A no-reference image quality measure based on CPBD and noise

Yufeng Gu 1, a, Kejian Yang 1,b

1Computer Science and Technology, Wuhan University of Technology, Wuhan, Hubei 430063, China

a854987418@qq.com, bykjcgs@whut.edu.cn

Keywords: No-reference; Image Quality Assessment; CPBD; Noise pollution.

Abstract: In order to meet the needs of the objective evaluation of no-reference image, a no-reference image quality measure is presented. The measure is based on edge analysis and is suitable for images with noise. Taking properties of the Human Visual System(HVS) into account, we compute the probability of blur after getting the edge width and the local contrasts. And at last the image quality probability can be got considering cumulative probability of blur detection and the noise pollution degree. Experimental results show that the metric has a wide application, good anti-noise ability, simple calculation, as well as in high consistence with the subjective evaluation results.

1. Introduction

A With the widespread use of image information technology, the image which contains a lot ofvaluable information is more and more valued as the source of visual information. In the process of image acquisition and transmission, it is inevitable that images will be affected by noise, blur, the loss of data and so on, resulting in a decrease of image quality. The image quality directly affects the amount of information acquired and the subjective feelings of people. So, an automated and objective no-reference image quality evaluation assessment is needed considering the errors caused by subjective judgment and the time-consuming staff.

So far research on reference and semi-reference image quality evaluation has achieved good results. And the non-reference evaluation which is most practical has gradually become the researches emphasis in recent years. Some scholars have studied the image quality evaluation method, and have achieved some results.

These algorithms are either based on edge analysis to evaluate the image blurriness as in literature [1-4], or simply estimate the variance of the noise as in [7-10]. For images which contain blur and noise in the same time, these motheds lack the ability to evaluate them. Considering the situation, this paper proposes an evaluation method combining the edge analysis and image noise level, which can be used more broadly.

2. The Improved Quality Probability Model

To evaluate the quality of an image with blur and noise, an improved probability model based onthe edge analysis is proposed by taking noise into account. In this section, the implementation procedures of the algorithm are given in detail.

Journal of Electronics and Information Science (2016) 1: 11-16 Clausius Scientific Press, Canada

11

2.1 Edge D

The betthe edge wthe gradiensearching extremum accurate mliterature [performancin a decreacompared w

Where Fig. 1. ∆∅direction aAlthough contribute satisfy the ∆∅ 2and largeraffected by

Fig. 1. Pa

2.2 Cumul

The JusSystem (Hcontrast. F

Where parameter

Detection a

tter the imagwidth largelynt, and thealong the tis the edge

measurement[11] points ce. And conased performwith horizo

an∅ denoteand the vertthe denomto errors inconditions2° will decr thresholdsy imprecise

art of an imaline rep

lative Prob

st NoticeablHVS). It defi

or a given c

w e iβ is obtain

nd Edge W

ge is, the my reflects the gradient dtwo ends oe width. Buts，only thout that in

nsidering thmance whic

ontal/vertica

nd es the angltical direct

minator in n the meas of ∆∅rease perfos ∆∅measureme

age line. Thpresents the

bability Mo

le Differenfines the mincontrast c, a

is the just nned by using

Width Meas

more obviouhe sharpnesdirection isf the gradieut to reducee approximncorporatinhe measuremch is due to

al direction.

are local mle differenction, whereEq. 1 com

surement pr∆∅

rmance as t15° also

ents. The thr

he wide dottlocal maxim

odel of Blur

ce (JNB) cnimum amoa probability

noticeable bg the least s

surement.

us the edgesss of an imas obtained. ent directioe the comp

mately verticng extra homents of dio the largerThe edge w

maximum ance betweene andmpensates frocess. Thuare conside

there are nodecreases presholds ∆

ted lines indma

r Detection

concept is pount of percty of blurrin

1

blurred edgesquares curv

s are, and thage. In gene

Then the on. And theputational cocal edges arorizontal meiagonal edgr quantizatiowidth w e

∆∅.

nd local min the gradied arefor angle dus, only appered. Expero enough mperformanc∅ 8°

dicate the de and local

.

proposed inceived blurrness can be c

1

e width depve fitting m

he narrowereral, the difflocal extre

e distance bomplexity ae added to teasurementse gradients on width (√is compute

inimum of ent ∅e the imagedifference, proximatelyriments sho

measuremente, due to can lead to

etected edgeminimum

[5] based riness arouncalculated a

pended on tmethod for d

r the edge wfference is uemum can between theand ensure the calculats do not imwill conve√2) of diaged as follow

the edge pitan

e gradient colarge devi

y vertical edow that lowts to be repthe error p

o a good per

es. The narr.

on the Humnd an edge as follows:

.

the local codifferent con

width is. Soused to findbe find bye two localreasonably

tion, for themprove theersely resultgonal pixelsws:

(1)

ixel , see/

omponents.iations willdges which

w thresholdsresentative,propagationrformance.

