EPSNR FOR OBJECTIVE IMAGE QUALITY MEASUREMENTS

Chulhee Lee, Guiwon Seo and Sangwook Lee

School of Electrical and Electronis Engineering,Yonsei University, Shinchon dong, Seoul, Korea Rep.

Keywords: EPSNR, Objective quality measurement, Perceptual image quality.

Abstract: In this paper, we explore the possibility to apply a recently standardized method for objective video quality

measurements to measure perceptual quality of still images. It has been known that the human visual system

is more sensitive to edge degradation. We apply this standardized method to several image data sets which

have subjective scores. The standardized method is compared with existing objective models for still images.

Experimental results show that the standardized method shows better performance than the conventional

PSNR and show similar performance compared to top-performance models.

1 INTRODUCTION

Objective quality assessment emerges an important

problem. As multimedia services, such as mobile

broadcasting, video on demand (VOD) and

videophones, over channels where bandwidth can

not be guaranteed become widely available, quality

monitoring becomes a critical issue. Traditionally,

perceptual video quality has been measured by a

number of evaluators who subjectively evaluate

video quality. Although this subjective evaluation is

considered to be the most accurate method, it is

expensive and cannot be done in real time. As a

result, efforts have been made to develop objective

models for perceptual video quality measurement.

These efforts have resulted in several international

standards (ITU-R, 2003, ITU-T, 2004). For example,

in (ITU-R, 2003, ITU-T, 2004), four objective

models are included. Although these models are

developed to measure perceptual quality of video

signals, they also might be used to measure

perceptual quality of still images.

Traditionally, codec optimization has been done

by minimizing mean square errors (equivalently

PSNR). However, it has been known that the

correlation between PSNR and perceptual quality is

not high and efforts are made to better metrics to

measure perceptual quality. These metrics can be

used for parameter optimization during encoding

process. They can be also used to evaluate new

codecs, video transmission systems, traffic

optimization, etc. In this paper, we explore the

possibility to apply the standardized method to

measure perceptual quality of still images. Among

the four models of (ITU-R, 2003, ITU-T, 2004), we

tested the model developed by Yonsei. The model is

easy to implement and very fast. Consequently, it

can be used for codec optimization which requires a

large number of computations of the metric. We

applied the method the database which has been

widely used for quality measurement of still images.

We also conducted our own subjective test and

evaluate the performance of the model.

2 EPSNR

It has been known that the human visual system is

more sensitive to edge degradation. Thus, in (ITU-R,

2003, ITU-T, 2004), an edge detection algorithm is

first applied to find edge areas. For example, the

horizontal gradient image and the vertical gradient

image are first computed using gradient operators.

Then, the magnitude gradient image is computed as

follows:

(,) (,) (,)

horizontal vertical

gmn g mn g mn=+. (1)

Then, thresholding operation is applied to the

magnitude gradient image to determine edge pixels.

In (ITU-R, 2003, ITU-T, 2004), the Sobel operator

is recommended.

Alternatively, it is possible to use the successive

edge detection procedure. First, a vertical gradient

operator is applied to the reference image, producing

a vertical gradient image. Then, a horizontal gradient

operator is applied to the vertical gradient image,

Lee C., Seo G. and Lee S. (2009).

EPSNR FOR OBJECTIVE IMAGE QUALITY MEASUREMENTS.

In Proceedings of the First International Conference on Computer Imaging Theory and Applications, pages 92-95

DOI: 10.5220/0001807800920095

 SciTePress

producing a horizontal and vertical gradient image.

Then, a thresholding operation produces edge areas.

Figs. 2-4 illustrate this procedure.

Figure 1: An original image.

Figure 2: A vertical gradient image.

Figure 3: A horizontal and vertical gradient image.

Figure 4: Edge pixels.

Next, errors between the edge pixels of the source

image and those of the processed image are

computed. For example, the squared error of edge

areas of the l-th frame can be computed as follows:

pixeledgeanisjiSif

jiPjiSse

lll

),(

}),(),({

∑∑

−=

(2)

where

),( jiS

is the l-th image of the source video

sequence,

),( jiP

represents the l-th image of the

processed video sequence. Finally, the global edge

mean squared error is computed as follows:

mse se

∑

where K is the total number of pixels of the edge

areas. Finally, the edge PSNR (EPSNR) is computed

as follows:

10log ( )

EPSNR

mse

where P is the peak value. If the number of edge

pixels is smaller than a certain value, the threshold

value is reduced by 20 until a sufficient number of

edge pixels are obtained. In the Yonsei model, this

EPSNR is used as a base objective video quality

metric after some posting-processing (C. Lee et al.,

2006). The model was validated by independent

laboratories and significantly outperformed PSNR

(Table 1).

