Lossy image compression based on contrast allocation and global precedence

1
Lossy image compression based on contrast
allocation and global precedence

• Damon Chandler and Sheila Hemami
• Visual Communications Lab
• School of Electrical and Computer Engineering
• Cornell University

2
Overview

• Wavelet subband quantization distortions
• Contrast sensitivity and contrast constancy
• Masked detection experiments results
• Application to compression

3
Lossy Compression Additive Noise Model
4
Discrete Wavelet Transform
5
Quantization of Transform Coefficients
6
Contrast Detection
Spatial Frequency (c/deg)
Georgeson et al. 1975
7
Contrast Detection and Constancy
Contrast
Spatial Frequency (c/deg)
Reference grating
Georgeson et al. 1975
8
Contrast Detection and Constancy
_at_ threshold

• Detection thresholds vary w/ frequency.
• Perceived contrast of suprathreshold sine-waves
  shows significantly less frequency variation.

Contrast
Spatial Frequency (c/deg)
Reference grating
Georgeson et al. 1975
9
How to proportion the contrasts?

• Visually lossless compression
• ? Proportion based on detection thresholds?
• (e.g., less contrast for high-frequency
  distortions)
• Visually lossy compression
• ? Scale detection thresholds?
• ? Proportion based on perceived contrast ratios?
• Wavelet distortions, not sine waves.

10
Masked Detection Experiments

• Task Increase contrast of f c/deg distortions
  until they are just visible.
• Stimuli
• f 1.15, 2.3, 4.6, 9.2, and 18.4 c/deg
• ? Horiz., Vert., and Diag. (LH, HL, HH
  subbands)
• 15 natural images maskers
• 8-bit grayscale
• 512x512 pixels

11
Masked Detection for Lossless Compression
12
Masked Detection for Lossless Lossy Compression
13
Masked Detection Results
14
Masked Detection Results
15
Masked Detection Results

• Detection thresholds vary w/ frequency.
• Less frequency dependence at suprathreshold
  contrasts.
• Subbands discarded in a fine-to-coarse-scale
  progression (global precedence).

16
Application to Compression

• Proportion contrasts to preserve global-to-local
  scale-space integration

17
Application to Compression

• Given C(s) , how to choose Q(s) ?
• Relate pixel-value to luminance
• Relate RMS contrast to MSE
• Compute Q(s) from MSE
• Std. Techniques Iteration, Q(s)?(12?MSE)½, etc.

18
JPEG-2000 CSF
JPEG-2000 Contrast-Based
_at_ 0.1 bits/pixel
19
Conclusions

• Selective effects on LFs observed with
  natural-image backgrounds.
• Must consider visual processing of
  image-structure global precedence.
• Visually optimal bit-allocation is accomplished
  by adjusting the contrasts of the distortions.

20
JPEG-2000 DAVPW
JPEG-2000 Contrast-Based
_at_ 0.1 bits/pixel
21
JPEG-2000 DAVPW
JPEG-2000 Contrast-Based
_at_ 0.1 bits/pixel
22
JPEG-2000 DAVPW
JPEG-2000 Contrast-Based
_at_ 0.25 bits/pixel
23
Distortions _at_ f c/deg
Distortions _at_ f /8 c/deg
Total C 0.18
X
X
X
X
o
X
o
X
X
X
24

• Subbands discarded in a high-to-low-frequency
  progression

Subband discarded
25
(No Transcript)
26
Coarse
Intermediate
Fine
27
Coarse
Intermediate
Fine
28
(No Transcript)
29
Coarse
Intermediate
Fine
30
(No Transcript)
31
Coarse
Intermediate
Fine
32
(No Transcript)
33
Contrast Detection
Contrast
Spatial Frequency (c/deg)
Georgeson et al. 1975
34
How to proportion the contrasts?

• Consider how image is visually processed
• Natural images ? many low frequencies
• (1/f amplitude spectra)
• Global Precedence Navon 1977
• Global-to-local temporal analysis
• Visual Scale-space Integration Hayes 1989
• Global-to-local integration of edge-structure