row dotted

man Visualby giving a

(2)

ontrast. Thentrasts c in

o d y l y e e t s

e . l h s ,

n

l a

)

e n

12

order to increase the consistency between Eq. 2 and the experimental results of the fuzzy psychological perception. The experimental results in literature [6] find that the probability is more accurate and credible when β has a median value of 3.6, and the JNB edge widths for different contrasts can be modeled as:

w e5, c 503, c 51. (3)

An improved method of JNB is proposed in literature [9]. The method divides an image into 64x64 blocks. Then the blocks whose edge pixels are accounted for more than 0.2% of the total number of pixels in the block are defined as edge blocks to be processed further. It is pointed out that, at the JNB with, w e w e corresponds to a probability of blur detection pP 63%. Thus, for a given edge pixel e , when p P , it is considered that there is no blur to be detected at that edge. With the increase of the blur degree of an image, the edges’ width increases, which leads to a bigger value of w e and to a higher probability of blur detection. Considering all of the edge blocks and the edges, the entire image’s sharpness probability is modeled as a cumulative probability of blur detection (CPBD) as follows:

CPBD P p P ∑ P p . (4)

Where is defined as the value of the probability distribution function at a given .

2.3 Measurement of Noise Effects.

Noise, as one of the most common image degradation factors, is often produced in the process of generation and propagation of an image. When evaluating the image quality with noise, people pay more attention to the intensity of noise rather than it’s random characteristic，for it is the intensity of noise that effects people’s subjective perception. The differences between the images before and after denoising reflect the noise pollution degree. Thus, after using the total variation (TV)[10] minimization method which can save the image details as much as possible to denoise, the signal noise ratio(SNR) which denotes the ratio of the original image signal intensity and the intensity of noise can be got :

SNR 10 ∙ log∑ ∑ ,

∑ ∑ , ,. (5)

Where and are the size of the image. , denotes the denoised image and , denotes the original image with noise. Considering the randomness of the noise, the probability that the detected edges are effected by the noise to reduce the quality of the image is 1/ 1 , that is, the ratio of the image quality to be reduced.

2.4 The Comprehensive Evaluation Probability Model.

There is a CPBD model which denotes the probability of edge sharpness of the whole image and the ratio noise affects the quality of an image. Since the noise is random and independent of the image signal or the edge sharpness, a new probability model can be obtained by putting them together:

P ∙ CPBD. (6)

Where P is the final quality evaluation value, that is, the probability of a high quality image. And the higher the image quality is, the higher the value will be.

13

3. Experi

In this no-referenccontents ob

As the bquality is proposed mthere is no we blurredwhich ranggenerated iwere carrieand (b), itdeviation iscore is albecome ba

To evaluof the propThe result of the subjDMOS is,

imental Re

section, expce image qbtained from

blurriness inreduced, lemethod for single imag

d images wges from 0images withed out on alt can be foincrease, thlmost 0, forad.

Fig. 3.imag

uate imagesposed methois shown in

bjective valuthe worse th

esults

perimental quality measm the LIVE

ncrease, theeading to a

three groupge that is be

with a circu0 to 8 withh the proposll the testedound that the probabilitr almost al

(a)

. The perforges in LIVEs with noiseod about non Fig. 4. Thue betweenhe image qu

results are sure. The p database[1

Fig.

probabilitylower prob

ps images(Feing Gauss bular-symmeh the step ssed metric f

d images, anhe proposety decrease l the edge

rmance of thE and (b) cae, noise effeoise was teshe Differencn the originuality is.

provided toproposed m2].

2. Test im

y of blur detbability scoFig. 2) whiblurred with

etric 2D Gasize 0.2. Fifor the seriend a similared metric is rapidly. Wwidths are

he proposedaps image blect is added sted againstce Mean Opnal image a

o illustrate metric was t

mages

tection is exore. Fig. 3(aich are beinh continuouaussian kerig. 3(b) shoes of caps imr performans behaving

When the stagreater tha

d method telurred with into the ass

t other threepinion Scorand the dist

the performtested with

xpected to ina) shows thng Gauss blus interleavernel of stanows the remages. The nce was obs

as expecteandard devian the JNB

(b)

sted on (a) bdifferent σ

sessment moe metrics: De (DMOS) torted imag

mance of thimages wit

ncrease, andhe performalurred in Led standard ndard deviaesult of eva

same two eserved. Fromed. With thiation is abo

B width and

blurred σ . odel. The p

DMOS, PSNdenotes the

ge. And the

he proposedth different

d the imageance of the

LIVE. Sincedeviations,

ation σ ,aluating theexperimentsm Fig. 3 (a)he standardout 3.0, the

d the image

erformanceNR, CPBD.e differencee larger the

d t

e e e , , e s ) d e e

e . e e

14

The Tabof blur, whnoise on thnoise, the account theeffect incrobtained isthe image w

Fig. 4. T

At last, six imagesto form a sand proposcan be foudistorted imlower vola

4. Conclu

Image qobjective eimproves tthe proposand it is po

Reference

[1] Ong E and Its Appl

[2] Ong E

Table 1.

sigma DMOS PSNR CPBD Our

ble 1 showshich is lackhe edge deteCPBD mete noise effereases and s consistentwith noise.

