Fast Exact Euclidean Distance FEED Transformation Theo Schouten

  • Slides: 12
Download presentation
Fast Exact Euclidean Distance (FEED) Transformation Theo Schouten Egon van den Broek Radboud University

Fast Exact Euclidean Distance (FEED) Transformation Theo Schouten Egon van den Broek Radboud University Nijmegen August 2004 FEED

Distance transformation • distance map D(p) = min { dist(p, q), q O }

Distance transformation • distance map D(p) = min { dist(p, q), q O } • approximation of Euclidean • Rosenfeld & Pfaltz – local, parallel or sequential • Borgefors – chamfer, weighted distances August 2004 FEED

Euclidean distance • not by local operations • disconnected Voronoi tile • often right,

Euclidean distance • not by local operations • disconnected Voronoi tile • often right, sometimes wrong ED • correction August 2004 Cuisenaire & Macq CVIU 76 (1999) FEED

Principle of FEED • D(p) = if (p O) then 0 else for each

Principle of FEED • D(p) = if (p O) then 0 else for each q O for each p O D(p) = min ( D(p), ED(q, p)) • inverse of definition • correct, terrible slow August 2004 FEED

Speed up, step 1 • reduce q O to consider • only the border

Speed up, step 1 • reduce q O to consider • only the border pixels of O x x x August 2004 Border: q O at least 1 4 -conn p O FEED

Speed up, step 2 • pre-computation of ED(q, p) • matrix, size of image

Speed up, step 2 • pre-computation of ED(q, p) • matrix, size of image translation, reflection invariant • M = fnon-decr( ED), like square • size can be reduced – in case max. dist. is known – only up to a maximum is interesting August 2004 FEED

Speed up, step 3 • reduce p O to update per B August 2004

Speed up, step 3 • reduce p O to update per B August 2004 FEED

Balance • time lost: – searching object pixels – administration bisection line • against

Balance • time lost: – searching object pixels – administration bisection line • against time gained: – not updating certain p O • optimum, distribution object pixels August 2004 FEED

Results • Shih & Liu 4 -scan ED (PR 31, 1998) – not their

Results • Shih & Liu 4 -scan ED (PR 31, 1998) – not their correction method • test images, object-like images • FEED is faster, up to 2. 7 – up to 4. 5 reduced M • random dot images, faster < 15% • FEED uses less memory August 2004 FEED

Applications • human color categories – black, white, gray, red, green, blue, yellow, brown,

Applications • human color categories – black, white, gray, red, green, blue, yellow, brown, purple, pink, orange • 216 web-safe colors • classify 2563 colors • RGB->HSI, SI: 3 /8, 3, HI: 8 • content based image retieval, texture a August 2004 FEED

Further developments • step 3: faster, simpler • formal proofs • partial maps, fixed

Further developments • step 3: faster, simpler • formal proofs • partial maps, fixed objects + moving objects in video • color space applications August 2004 FEED

FEED conclusions • • • EDT inverse definition simple, correct, slow 3 speed up

FEED conclusions • • • EDT inverse definition simple, correct, slow 3 speed up approaches faster than 4 -scan method up to maximum, partial maps human-centered color space August 2004 FEED