Color spaces CIE RGB space HSV space CIE

Color spaces CIE - RGB space. HSV - space. CIE - XYZ space. L*A*B* - space. YUV, YIQ CMY, CMYK

CIE-XYZ Gamut • Only the area between R, G, B can be reproduced by R, G, B primaries.

Considerations in determining the R, B, G primitives • Producing a wide range of colors • Dynamic range considerations: don’t use colors that you don’t need. • “System” considerations: Colors that are easy to produce by color monitors.

cia pa n in t The CMY(K) model

Perceptual Models • L*ab color space normalizes the color space such that Euclidian distances will fit the perceptual ones (using JND) • For example: Human vision has a nonlinear perceptual response to brightness. • In general, I needed for just noticeable difference (JND) over background I was found to satisfy: I /I = const • Weber’s low: lightness perception is roughly logarithmic. • This is comparable to the L* component of the perceptual L*ab color model:

Color segmentation First try Segment the image according to its color histogram- Find “Clusters” of colors in the RGB space (or make a color quantization) Problem: In the RGB space, there is information which is not relevance to the chromaticity : lightening, texture and shadows. We need to eliminate the intensity !

Color segmentation (cont’) Solution: Convert the RGB image to another image space containing an intensity component- omit this component from the segmentation process. Examples: YIQ: Omit Y HSV (Hue, Saturation, Value): Omit V Original image Color segmentation using the HSV model

Skin Color Task: We would like to detect faces according to their color histograms Skin Color: The chromaticity of skin is very restricted (mainly the “Hue” component )* [*Skin Color is due to the amount of the pigment melanin]

Skin Color (cont’) Hue: Saturation: Value:

White balancing Problem: The color image is effected by the color of the light. Example: Outdoor scenes are “more blue” than indoor scenes. Possible solution: (far from perfect) Perform white balancing: find “white” regions, and change the color map such that these regions will become real white. Sometime, color skin can also be used for this purpose.

Simplified Physical Model: • An image is a function of many parameters: Lighting Geometry Surface albedo Image Viewing

Simplified Physical Model: • It is common to divide it into Lambertian components and specular ones. • Lambertian: The light is returned to all direction. • Specular: The light is returned in approx’ one direction. • Most objects are mainly Lambertian, but have a small Specular component.

Simplified Physical Model: • For Lambertian objects, the specular component is zero therefore we have a linear function. • Pixels belonging to a region with homogenous color should lie upon a line throw the origin in the RGB histogram. • With the Specular component, these pixels will lie on a plane (but most of the pixels will still lie on the original line)

The T-Shape model. • The T shape model introduced by Klinker et al. is widely used to model specularity. • The model assumes a large n in the previous equation => for each pixel there is only one dominant component 14

The color line model An ongoing work of Ido Omer and Mike Werman.

“Real Histogram” properties. • The lines best describing the color clusters don’t intersect the origin.

Cut Off: • One of the possible causes for the inaccuracy of the linear model is the “cut off “ phenomena in image sensors.

Looking at the histogram:

Looking at the histogram: 19

Looking at the histogram: 20

Looking at the histogram: 21

Comparing color segmentation using different color models. Original HSV Lab Color Lines

Color segmentation with “color lines” • slice the histogram perpendicularly to the origin. • search for local maxima. • combine these maxima to color lines

Color segmentation with “color lines” • Since we look only at the histogram, we are not effected by local image properties like texture. • The number of colors in the original image > 80, 000, yet it has been described using ~40 lines. • Conclusion: The color histograms of natural images are very sparse.

Other distortions… Be aware: Most cameras apply various color enhancements that distorts the linear color model.

Talking about mosaicing, the “opposite” problem also exists. • Most digital cameras use filter arrays to sample red, green, and blue according to the Bayer pattern or similar ones. • At each pixel only one color sample is taken, and the values of the other colors must be interpolated from neighboring samples. 26

Demosaicing. • Many demosaicing techniques refer to the green channel as the “detail channel “ and to the red/blue channels as chrominance channels. • These techniques start by interpolating green values and then interpolate red/blue values according to the green one. 27