The perform

we test the picked in eset of imagesed metric tund that thmages in LIatility.

usions

quality affecevaluation othe CPBD msed algorithossessed of

es

P, Lin W, Luications, 200

P, Lin W, Lu

standard de

0.0000 0.00.000 24+∞ 380.8446 0.80.8224 0.7s that the Ck of the abilection and cthod even cect resulting

dominates t with the s

mance of CP

performanceach type ofes whose Dto evaluate te proposedIVE. Compa

cts the acquof no referemodel by tam is close stronger app

u Z, et al. A n3. Proceeding

u Z, et al. No-

eviation sig

019531 0.4.376 338.9170 338135 0.7568 0.PBD metholity to evalu

calculation ocompletely g that when t

and correcsubjective e

PBD and pr

ce of the prf distortion,

DMOS distrithese image

d method haaring with C

uisition of inence image

aking into acto the subjeplicability a

no-reference gs. Seventh In

-reference JP

gma of white

.035156 03.502 43.9107 2.8362 0.6472 0od is only buate the noof the edge lost the refthe image icts the finaevaluation o

roposed met

roposed met 30 images ibution is rees, the resulas a more CPBD, it is

nformation,e is deeply ccount the iective perceand general

quality metrnternational S

PEG-2000 ima

e noise and

.140625 07.815 52.2460 1.8641 0.3964 0

based on theise level inwidth, when

ference. Buts polluted hal score mof HVS to a

tric on diffe

thod on otheare selected

elatively wilt shows in Faccurate ev closer to th

, thus, a genneeded. In influence ofeption of thity.

ic for measurSymposium o

age quality m

correspond

0.234375 153.846 518.3455 10.9053 00.2316 0e edge width

the image.n the imaget the propo

heavily by naking it poaccurately e

erent types o

er types of id from all thde and unifFig. 4. Throvaluation abhe subjectiv

neral and acthis paper,

f noise. Thehe quality o

ring image blon. IEEE, 200

metric[C]// Mu

ding score

1.1996094 57.678 10.0190 0.9249 0.0487 th to give a . Due to thee is pollutedosed methodnoise, the raossible thatevaluate the

of image dis

image distohe five distoform. Usingough the expability for ave perceptio

ccurate met, the propose test resultof all kinds

lur[C]// Signa03:469-472 v

ultimedia and

probabilitye impact of

d heavily byd takes intoatio of noiset the resulte quality of

stortion.

ortion. Withortion typesg the CPBDperiment, it

all kinds ofon and has a

thod for thesed methods show thatof images,

al Processingvol.1.

d Expo, 2003

y f y o e t f

h s

D t f a

e d t ,

g

3.

15

ICME '03. Proceedings. 2003 International Conference on. IEEE, 2003:I-545-8 vol.1.

[3] Yan Q, Xu Y, Yang X. No-reference image blur assessment based on gradient profile sharpness[C]// Broadband Multimedia Systems and Broadcasting (BMSB), 2013 IEEE International Symposium on. IEEE, 2013:1-4.

[4] Wang Y, Du H, Xu J, et al. A no-reference perceptual blur metric based on complex edge analysis[C]// Network Infrastructure and Digital Content (IC-NIDC), 2012 3rd IEEE International Conference on. IEEE, 2012:487-491.

[5] Ferzli R, Karam L J. Human Visual System Based No-Reference Objective Image Sharpness Metric[C]// IEEE International Conference on Image Processing. 2006:2949-2952.

[6] Ferzli R, Karam L J. A no-reference objective image sharpness metric Based on just-noticeable blur and probability summation[C]// IEEE International Conference on Image Processing. 2007(3): III-445−III-448

[7] Immerkær J. Fast Noise Variance Estimation[J]. Computer Vision & Image Understanding, 1996, 64(2):300-302.

[8] Laligant O, Truchetet F, Fauvet E. Noise estimation from digital step-model signal.[J]. IEEE Transactions on Image Processing A Publication of the IEEE Signal Processing Society, 2013, 22(12):5158-67.

[9] Narvekar N D, Karam L J. A no-reference perceptual image sharpness metric based on a cumulative probability of blur detection[C]// Quality of Multimedia Experience, 2009. QoMEx 2009. International Workshop on. 2009:87 - 91.

[10] Rudin L I, Osher S, Fatemi E. Nonlinear total variation based noise removal algorithms[J]. Physica D Nonlinear Phenomena, 1992, 60(1-4):259-268.

[11] Feichtenhofer C, Fassold H, Schallauer P. A Perceptual Image Sharpness Metric Based on Local Edge Gradient Analysis[J]. IEEE Signal Processing Letters, 2013, 20(4):379-382.

[12] H. R. Sheikh, Z. Wang, L. Cormack and A. C. Bovik, "LIVE Image Quality Assessment Database Release 2", http://live.ece.utexas.edu/research/quality.

16