A Comparison of Still-Image Compression Standards and Proposed Methods for Improving Lossy Image Quality

1
A Comparison of Still-Image Compression Standards
and Proposed Methods for Improving Lossy Image
Quality

• MDDSP Literature Survey Presentation
• Eric Heinen

2
Project Goals

• Investigate existing methods for improving JPEG
  image quality
• Consider adapting these methods to JPEG2000
• Novel approaches?
• Compare image quality of JPEG and JPEG2000 (with
  without improvements) across
• Image set
• Compression ratios

3
Joint Thresholding and Quantizer Selection for
JPEG Encoding(M. Crouse and K. Ramchandran, 1997)

• Optimize encoding without affecting decoding
• Rate-Distortion optimization framework

4
Improved JPEG Compression Using Human Visual
System (HVS) Model(G. Sreelekha and P. S.
Sathidevi, 2005)

• Exploits contrast sensitivity of HVS
• Sinusoid at given frequency requires certain
  amount of contrast to elicit a response
• HVS is insensitive to very low and very high
  frequencies
• Contrast Sensitivity Function (CSF)
• CSF Thresholding
• Discard DCT coefficients below CSF value

5
Improved JPEG Compression Using Human Visual
System (HVS) Model(G. Sreelekha and P. S.
Sathidevi, 2005)

• Also exploits masking property of HVS
• Vision less sensitive to local variation in
  brighter regions
• Compute luminance masking threshold based on
  blocks DC coefficient
• DCT coefficients below this threshold discarded
• Quantization after thresholding
• different from JPEG standard
• incompatible with JPEG standard decoder

6
JPEG2000 vs. JPG

• Arbitrary rectangular tile size allowed
• Color space transformation can be
• Lossless (YUV)
• or lossy (YCbCr)
• Discrete Wavelet Transform (DWT)
• More sophisticated entropy encoding
• Improvements to JPEG can be applied

7
Image Quality Assessment

• Full-reference image quality metrics (IQM)
• Estimate DMOS
• Convenient
• Popular IQMs
• Peak Signal to Noise Ratio (PSNR)
• Universal Image Quality Index
• Structural Similarity (SSIM)

8
Statistical Evaluation of Recent IQMs(H. R.
Sheikh, M. F. Sabir, and A. C. Bovik, 2006)

• Image database
• Set of distortions
• JPEG, JPEG2000, White Noise, Gaussian Blur,
  simulated wireless channel
• Collect mean opinion scores
• 3 different performance metrics
• RMSE between IQM score and DMOS after non-linear
  regression

IQM RMSE
PSNR 13.4265
JND 10.2649
DCTune 16.2150
PQS 10.4221
NQM 11.4698
Fuzzy S7 15.3544
BSDM (S4) 9.7934
SSIM (MS) 9.3691
IFC 9.0007
VIF 8.2459
9
Questions?