IMAGE QUALITY ASSESSMENT: FROM ERROR MEASUREMENT TO STRUCTURAL SIMILARITY

1

• IMAGE QUALITY ASSESSMENT FROM ERROR MEASUREMENT
  TO STRUCTURAL SIMILARITY
• Z. Wang, A. C. Bovik,
• H. R. Sheikh, and E. P. Simoncelli
• IEEE Transactions on Image Processing
• Vol. 13. No. 4, April 2004

2
INTRODUCTION

• Digital images are subject to a wide variety of
  distortions during
• Acquisition
• Processing
• Compression
• Storage
• Transmission
• Reproduction
• In some applications, the distorted images are
  viewed by human observers.
• Subjective evaluation is inconvenient,
  time-consuming, and expensive.
• The goal of research is to develop a quantitative
  measure that can automatically predict the
  perceived image quality.
• Classification of quantitative measures
• Full-reference
• Non-reference
• Reduced-reference
• MSE and PSNR are appealing because they are
  simple to calculate, have clear physical
  meanings, and are mathematically convenient.

3
STRUCTURAL SIMILARITY BASED ON IMAGE QUALITY
ASSESSMENT

• Natural images are highly structured
• Their pixels exhibit strong dependencies,
  especially when they are spatially close.
• The Minkowski metric
• is based on pointwise signal differences, which
  are independent of the underlying signal
  structure.
• Although most quantitative measures are based on
  the Minkoswki metric, they do not provide a good
  correlation with human observation.
• The assumption is that the HVS is highly adapted
  to extract structural information from the
  viewing field.
• The error sensitivity approach estimates
  perceived errors to quantify image degradations.
• The new approach considers image degradations as
  perceived structural information variation.

4
BOAT DISTORTED WITH DIFFERENT TYPES OF DISTORTIONS
All the distorted images have the same MSE values
but different MSSIM values (b) 0.9168 (c)
0.9900 (d) 0.6949 (e) 0.7052 (f) 0.7748
5
THE UNIVERSAL QUALITY INDEX, UQI

• ,
  where x and y are vectors, representing the
• original and distorted signals, and their
  elements are .
  
• 
  

Overall quality index
6
THE STRUCTURAL SIMILARITY (SSIM) INDEX

• UQI produces unstable results when either of the
  two terms in the denominator becomes very close
  to zero.
• Luminance comparison
• Contrast comparison
• Structure comparison
• 
  

To simply the expression, ???1 and C3C2/2
7
TEST ON JPEG AND JPEG2000 IMAGE DATA BASE

• 29 high-resolution 24bits/pixel RGB color images
  were compressed.
• 175 JPEG images
• 169 JPEG2000 images
• Bits rates for JPEG images 0.150 0.3.336
  bits/pixel
• Bits rates for JPEG2000 images 0.028 3.150
  bits/pixel
• Subjects viewed the images from comfortable
  seating distances.
• They were asked to use a scale with bad, poor,
  fair, good, and excellent.
• Each JPEG image was viewed by 13-20 subjects.
• Each JPEG2000 image was viewed by 25 subjects.
• The subjects were mostly male college students.
• In the experiments, only the luminance layer (Y)
  of the YUV color model is used.
• The two chrominance layers (U and V) do not
  significantly change the performance of the model.

8
SCATTERS PLOTS OF FOUR IMAGE QUALITY MEASURES
9
PERFORMANCE COMPARISON OF FOUR IMAGE QUALITY
MEASURES
10
CONCLUSIONS

• The traditional approach to image quality
  assessment based on error sensitivity (i.e., the
  Minkowski error metric) is not good!
• Four image quality assessment models are used
• PSNR
• Sarnoff
• UQI
• MSSIM
• SSIM index compares favorably with other 3
  models.
• The scatter plot for MSSIM appears to be the
  best.
• In Table I, MSSIM is better than the other
  models.
• In the experiments 344 JPEG and JPEG2000 images
  were used.