel set with this 3 * 3 block of pixels, are referred to as subregion, as shown in Figure 2.

Among Fig. 2, pixel is labeled as belonging to of same numeral of same subregion, and what pixel was labeled as two numerals shows that it belongs to two sub regions simultaneously.For example, the pixel that is labeled as " 12 " had both belonged to subregion 1, also belonged to subregion 2.Like this, 3 * 3 zones that surround edge pixel point are divided into four sub regions along four direction, and each subregion and edge pixel point have same distance.

Fig. 3 has listed the four sub regions definition of 5 * 5 block of pixels that comprise boundary pixel point, and each subregion all comprises 6 pixels, all edge pixel point in center is had same contribution.

(2) produce blurred picture again

Use the filter of 3*3

Do convolution algorithm with original image, regeneration one width of cloth blurred picture on original testing image.

(3) calculate the regional area variation

The set of supposing the regional area pixel is:

L＝{l _i；1≤i≤M} (1)

M is the number of regional area pixel in the formula, l _iThe gray value of representing i pixel.

At original image to be tested with calculate the variation of regional area pixel again in the blurred picture respectively:

$\mathrm{SV}={\left(\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{(l_i-\mathrm{\μ})}^2\right)}^{\mathrm{\α}}---\left(2\right)$

In the formula

Represent this area grayscale average; Parameter alpha is a normal amount, preferably gets 0.3＜α＜10, obtains α among the present invention by experiment and be reaching best effect at 4.3 o'clock.As shown in Figure 4, abscissa is the different value of parameter alpha, and ordinate is the coefficient correlation between the inventive method score and the subjective evaluation and test score, reaches the highest coefficient correlation when α is 4.3.

(4) suitably revise regional area variation statistics

In experiment, we more again blurred picture and original image to be tested find that most of regional area variation statistics will reduce or remain unchanged, but also have the minority regional area make an exception (this exception occur in many marginal points around regional area).We think the fuzzy variation statistics that should reduce regional area, and this exception can not truly reflect fuzzy influence, so the variation statistics that we adjust these exception areas is 0.

(5) make up fuzzy evaluation and test yardstick

After finishing all regional area variation statistics, the partial statistics in two zones of adding up respectively then promptly obtains With

Ambiguity in definition image evaluation and test metric is:

$\mathrm{SVBM}=\frac{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{ori}}-\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{re}-\mathrm{blur}}}{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{ori}}+\mathrm{\ϵ}}---\left(3\right)$

N is a regional area number in the image in the formula,

With

Be respectively the original testing image and the variation statistics of j regional area of blurred picture again.ε is a very little normal amount (such as 0＜ε≤0.001), in order to avoid denominator is removed by zero, ε is taken as 0.001 in the experiment.Ambiguity value SVBM is between 0～1, and it is big more to be worth its fog-level of more little expression.

Compared to existing technology, the present invention has the following advantages:

(1) the evaluation and test performance has a distinct increment, and is better with human subjective vision perception consistency.Test by LIVE laboratory, University of Texas blurred picture quality assessment database, the inventive method is respectively 0.9548 and 0.9616 with the Spearman coefficient of relationship and the nonlinear regression coefficient of subjective evaluation and test score, and the best way in the current pertinent literature report, i.e. method (the R.Hassen that measures based on the non-reference picture sharpening of local phase unanimity, Z.Wang and M.Salama, No-reference image sharpness assessment based on local phase coherence measurement, in Proc.IEEE Int.Conf.Acoustics, Speech ﹠amp; Signal Processing, 2010), only be 0.9239 and 0.9368, improved about 0.031 and 0.025 respectively.

(2) method is simple, and amount of calculation is little.The inventive method only need be carried out rim detection and simple statistical operation in the spatial domain, and comparing the consistent method of measuring of local phase needs complicated wavelet transformation and phase calculation, and amount of calculation reduces greatly.

(3) being easy to hardware realizes.Because the present invention only needs simple rim detection and some statistical operations, be easy to realize by hardware, can be used for image/video and handle associated equipment.

