Beruflich Dokumente
Kultur Dokumente
IMAGES
Laboratory for Image and Video Engineering (LIVE), Dept. of Electrical and Computer Engineering,
The University of Texas at Austin, Austin, TX 78712-1084, USA
Email: zhouwang@ieee.org, {sheikh, bovik}@ece.utexas.edu
ABSTRACT a great deal of effort has been made to develop new ob-
jective image/video quality metrics that incorporate percep-
Human observers can easily assess the quality of a distorted
tual quality measures by considering Human Visual System
image without examining the original image as a reference.
(HVS) characteristics [1]–[4].
By contrast, designing objective No-Reference (NR) quality
Most of the proposed image quality assessment approa-
measurement algorithms is a very difficult task. Currently,
ches require the original image as a reference. Interestingly,
NR quality assessment is feasible only when prior knowl-
human observers can easily assess the quality of distorted
edge about the types of image distortion is available.
images without using any reference image. By contrast, de-
This research aims to develop NR quality measurement
signing objective No-Reference (NR) quality measurement
algorithms for JPEG compressed images. First, we estab-
algorithms is a very difficult task. This is mainly due to the
lished a JPEG image database and subjective experiments
limited understanding of the HVS, and it is believed that
were conducted on the database. We show that Peak Signal-
effective NR quality assessment is feasible only when the
to-Noise Ratio (PSNR), which requires the reference im-
prior knowledge about the image distortion types is avail-
ages, is a poor indicator of subjective quality. Therefore,
able. Although only a limited number of methods have been
tuning an NR measurement model towards PSNR is not an
proposed in the literature [5]–[9] for objective NR quality
appropriate approach in designing NR quality metrics. Fur-
assessment, this topic has attracted a great deal of attention
thermore, we propose a computational and memory efficient
recently. For example, the Video Quality Experts Group
NR quality assessment model for JPEG images. Subjective
(VQEG, http://www.vqeg.org) considers the standardization
test results are used to train the model, which achieves good
of NR and Reduced-Reference (RR) video quality assess-
quality prediction performance. A Matlab implementation
ment methods as one of its future working directions, where
of the proposed method is available at http://anchovy.ece.ute
the major source of distortion under consideration is block
xas.edu/˜zwang/research/nr jpeg quality/index.html .
DCT-based video compression.
The purpose of this research is to develop objective NR
1. INTRODUCTION quality assessment algorithms for JPEG compressed images.
Such algorithms must have the capability to effectively pre-
In recent years, there has been an increasing need to develop dict perceived JPEG image quality. We consider blurring
objective measurement techniques that can predict image and blocking as the most significant artifacts generated dur-
/video quality automatically. Such methods can have var- ing the JPEG compression process. An efficient way is pro-
ious applications. First, they can be used to monitor im- posed to extract features that can be used to reflect the rela-
age/video quality for quality control systems. Second, they tive magnitudes of these artifacts. The extracted features are
can be employed to benchmark image/video processing sys- combined to constitute a quality prediction model. Subjec-
tems and algorithms. Third, they can also be embedded tive experimental results on JPEG compressed images are
into image/video processing systems to optimize algorithms used to train the model, which achieves very good quality
and parameter settings. The most widely used objective im- prediction performance.
age quality/distortion metrics are Peak Signal-to-Nose Ra-
tio (PSNR) and Mean Squared Error (MSE), but they are
2. SUBJECTIVE EXPERIMENTS
widely criticized as well for not correlating well with per-
ceived quality measurement. In the past three to four decades, The subjective test was conducted on 8 bits/pixel gray level
This research was supported in part by Texas Instruments, Inc., and by images. There are 120 test images in the database. Thirty of
the State of Texas Advanced Technology Program (ATP). them are original images, which are randomly divided into
Fig. 1. Group I images. Fig. 2. Group II images.
4
10
3
9
2
Group 1 images 8
Group 2 images
1
20 22 24 26 28 30 32 34 36 38 40 7
PSNR
MOS
Fig. 3. PSNR results compared with MOS. 5
1 4
original image
JPEG coded image
0.9
3
0.8
2
Group 1 images (training images)
0.7 Group 2 images
1
Power (log10(1+P[l]))
1 2 3 4 5 6 7 8 9 10
0.6 MODEL PREDICTION
0.5
Fig. 5. Model prediction results using Group I images as
0.4
the training images.
0.3
0.2
The second activity measure is the zero-crossing (ZC) rate.
