Measurement of Ringing Artifacts in
JPEG Images*
Xiaojun Feng
Jan P. Allebach
Purdue University - West Lafayette, IN
* Research supported by the Hewlett-Packard Company
EI 2006 - San Jose, CA
Slide No. 1
Motivation
 
Applications
  Image quality assessment through measuring JPEG
artifact perceptibility
  Automated workflow for variable data printing
 
Goal
  Develop objective no-reference measurement of visual
impact for ringing artifacts in JPEG compressed
images.
EI 2006 - San Jose, CA
Slide No. 2
Prior Art
 
Full-reference approach – Marziliano et al.
(2004)
  Usually perfect reference image is not available
 
No-reference approach – Oguz (1999)
  Texture of ringing artifacts is not compared with a
neighboring smooth region
EI 2006 - San Jose, CA
Slide No. 3
Outline
 
What is ringing artifact
 
Proxy object
 
System detail
 
Sample results
 
Conclusions and future work
EI 2006 - San Jose, CA
Slide No. 4
Overview of ringing artifacts
 
Characteristics of ringing artifact
  Ringing artifact is seen to be noise-like variations in the
vicinity of major edges.
  Activity of the ringing region is higher than that of
neighboring smooth region
 
Source of ringing artifact
  DCT coefficient quantization
EI 2006 - San Jose, CA
Slide No. 5
Example of ringing artifact
JPEG Compressed Image*
EI 2006 - San Jose, CA
Enlarged Lighthouse Top
Slide No. 6
* Image source: http://www.beaconphotographics.com/AmericanLights.htm
Ringing artifact measurement
 
Compare the activities of the ringing artifact and
the neighboring smooth region
 
Two masking effects are taken into account in
the model:
  Texture masking
  Luminance masking
EI 2006 - San Jose, CA
Slide No. 7
Proxy smooth object
 
Isolated ringing regions need proxy objects with which activities
may be compared
 
Proxy object is the smooth object to which ringing region
belongs
 
Proxy object can be assigned to ringing region by color similarity
Proxy
object
Isolated ringing
region
EI 2006 - San Jose, CA
Slide No. 8
Ringing region
with smooth
neighborhood
Ringing artifact detection
Step 1:
Edge
detection
JPEG
image
Step 2:
Smooth
region
clustering
Step 3:
Ringing
proxy
assignment
Step 4:
Local ringing
Feature
calculation
 
Step 1: Detect edges using Sobel operator
 
Step 2: Cluster smooth regions into different object classes
according to their color and texture similarity
 
Step 3: Assign a proxy class to each ringing region
 
Step 4: For each edge pixel (x,y), compute a local ringing feature
by pooling over the visibility of ringing regions in a local window
centered at (x,y)
EI 2006 - San Jose, CA
Slide No. 9
Ringing
map
Smooth region identification
 
Segment an image into
  Edges
  Potential ringing regions – regions surrounding edges
  Smooth regions – regions other than edges and ringing artifact
JPEG image
Edge / ringing / smooth segmentation
Edge
Ringing region
Smooth region
EI 2006 - San Jose, CA
Slide No. 10
Smooth region clustering
Smooth
regions
Color
clustering
Map
fusion
Texture
clustering
EI 2006 - San Jose, CA
Slide No. 11
Overall
clustering
Region activity calculation
 
The activity of a region is computed based on the
luminance changes of neighboring pixels.
 
- set of neighboring pixel pairs for region
 
-
EI 2006 - San Jose, CA
channel value for pixel
Slide No. 12
Luminance masking
 
Noise perceptibility is affected by
background luminance. Noise
with very bright or very dark
background is not easily
detected
 
Chou measured just-noticeable
noise levels under various
background luminance (1995)
 
We adopt Chou’s model and
convert the noisy level threshold
into region activity threshold
EI 2006 - San Jose, CA
Slide No. 13
Ringing visibility feature
 
 
 
For a ringing region
, the overall masking effect is involved as:
 
− region activity of proxy object (texture masking)
 
− luminance masking function
Ringing visibility feature for region
 
− size of region
 
− size of a block (64 for JPEG)
Local ringing feature of edge pixel
EI 2006 - San Jose, CA
Slide No. 14
Example of ringing visibility
feature
JPEG image
EI 2006 - San Jose, CA
Activity contrast
Slide No. 15
Sample Result
JPEG image
EI 2006 - San Jose, CA
Ringing map
Slide No. 16
Sample result
Quality factor=90
Quality factor=70
Quality factor=50
Quality factor=30
Compression ratio increases
EI 2006 - San Jose, CA
Slide No. 17
Conclusion and future work
 
Our ringing artifact measurement
  Uses no reference image
  Involves both texture masking and luminance masking
effects to measure the visual impact of ringing artifact
 
Future work
  Conduct psychophysical experiment
  Generate a global ringing value from the ringing map
EI 2006 - San Jose, CA
Slide No. 18
Thank you!
EI 2006 - San Jose, CA
Slide No. 19