Tschumperl´ e, Olivier L´ ezoray GREYC, CRNS UMR 6072 ´ Equipe Image - ENSICAEN 6, boulevard du Mar´ echal Juin 14050 Caen cedex {maxime.daisy, david.tschumperle, olivier.lezoray}@ensicaen.fr 20th of June, 2013
be removed, content. • Kind of interpolation. • Patch-based inpainting: • Usage of pieces of image for the reconstruction. • Large area reconstruction. • Artefact apparition during reconstruction. Problem : How to remove seams from a reconstructed image ?
Conclusion Artefact Detection Break detection • Patch-based inpainting result artefacts • Sharp luminosity variation • Texture changes • Detection from only two dataset I and U ? • High image gradient in I Problem ⇒ Real image contour ? • High divergence in U Problem ⇒ Flat area ? • High gradient in I and high divergence in U ⇒ Allow to detect probable incoherence Figure: Positionmap explanation Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 12 / 44
Conclusion Artefact Detection Image Figure: Patch-based inpainting result from which detect artefacts. Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 13 / 44
Conclusion Spatial Patch Blending Blending principle in four images Figure: First image. Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 17 / 44
Conclusion Spatial Patch Blending Blending principle in four images Figure: Second image. Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 18 / 44
Conclusion Spatial Patch Blending Blending principle in four images Figure: Overlapping region exists (necessary condition). Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 19 / 44
Conclusion Spatial Patch Blending Entry-level Method Ji (p) = c∈C w(c, p) Ψi c (p − c) ε + c∈C w(c, p) (1) with w(c, p) = e d1(c,p) σ(p)2 (2) and d1 (c, p) = min q∈ψc q − p 2 (3) • C set of reconstruction points. • Ψ set of patches used for the reconstruction. • Ψc patch used for the reconstruction of a point c. • σ : blending band width. • Necessity of computing as many distance map as number of patch pieces. Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 21 / 44
Conclusion Spatial Patch Blending Optimised Method Ji (p) = c∈C w(c, p) Ψi c (p − c) ε + c∈C w(c, p) (4) with w(c, p) = e d2(c,p) µ(σ(p))2 (5) and d2 (c, p) = c − p 2 (6) • µ : quantization function. • No distance map to compute → much faster than previous method Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 23 / 44
Conclusion Spatial Patch Blending Optimisations - Blend Scales 1 Quantization of the blending amplitude map σ with N levels. 2 Computation of the blending of N images. 3 Composition of the final image J with all N blending images. Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 25 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 26 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 27 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 28 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 29 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 30 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 31 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 32 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 33 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 34 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 35 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 36 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 37 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 38 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 39 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 40 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 41 / 44
Conclusion Results and Conclusion Results Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 42 / 44
Conclusion Results and Conclusion Conclusion • Fast seams removal • Lightweight changes to apply to inpainting algorithm. • Few parameters, and easily adjustable. Maxime Daisy, David Tschumperl´ e, Olivier L´ ezoray (GREYC) Spatial Patch Blending 20th of June, 2013 43 / 44