Sharpening from Shadow:

Colour From Grey by Optimized Colour OrderingArash VahdatMark S. Drewavahdat@cs.sfu.camark@cs.sfu.caSchool of Computing ScienceSimon Fraser University

November 20101OutlineProblem definitionGrey to Colour TransformationOur SolutionParametric CurveOptimizationResultsConclusion22Problem DefinitionThe problem is to recover colour image from grey level image using the minimum amount of information.

3

( + extra information )Encoder SideDecoder Side3Problem Definition- ContOur ProblemColorization ProblemColour image at encoder side.No user interaction.No colour image as input.Colour hints are provided by human.4

* Images from Drew & Finlayson (ICIP 08)

( + extra information )4Colour from Grey is Hard!There are many colours that can be assigned to a single grey level value.5Y= 0.299 R + 0.578 G + 0.114 B

RBG

u=(0.299,0.587,0.114)RGB to Grey

Ye.g., simple definition from Multimedia:5SolutionAssume each grey level value represents particular fixed point in colour space.6

RBGabcg-1gg+1dEncoder: Colour to GreyFor each pixel in colour image we assign grey value of closest fixed point in colour space as its grey value.(rd ,gd ,bd )gg-1RGBGreyra ,ga ,bag-1rb ,gb ,bbgrb ,gb ,bbg+1Decoder: Grey to ColourFor each grey value use designated colour for that value.(ra ,ga ,ba )g-16ProblemsBoth procedures in Encoder and Decoder add error to recovered colour image.

7(rd ,gd ,bd )gg-1(ra ,ga ,ba )We need to encapsulate colour lookup table with the data, which is overhead.

Our Solution: A Parametric Curveminimize error by tuning parametersattach a few parameterParametric Curve: C(g) : maps grayscale values to colour points. The curve should traverse different regions of colour space.perceptual colour difference is reflected well in CEILAB Colour space.8

RBGabcg-1gg+1u=(0.299,0.587,0.114)

a *b*L*

abcg-1gg+1

Parametric Curve: 9

9OptimizationColour to Grey:for pixel p with colour (p) approximated grey scale value is:10Grey to Colour:use the corrospondonding colour point on the curve.Minimize Error:

Results11

input imagegrey level imageour grey level imageour recovered colour image

Results12Gamut encompassed by parametric curve

GIF paletteOrdered Colours along the curveL*a*b*

Results13

3 bits4 bits6 bits8 bits3 bits4 bits6 bits8 bits

our methodGIFResults14

3 bits4 bits

6 bits8 bitsour method

3 bits4 bits6 bits8 bitsGIFResults15

input imagegrey image

Colour output with 4 bppColour output with 8 bppResults16

input imagegrey imageColour output with 4 bppColour output with 8 bpp

CIELAB errors17

Grey imageColour imageConclusionsWe propose a novel method to reconstruct colour from greyscale images, by optimizing a mapping from greyscale to colour using a parametric curve.Almost always, grey version is better than GIF.The colour image has comparable or lower error especially for low bitrate.Future work:non-constant quantization rate.different curve form. 18Questions?Thank you.19Thanks!To Natural Sciences and Engineering Research Council of Canada