0.1
We define for n ∈ [1, N − 2],
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 ½
Frequency (l/N) 1 horizontal ZC at dh (m, n)
zh (m, n) = (4)
0 otherwise
Fig. 4. Power spectrum comparison of the original and
JPEG compressed “Lena” images. The horizontal ZC rate then can be estimated as:
XM NX−2
1
Zh = zh (m, n) (5)
The features are calculated horizontally and then vertically. M (N − 2) i=1 j=1
First, the blockiness is estimated as the average differences
across block boundaries: Using similar methods, we calculate the the vertical features
M bN/8c−1
of Bv , Av , and Zv . Finally, the overall features are given by:
1 X X
Bh = |dh (i, 8j)| (2) Bh + Bv Ah + Av Zh + Zv
M (bN/8c − 1) i=1 j=1 B= , A= , Z= . (6)
2 2 2
Second, we estimate the activity of the image signal. Al- There are many different ways to combine the features to
though blurring is difficult to be evaluated without the ref- constitute a quality assessment model. One method we find
erence image, it causes the reduction of signal activity, and that gives good prediction performance is given by
combining the blockiness and activity measures gives more
insight into the relative blur in the image. The activity is S = α + βB γ1 Aγ2 Z γ3 , (7)
measured using two factors. The first is the average abso-
lute difference between in-block image samples: where α, β, γ1 , γ2 , and γ3 are the model parameters that
must be estimated with the subjective test data. The non-
XM N X−1 linear regression routine “nlinfit” in the Matlab statistics
1 8
Ah = |dh (i, j)| − Bh (3) toolbox is used to find the best parameters for (7). It is
7 M (N − 1) i=1 j=1 important that the model is not overtrained, in which case,
10 10
9 9
8 8
7 7
6 6
MOS
MOS
5 5
4 4
3 3
2 2
Group 1 images Group 1 images
Group 2 images (training images) Group 2 images
1 1
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
MODEL PREDICTION MODEL PREDICTION
Fig. 6. Model prediction results using Group II images as Fig. 7. Model prediction results using both groups of im-
the training images. ages as the training images.
although very good fitting is obtained for the training data, 5. REFERENCES
the model’s generalization ability is affected. Therefore, we
[1] VQEG, “Final report from the video quality experts
use different groups of images as the training images. The
group on the validation of objective models of video
results shown in Figs. 5, 6 and 7 are obtained using Group I
quality assessment,” http://www.vqeg.org/, Mar. 2000.
(Fig. 1), Group II (Fig. 2), and both groups of images as the
training images, respectively. The model performs well in [2] T. N. Pappas and R. J. Safranek, “Perceptual criteria
all three tests, which implies that the model is robust. This for image quality evaluation,” in Handbook of Image &
is confirmed by Table 1, where the Root Mean Squared er- Video Proc. (A. Bovik, ed.), Academic Press, 2000.
ror (RMS) between the model prediction score and MOS
is given. The parameters obtained with all test images are [3] C. J. van den Branden Lambrecht, Ed., “Special issue
α = −245.9, β = 261.9, γ1 = −0.0240, γ2 = 0.0160, and on image and video quality metrics,” Signal Processing,
γ3 = 0.0064, respectively. vol. 70, Nov. 1998.
[4] Z. Wang and A. C. Bovik, “Why is image quality
assessment so difficult?,” in Proc. IEEE Int. Conf.
4. CONCLUSIONS
Acoust., Speech, and Signal Processing, May 2002.
We demonstrate a novel NR perceptual quality assessment [5] H. R. Wu and M. Yuen, “A generalized block-edge im-
scheme for JPEG compressed images. Subjective experi- pairment metric for video coding,” IEEE Signal Pro-
ments were conducted to evaluate the quality of JPEG com- cessing Letters, vol. 4, pp. 317–320, Nov. 1997.
pressed images. The features described in the paper effec-
tively capture the artifacts introduced by JPEG, and the non- [6] Z. Wang, A. C. Bovik, and B. L. Evans, “Blind mea-
linear fitting gives good agreement with MOS scores. surement of blocking artifacts in images,” in Proc. IEEE
The method is computationally efficient since no com- Int. Conf. Image Proc., vol. 3, pp. 981–984, Sept. 2000.
plicated transforms are computed and the algorithm can be [7] A. C. Bovik and S. Liu, “DCT-domain blind measure-
implemented without storing the entire image (or even a row ment of blocking artifacts in DCT-coded images,” in
of pixels) in memory, which makes embedded implementa- Proc. IEEE Int. Conf. Acoust., Speech, and Signal Pro-
tions easier. The basic methodology of the proposed method cessing, vol. 3, pp. 1725 –1728, May 2001.
can also been used to develop NR quality assessment meth-
ods for H.26x/MPEG compressed video. [8] P. Gastaldo, S. Rovetta, and R. Zunino, “Objective as-
A Matlab implementation of the proposed method is sessment of MPEG-video quality: a neural-network ap-
available at http://anchovy.ece.utexas.edu/˜zwang/research/ proach,” in Proc. IJCNN, vol. 2, pp. 1432 –1437, 2001.
nr jpeg quality/index.html . We are also continuing our sub- [9] M. Knee, “A robust, efficient and accurate signal-
jective experiments with more test images, subjects and ty- ended picture quality measure for MPEG-2.” available
pes of image distortions, and will make the test database at http://www-ext.crc.ca/vqeg/frames.html, 2001.
available to the public in the near future.