Colour Edges

Colour edges

Goal: find boundaries between colours of similar brightness, such as pink hair on pale skin, which luminance-based line extraction can't see.

Code: LineExtractor.MarkColorEdges, RgbImage.ToLabChroma

CIE Lab and ΔE

RGB distance doesn't match how different two colours look. CIE L*a*b* separates:

• L*: lightness
• a*: green ↔ red
• b*: blue ↔ yellow

Euclidean distance in Lab, ΔE, approximates perceived difference: ΔE ≈ 1–2 is just noticeable. Using only the a*b* plane measures hue and saturation difference independent of brightness, which is exactly what luminance misses.

Conversion: sRGB → linear (gamma lookup table) → XYZ (D65) → Lab. It costs about 0.17 s on a 3 MP image.

Algorithm

Run the same center-surround comparison on the a* and b* planes. A pixel is a colour-edge candidate when |center_ab − surround_ab| ≥ threshold.

Pitfall 1: a step only gives about 24% of its height

For a line, center-surround responds strongly. For a step between two flat regions, both blurs straddle the edge, and the difference peaks at only

max over d of  Φ(d/σ) − Φ(d/3σ)  ≈ 0.24   (about 1.6σ from the edge; independent of σ)

of the step height (Φ is the normal CDF). The first version used the raw difference as the threshold, so ColorContrast = 10 effectively required ΔE ≈ 40. A test pink/grey pair at ΔE 28.5 wasn't detected. Fix: the option is now the ΔE between the two regions, and internally the threshold is ColorContrast × 0.24.

Pitfall 2: doubled outlines

Most colour boundaries in artwork already have a drawn line. Adding a colour band beside it doubled the outline. Things tried, in order:

1. Put the line on the more saturated side. This failed: the blue background is more saturated than pink hair, so the colour line landed on the opposite side from the luminance line.
2. Compare luminance and chroma response strength per pixel. This failed: a drawn outline is a line in luminance but a step in chroma, so the two responses have different spatial shapes and the ratio isn't consistent across the band.
3. Judge whole pieces (the version kept). Group colour pixels into connected pieces and drop any piece where ≥ 50% of its pixels lie within ceil(3σ)+1 px of a luminance line. That piece is the same boundary seen twice. A genuine colour-only boundary touches luminance lines at most at its ends.

Side rule (kept): put the colour line on the darker side, where a luminance line would be, or on the more saturated side when brightness is equal (|ΔL| < 0.5). Both are where an artist would draw the line, and both keep it one-sided.

Outcome and honest assessment

On the three sample illustrations, colour edges add very little: a short line near an eye and a few specks. These artworks outline nearly every region with a line that's darker than its surroundings, so luminance already catches them. The feature matters for lineless or flat-colour styles, logos, and soft hair-to-skin transitions. The default ColorContrast = 20 is safe: it rarely adds clutter, and a dedicated test proves it finds an equal-luma pink/grey boundary that luminance alone can't.