Colour Constancy

T.W. Hung

Colour Constancy – Human

• A mechanism enables human to perceive constant colour of a surface over a wide range of lighting conditions.

• Land studied and attempted to explain the mechanism, Retinex theory.

Colour Histogram 1 (picture 3)Image under reference lighting condition

Colour Histogram 2 (picture 4)Image under biased lighting condition

Colour Constancy - Machine

Problem:• Model What form of a transform used in the

algorithm is sufficient to discount the effect of change of illuminant

• MechanismHow to compute the parameters of the desired

transformation

Common Assumption

• A set of squared colour patches

• Reasonable number of different coloured patches

• Surface property is Lambertian

• Single source of illumination

• Even illumination over the scene

Model

• Coefficient Model [Von Kries, Finlayson]

Ci,j,1 = T Ci,j,2

Mechanism

• White patch algorithm Maximum to predefined maximum

• Grey world algorithm Average to predefined average

• Gamut mapping methodFeasible mapping with Surface colour constraints, illumination constraints

• Specular highlight detection methodDetect the presence of the specular highlight

• Retinex algorithmAverage of logorithmic ratio of surface i and its neighborhood colours

“All objects that are known to us from experience, or that we regard as familiar by their colour, we see through the spectacles of memory colour”

Hering 1878

Problem

To compute a transformation to match the colours in the image to a database of ‘known’ colours.

1. Form of transformation

2. Fitness measure for the matching criteria

3. Matching method

Mahanobis eqn.

Experimental setting

Graph of lighting

Picture

RED(database)GREEN (database)BLUE(database) RED (under bias illuminant)GREEN (under bias illuminant)BLUE (under bias illuminant)RMS Diff.

144 176 92 51 127 57 110.792599196 200 148 68 139 100 149.6963593208 176 136 72 125 95 150.9238218188 164 160 65 119 113 139.1509971144 108 104 52 78 72 101.9215384160 116 140 57 85 99 115.112988148 188 164 54 133 109 122.0081964104 116 180 42 85 120 91.67878708104 148 124 41 109 89 81.9451035888 132 104 38 98 75 67.0596749292 148 204 39 110 140 91.37286249

224 204 212 79 141 143 172.4963768132 152 148 50 113 100 102.708324980 80 136 36 52 97 65.1229606292 88 100 39 60 68 67.9485099296 76 84 40 47 57 68.60029154

216 80 96 74 51 66 148.0033783224 168 216 78 123 145 168.4695818144 156 208 52 115 132 126.178445188 100 156 38 72 11 155.9134375

Conclusion

• Description of the colour constancy problem: model and mechanism

• Description of some of the existing algorithms

• Simple linear transformation

• Minimum error matching criteria

• Match to ‘normalised’ colour model

• Adaptation to relative colour match, seems robust