Color Balancing Using Color by Correlation

1
Color Balancing Using Color by Correlation

• PSYCH221
• Image and Vision Systems Project
• by
• Srinivasa Rangan

2
Color Balancing

• Images viewed under different illuminants look
  different
• Global adjustment of color intensities

3
Color Balancing

• Applications
• Machine vision, object recognition
• Digital Photography
• Algorithms
• Gray World, White balance
• Sensor Correlation, Bayesian model
• Color by correlation
• Proposed by Finlayson et. al

4
Color by Correlation Method

• Based on illuminant classification
• Set of commonly occurring indoor/outdoor
  illuminations.
• Partition chromaticity space into distinct
  regions
• Color space is of finite extent (8 bits)
• Device measurement error, eye sensitivity etc.
• Chromaticity space taken as x(1/3), y(1/3)
• Ensures division into uniform regions

5
Probability Matrix

• Build probability matrix M
• Each entry denotes
• log( Pr( chromaticity i illuminant j) )

6
Color balancing an image

• To color balance an image
• Create a chromaticity vector v
• v x1y1 x1y2 . xnyn
• xiyj
• 1 if chromaticity level xi yj occurs in the image
• 0 otherwise
• Maximize p( illuminant i Cim )
• Bayes rule
• p( illuminant i Cim ) p(illuminant i)
  p( Cim ill i)
• Compute v M
• indicative of Pr( illuminant i Cim )

7
Color balancing an image

• Choose most probable illuminant
• Once illuminant is known apply appropriate linear
  transformation to image colors

8
Creating M

• Take wide range of surfaces
• 1995 surfaces from SimonFraser data set
• Find response of each surface under illuminants
  of interest
• Count the number of occurrences of each
  chromaticity interval
• Obtain the probability of chromaticity intervals
  given illuminant

9
Results Synthetic Images

• Choose 18 illuminants available in ISET as set of
  possible illuminants
• Randomly select a set of surface patches and
  illuminants
• Apply Gray World and Color by correlation color
  balancing techniques to generated images
• Use ?E LAB metric for evaluation
• Compare image rendered under daylight (D65) with
  color balanced image

10
(No Transcript)
11
(No Transcript)
12
Real World Examples
13
Gray World
Color By Correlation
14
(No Transcript)
15
Gray World
Color By Correlation
16
(No Transcript)
17
Gray World
Color By Correlation
18
Conclusions

• Color by correlation seems to perform better than
  Gray World
• Estimating the class of the illuminant seems
  more important than estimating the actual
  illuminant.
• The algorithm is susceptible to some of the same
  problems as gray world for images that dont have
  many surfaces.

19
Thank You
20
References

• G. D. Finlayson, P. H. Hubel, and S. Hordley,
  Color by Correlation, Proc. IST/SID Fifth
  Color Imaging Conference Color Science, Systems
  and Applications, pp. 6-11
• D. H. Brainard and W. T. Freeman, Bayesian color
  constancy, Journal of the Optical Society of
  America A, vol. 14, pp. 1393-1411, 1997

21
(No Transcript)