Advantage of the present invention can further prove by following experiment:

This experiment is carried out on the LIVE of University of Texas blurred picture quality assessment database, has 156 width of cloth to blur distorted image in this database, has provided the subjective score MOS value of this 156 sub-picture simultaneously.In order to test the consistency of the present invention and subjective perception, we have selected following two kinds of measurement criterions: (1) Spearman rank order coefficient of relationship (SROCC), the monotonicity of reflection objective evaluating prediction achievement; (2) coefficient correlation (CC), the accuracy of reflection objective evaluating.Table 1 provided the inventive method and pertinent literature method the contrast situation (a kind of based on local phase the consistent non-reference picture sharpening of measuring estimate (LPCM) " R.Hassen; Z.Wang and M.Salama; No-reference image sharpness assessment based on local phase coherence measurement; in Proc.IEEE Int.Conf.Acoustics, Speech ﹠amp; Signal Processing, 2010. "; A kind of nothing based on just noticeable fuzzy concept is with reference to objective image sharpening yardstick (JNBM) " R.Ferzli and L.J.Karam; A No-Reference Objective Image Sharpness Metric Based on the Notion of Just Noticeable Blur (JNB); IEEE Transactions on Image Processing; vol.18; no.4; pp.717-728,2009. "; A kind of perceived blur and the ring scale of measurement (PBRM) " P.Marziliano; F.Dufaux; S.Winkler and T.Ebrahimi.Perceptual blur and ringing metrics:Applications to JPEG2000; Signal Processing:Image Communication; vol.19; no.2, pp.163-172,2004. " at the JPEG2000 image.

The subjective and objective consistency evaluation and test comparative result of table 1 the inventive method and pertinent literature method

Model	Coefficient correlation	The rank order coefficient of relationship
			PBRM	0.9105	0.8919
JNBM	0.8168	0.7774
			LPCM	0.9368	0.9239
The present invention	0.9616	0.9548

As can be seen from the table, the existing relatively method of the inventive method has superiority preferably: (1) coefficient correlation is the highest, illustrates that this method has higher accuracy of forecast; (2) the rank order correlation coefficient is also the highest, illustrates that it has stricter prediction monotonicity.

Fig. 5 has shown the scatter diagram of the inventive method and subjective scoring, and abscissa is the inventive method objective evaluating picture quality score, and ordinate is subjective evaluation and test picture quality score.This figure has also illustrated the height consistency of the inventive method and subjective perception.

Claims (2)

1. the blurred picture based on the image local statistical nature does not have with reference to evaluating method, it is characterized in that by treat test pattern and again blurred picture fringe region partial statistics characteristic make up the fuzzy scale of measurement, process is as follows:

(1) selects the regional area set: use the Sobel operator to detect original edge of image pixel to be tested, distribute the center of each edge pixel to one block of pixels then;

(2) produce blurred picture again: the filter that uses 3*3

Do convolution algorithm with original image, regeneration one width of cloth blurred picture on original testing image;

(3) calculate the regional area variation:

If the set of the regional area pixel of selecting according to step (1) is L={l _i1≤i≤M},

M is the number of regional area pixel in the formula, l _iThe gray value of representing i pixel;

At image to be tested with calculate the variation of regional area pixel again in the blurred picture respectively

$\mathrm{SV}={\left(\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{(l_i-\mathrm{\μ})}^2\right)}^{\mathrm{\α}},$

In the formula

Represent this area grayscale average, parameter alpha is a constant, 0.3＜α＜10;

(4) will be again the blurred picture variation statistics of comparing the zone that regional area variation statistics raises with original image to be tested be changed to 0;

(5) make up fuzzy evaluation and test yardstick: after finishing all regional area variation statistics, the partial statistics in two zones of adding up respectively promptly obtains

With

Ambiguity in definition image evaluation and test metric is

$\mathrm{SVBM}=\frac{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{ori}}-\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{re}-\mathrm{blur}}}{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{SV}}_j^{\mathrm{ori}}+\mathrm{\ϵ}},$

N is a regional area number in the image in the formula,

With

Be respectively the original testing image and the variation statistics of j regional area of blurred picture again; ε is a constant, 0＜ε≤0.001.

2. do not have with reference to evaluating method based on the blurred picture of image local statistical nature according to claim 1, it is characterized in that parameter alpha=4.3.

Images