NOREFERENCE PERCEPTUAL QUALITY ASSESSMENT OF JPEG COMPRESSED IMAGES

1

• NO-REFERENCE PERCEPTUAL QUALITY ASSESSMENT OF
  JPEG COMPRESSED IMAGES
• Z. Wang, H. R. Sheikh and A. C. Bovik Laboratory
  for Image and Video EngineeringDepartment of
  Electrical and Computer EngineeringThe
  University of Texas at Austin, Austin, Austin, TX
  78712
• Proceedings of IEEE 2002 International Conference
  on Image Processing
• Rochester, NY, September 22-25, 2002

2
INTRODUCTION

• In recent years, there has been an increasing
  need to develop objective measurement techniques
  that can predict image/video quality
  automatically.
• Such methods an have various applications
• They can be used to monitor image/video quality
  for quality control systems.
• They can be employed to benchmark image/video
  processing systems and algorithms.
• They can be embedded into image/video
  processing systems to optimize algorithms and
  parameter settings.
• The most widely used objective measures are PSNR
  and MSE.
• However, they do not correlate well with
  perceived quality measurement.
• Most of the objective measures require the
  original image as a reference.
• Human observers are able to assess the quality of
  distorted images without using any reference
  image.
• Designing an no-reference (NR) objective measure
  is a difficult task.

3
SUBJECTIVE EXPERIMENTS

• The subjective test was conducted on 8 bits/pixel
  gray level images.
• There are 120 test images in the database.
• 30 of them are original images.
• These 30 images are randomly divided into two
  groups.
• Each group contains 15 images.
• The rest of the images are JPEG compressed
  using Matlab.
• The quality factors are selected randomly between
  5 and 100.
• The resulting bits rates range from 0.2 to 1.7
  bits/pixel.
• 53 subjects were shown the database.
• Most of the subjects were college students.
• The subjects were asked to assign each image a
  quality score between 1 and 10 (10 is the best,
  1 is the worst).
• the 53 scores were averaged to obtain a MOS.

4
TWO GROUPS OF IMAGES

5
PSNR VS. MOS
Correlation coefficient 0.3267

6
JPEG COMPRESSION

• JPEG is a block DCT-based lossy image coding.
• JPEG works as follows
• DCT is applied to 8x8 blocks.
• Each block goes though quantization and entropy
  coding.
• Both blurring and blocking artifacts may be
  created during quantization.
• The blurring effect is mainly do to the loss of
  high frequency DCT coefficients.
• The blocking effect occurs due to the
  discontinuity at block boundaries.
• One effective way to examine both blurring and
  blocking effects is to transform the signal into
  the frequency domain (DFT).
• A disadvantage of the frequency domain method
  is the involvement of the Fast Fourier Transform
  (FFT). This is expensive.
• FFT also requires more storage.

7
POWER SPECTRUM IN THE DFT DOMAIN
The blocking effect can easily be identified by
the peaks at frequencies 1/8, 2/8, 3/8, and
4/8. The blurring effect is characterized by the
energy shift from high frequency to low frequency
bands.

Test image signal x (m,n), where m ? 1,M and
n ? 1,N. Calculate a differencing signal along
each horizontal line dh(m,n)x(m,n1)-x(m,n),
where n ?1,N-1. fm(n)dh(m,n) 1-D
horizontal signal for a fixed value of
m. Compute the power spectrum of fm(n) for
m1,,M, and average them together.
8
OBJECTIVE NR QUALITY ASSESSMENT

• The authors attempt to design a computationally
  inexpensive and memory efficient feature
  extraction method.
• The features are calculated horizontal and then
  vertically.
• First, the blockiness is estimated as the average
  difference across block boundaries.
• Second, the activity of the image signal is
  estimated.
• The activity is measures using 2 factors
• The first is the average absolute difference
  between in-block image samples
• The second activity measure is the
  zero-crossing (ZC) rate.
• We define for n ? 1N - 2

9
OBJECTIVE NR QUALITY ASSESSMENT

• The horizontal ZC rate then can be estimated as
• Using similar methods, we calculate the vertical
  features
• of Bv, Av, and Zv.
• The overall features are given as
• There are many ways to combine the features.
• One method that gives good prediction performance
  is

10
SCATTER PLOTS FOR GROUP I AND GROUP II IMAGES

11
SCATTER PLOT FOR BOTH GROUPS OF IMAGES

12
CONCLUSIONS

• A novel NR perceptual quality assessment method
  for JPEG compressed images is presented.
• Subjective evaluation was conducted to evaluate
  the quality of JPEG compressed images.
• The features described effectively capture the
  artifacts introduced by JPEG.
• The agreement between PSNR and MOS is not good.
• Nonlinear curve fitting gives good agreement with
  MOS.
• The method is computationally efficient.
• No complicated transforms.
• The algorithm can be implemented without string
  the entire image in memory.