Evolving Color Constancy

1
Evolving Color Constancy

• Marc Ebner
• Universität Würzburg, Germany
• Pattern Recognition Letters 27(2006)1220-1229
  Elsevier

2
Algorithms for color constancy

• Gamut constraint methods
• Perspective color constancy
• Color by correlation
• The gray world assumption
• Recovery of basis function coefficients
• Mechanisms of light adaptation coupled with
  movements
• Neural networks
• Comprehensive color normalization
• Committee based methods
• Algorithms based on the dichromatic color model
• Computation of intrinsic images

3
PE(Articial Retina)

• PEa rectangular grid of processing elements
• Better than neural nets, quite complicated.

4
Processing elements

• 1 PE for 1 image pixel
• 3 layers of PEs carrying out results on the 3
  image bands red, green, and blue.
• Estimate of the illuminant (color of input
  pixel)
• The data from other neighboring PEs
• Initially, ( pixel value )

5
(No Transcript)
6
Conclusion

• Only the current color channel(band)is used.
• Average data from neighboring elements.

7
Parallel algorithm

• The gray world assumption
• The reflectance ,
• distributed over the interval 0,1
• From PE,
• Nthe number of image pixels.

8
Parallel algorithm

• a(x, y)an estimate of local space average color
  for each image pixel
• N(x, y)a set of neighboring elements
• (1)Average the data
• (2)Slowly add the color of the current
  pixel(psmall percentage)

9
Parallel algorithm

• The two equations, (1)(2),are carried out until
  convergence.
• 1000, 2000, 3000, 4000, 5000

10
(No Transcript)
11

• Local space average color
• 1 50 200 1000
• The parallel algorithm
• 1000

12
Reference

• Ebner, M., 2001. Evolving color constancy for an
  artificial retina. Genetic Programming Proc. of
  the 4thEuropean Conference, EuroGP 2001, Lake
  Como, Italy. Springer-Verlag, Berlin, pp. 1122.
• Ebner, M., 2004. A parallel algorithm for color
  constancy. J. Parallel Distributed Comput. 64
  (1), 7988.

13
Why Mondrian has been chosen

• First introduced by Edwin Land
• No curve and angle. No shade and texture.
• Neither uniformly colored nor uniformly bright.
• Resemble better the more colorful work of Klee or
  Lohse.
• Anya Hurlbert, 1999

14
Paul Klee

• ?????????  Tunisian Gardens1919
• Ref. www.writedesignonline.com/history-culture/bau
  haus.htm

15
Richard Paul Lohse

• Thematic series in 18 colours A, 1982
• Squares formed by colour groups 1944/2
• Ref. www.lohse.ch/bio_e.html

16
Mondrian

• Piet Mondrian, Composition A, 1923
• www.cartage.org.lb/en/themes/Arts/painting/20th-ce
  ntury/art-sake/artsake.htm

17
Typical Mondrian stimuli

• Yellowish daylight bluish daylight
• 2 grey papers(third from the top on the left)

18
The experiment of Kraft and Brainard

• Look through a window into a box
• A grey test surface against the back wall
• A Mondrian-like panel
• A tube wrapped in
• tin foil
• A cube, pyramid and
• tube made from grey
• cardboard

19
Local surround

• Neutral-illuminant Orange-red

20
Spatial Mean

• Neutral-illuminant pale-red

21
Maximum Flux

• Neutral-illuminantyellow-illuminant

22
Results

• Color constancy

23
Anya Hurlbert, 2007

• Unknown why humans need color constancy. Color?
  Shape?
• How is color constancy measured? with difficulty.
  Mondrians?
• How is color constancy achieved? More than one
  mechanism. Color processing in the brain.
• Retinex

24
Reference

• Hurlbert A (1999) Colour vision is colour
  constancy real? Current Biology 9R558R561.
• Hurlbert, A. (2007). Colour constancy. Current
  Biology, 17(21), R906-7.
• JM Kraft and DH Brainard, Mechanisms of color
  constancy under nearly natural viewing. Proc Natl
  Acad Sci USA 96 (1999), pp. 307312.