Transferring Color to Greyscale Images Tomihisa Welsh Michael

  • Slides: 24
Download presentation
Transferring Color to Greyscale Images Tomihisa Welsh, Michael Ashikhmin, Klaus Mueller (Stony Brook University)

Transferring Color to Greyscale Images Tomihisa Welsh, Michael Ashikhmin, Klaus Mueller (Stony Brook University) SIGGRAPH 2002

What is Colorization? • Woody Allen: – "colorization" of films "is a `monstrous, '

What is Colorization? • Woody Allen: – "colorization" of films "is a `monstrous, ' disgusting, ' horrible, ' `sinful, ' `absurd, ' `humiliating, ' `preposterous, ' and `insulting' mutilation and defacing of genuine works of art, in which computers are used to `doctor' and `tamper' with the `great originals, ' thereby creating `degraded, ' `cheesy, ' artificial symbols of one society's greed. "

What can it do? • “Colorize” old movies • Make black and white pictures

What can it do? • “Colorize” old movies • Make black and white pictures color • Adds a dimension to MRI’s or airport luggage scans • Lifelike electron microscope scans

Why is it difficult? • Grayscale images consist of one dimensional data – Luminance

Why is it difficult? • Grayscale images consist of one dimensional data – Luminance data – No Saturation or Hue • 3 D HSL colors have 256 x 256 or ~66, 000 colors with a given Luminance

It’s still not easy… • Different sections of an image might have same lumination

It’s still not easy… • Different sections of an image might have same lumination but very different hue.

But let’s give it a shot • We need a color space • HSL

But let’s give it a shot • We need a color space • HSL and RGB just wont cut it • Domo Arigato Mr. Ruderman et. Al. 1998 – Created l, a, B color space – Decorellated space • • Linearly independent Luminance ( l ) Yellow-blue ( a ) Red-Green ( B )

Come again? • Conversion: • LMS = Long, medium, and short wavelength • Now

Come again? • Conversion: • LMS = Long, medium, and short wavelength • Now that we have our new color space, we turn to some sadis…I mean statistics

Scale and correct l, a, and B

Scale and correct l, a, and B

RGB to LAB

RGB to LAB

Luminance Mapping • Source image = color image • Target image = grayscale image

Luminance Mapping • Source image = color image • Target image = grayscale image

Luminance Mapping part Deux • Hertzmann et. Al. 2001 – Standard deviation of 5

Luminance Mapping part Deux • Hertzmann et. Al. 2001 – Standard deviation of 5 x 5 pixel neighborhood – Match source and target based on combination of luminance (50%) and std. dev. (50%) – Map the match

Map it Source Luminance Channel Target Remapped

Map it Source Luminance Channel Target Remapped

Houston… • There is a problem with this. • What if there are different

Houston… • There is a problem with this. • What if there are different sections of the picture?

Swatches • Indicate similar parts of a scene

Swatches • Indicate similar parts of a scene

Examples

Examples

More Examples

More Examples

Video • • • Why not extend this to video? Industry born in 1970

Video • • • Why not extend this to video? Industry born in 1970 (Wilson Markle) Recolorized movies use these techniques Source frame used every camera change Once have first target, use that as source User-defined swatches can track movement

Waves

Waves

Horses

Horses

This is your brain

This is your brain

At what cost? • In 1987 $3, 000/min or $300, 000/movie • One Critic

At what cost? • In 1987 $3, 000/min or $300, 000/movie • One Critic calls it the “Bastardization” of film • Brings in on average $500, 000/release

Oh, Cool

Oh, Cool

Wrap-Up • • • Get source and target Convert colors to l, a, B

Wrap-Up • • • Get source and target Convert colors to l, a, B Map colors Use swatches for faster better results Process takes between 15 seconds and 4 minutes on Pentium III 900 MHz CPU • Use MATLAB