Table 1: Performance comparison on the VQEG Phase II

Data.

correlation

The Yonsei

model

525 0.9086

625 0.8519

PSNR

525 0.7573

625 0.6954

3 EXPERIMENTAL RESULTS

We applied the Yonsei model to the database

provided by LIVE Lab (H. R. Sheikh, et al. 2005). In

(H. Sheikh, 2006), a number of perceptual quality

metrics for still images are compared using the same

database. Thus, we compared the Yonsei model with

the models of (H. Sheikh, 2006). Table 2 shows

performance comparison. It is noted that all the other

metrics are taken from (H. Sheikh, 2006) except for

EPSNR FOR OBJECTIVE IMAGE QUALITY MEASUREMENTS

EPSNR. As can be seen, EPSNR provides high

correlations.

Table 2: Performance comparison (H. Sheikh, 2006).

jp2k#2 jpeg#2

PSNR 0.8740 0.9167

JND 0.9734 0.9870

DCTune 0.7834 0.9418

PQS 0.9364 0.9777

NQM 0.9463 0.9783

Fuzzy 0.9133 0.9509

BSDM 0.9450 0.9747

SSIM(MS) 0.9711 0.9879

IFC 0.9626 0.9744

VIF 0.9787 0.9885

EPSNR 0.9515 0.9457

In the next test, we printed some of the LIVE data

set using two colour printers. Using the printed

images, we performed subjective tests. Fig. 5 shows

the scatter plots between the subjective scores and

EPSNR. As can be seen, EPSNR shows high

correlations.

5 CONCLUSIONS

In this paper, we applied the recently standardized

method for objective video quality measurements to

measure perceptual quality of still images.

Experimental results show that it would be possible

to use the model for still images. By taking into

account the characteristics of still images, it would

be also possible to improve the performance of the

model.

REFERENCES

Recommendation ITU-R BT.1683, “Objective perceptual

video quality measurement techniques for digital

broadcast television in the presence of a full

reference,” 2003. Smith, J., 1998. The book, The

publishing company. London, 2

edition.

ITU-T Recommendation J.144, “Objective perceptual

video quality techniques for digital cable television in

the presence of a full reference,” 2004. 11

C. Lee, S. Cho, J. Choe, T. Jeong, W. Ahn and E. Lee,

“Objective Video Quality Assessment,” Optical

Engineering, vol. 45, no. 1, pp. 1-11, 2006.

(a)

(b)

Figure 5: Scatter plot between EPSNR and subjective

scores of the printed images (a) Printer A, correlation

coefficient: 0.894179 (b) Printer B, correlation coefficient:

0.918683.

H. Sheikh, M. Sabir, and A. Bovik, “A Statistical

Evaluation of Recent Full Reference Image Quality

Assessment Algorithms,” IEEE Trans. Image

Processing, Vol. 15, No. 11, pp. 3441-3452, 2006.

JNDmetrix Technology Sarnoff Corp., evaluation Version,

2003 [Online].Available:

http://www.sarnoff.com/products-services/video-

vision/jndmetrix/downloads.asp

A. B. Watson, “DCTune: A technique for visual

optimization of DCT quantization matrices for

individual images,” Soc. Inf. Display Dig.Tech. Papers,

vol. XXIV, pp. 946–949, 1993.

M. Miyahara, K. Kotani, and V. R. Algazi, “Objective

Picture Quality Scale (PQS) for image coding,” IEEE

IMAGAPP 2009 - International Conference on Imaging Theory and Applications

Trans. Commun., vol. 46, no. 9,pp. 1215–1225, Sep.

1998.

N. Damera-Venkata, T. D. Kite, W. S. Geisler, B. L.

Evans, and A.C. Bovik, “Image quality assessment

based on a degradation model,”IEEE Trans. Image

Process., vol. 4, no. 4, pp. 636–650, Apr. 2000.

D. V. Weken, M. Nachtegael, and E. E. Kerre, “Using

similarity measures and homogeneity for the

comparison of images,” Image Vis.Comput., vol. 22,

pp. 695–702, 2004.

I. Avcibas¸, B. Sankur, and K. Sayood, “Statistical

evaluation of image quality measures,” J. Electron.

Imag., vol. 11, no. 2, pp. 206–23, Apr., 2002.

Z. Wang, E. P. Simoncelli, and A. C. Bovik, “Multi-scale

structural similarity for image quality assessment,”

presented at the IEEE Conf. Signals, Systems, and

Computers, Vol. 2, pp. 1398-1402, Nov. 2003.

H. R. Sheikh, A. C. Bovik, and G. de Veciana, “An

information fidelity criterion for image quality

assessment using natural scene statistics,”IEEE Trans.

Image Process., vol. 14, no. 12, pp. 2117–2128, Dec.

2005.

H. R. Sheikh and A. C. Bovik, “Image information and

visual quality,”IEEE Trans. Image Process., vol. 15,

no. 2, pp. 430–444, Feb. 2006.

H. R. Sheikh, Z. Wang, L. Cormack, and A. C. Bovik,

LIVE Image Quality Assessment Database, Release 2

2005(Online). Available:

http://live.ece.utexas.edu/research/quality.

EPSNR FOR OBJECTIVE IMAGE QUALITY MEASUREMENTS