Presentation on theme: "Filtration based on Color distance"— Presentation transcript:
1 Filtration based on Color distance Filter designColor distanceUniform color spaceHVSHVS based filter design
2 Filter design Median filter: Neighborhood values are: 115, 119, 120, 123, 124, 125, 126, 127, 150Median value is: 124
3 Median filter for color space Approach #1Separate true-color image into color planesApply median filter separately for each color plane.blueredgreen
4 Median filter for color space The drawback of this method is that the separate elements are almost always correlated and such usage of median filter does not utilize this property
5 Well known method Vector median filter Each image pixel is treated as a vector.Case 1: For each pixel within a window calculate vector norm.Case 2: Calculate angle differences between the vectors within a window
6 Vector median filter 3D information is converted into 1D Then processed.
7 Color difference How colors are really different from each other? RGB(255,0,0) – redRGB(255,153,255) – pinkRGB(204,204,255) – violet
11 CIE color space CIE - Commission Internationale de l'Eclairage CIE developed a standard of three imaginary primariesReferred to as XYZ color
12 CIE chromacity diagram Normalized CIE primaries define x, y, zx+y+z = 1This graph is projection on xy plane. (dropping z)
13 CIE chromacity diagram Shows a special projection of 3dCIE color space XYZ.This is the base for all color management systems.The color space includes all distinguishable colors.Many of them cannot be shown on screen or printed.The diagram visualizes however the concept
14 CIE white pointThe black line follows the blackbody spectrum, and is the color carbon glows when heated to the corresponding temperature in Kelvintungsten light (A) SunsetAverage daylight (D65) 10K - blue sky
15 RGB XYZ RGB R = + 2.36461 · X - 0.89654 · Y - 0.46807 · Z G = · X · Y · Z ( 2 )B = · X · Y · ZX = · R · G · BY = · R · G · B ( 1 )Z = · R · G · B
16 Uniform color spaces La*b* color space Where Xn, Yn, Zn define the whitepoint
18 JND Actual size of ellipses is 10 times smaller ∆e=3 visually indistinguishable∆e=5 acceptable error (most printers)∆e=10 bad∆e=15 unacceptable
19 Median filterAt each point of the window calculate difference between the point and backgroundProceed with medianSwap corresponding colors
20 HVS (Human Visual System) Which square is brighter?They have equal luminancesThe reason is that our perception is sensitive to luminance contrast, rather than to absolute luminance.
21 Luminance v.s. Brightness Luminance Brightness(intensity) vs (Lightness)Y in XYZ V in HSVEqual intensity steps:LuminanceDI1DI2I2I1Equal brightness steps:I1 < I2, DI1 = DI2
22 Weber’s law In general, DI needed for just noticeable difference (JND) over background I was found to satisfy :Weber’s Law:Perceived Brightness = log (I)DII⋍ constant=0.02(I is intensity, DI is change in intensity)IntensityPerceived BrightnessThis equation states that equal increments in the log of luminance should be perceived to be equally different.This model partly explains why a uniform level of random noise is more visible in a darker region than in a bright region.
23 HVS filter design Example: Using defined window 3x3, 5x5,… calculate background luminanceConsider different behavior of the filter in darker areas, midtone areas and bright areas.
24 Applications Filtering artifacts introduced by JPEG. Improving quality of scanned images.…
25 ImportantColor and spatial information about the image should not be considered separately.