Automatic Image Restoration

Back to my research projects, you also can find the paper of this algorithm from my research projects page.


Paper and Data


Motivation


Diffusion works well in narrow pixels

When the defect region is only narrow pixels, diffusion (solving partial differential equation) works very well.

Original Image:

3 children : input volley ball : input
downarrow

Result of Fast Digital Image Inpainting (Oliveira, Manuel M., Brian Bowen, Richard McKenna and Yu-Sung Chang, ``Fast Digital Image Inpainting,'' Proceedings of the International Conference on Visualization, Imaging and Image Processing (VIIP 2001), Marbella, Spain. September 3-5, 2001, pp. 261-266.):

3 children : Fast Digital Image Inpinting volley ball : Fast Digital Image Inpinting

Result of Image Inpainting (Marcelo Bertalmio and Guillermo Sapiro and Vincent Caselles and Coloma Ballester, ``Image inpainting,'' SIGGRAPH 2000, pp. 417-424):

3 children : Image Inpinting volley ball : Image Inpinting

Large region is problematic.

The original brick texture comes from (VisTex data set Copyright (C) 1995 MIT) http://vismod.media.mit.edu/vismod/imagery/VisionTexture/vistex.html

From left to right : Input Image, Fast Digital Image Inpainting. Notice: This result is only isotropic diffusion.

brick homogeneous input rightarrow brick homogeneous fd result

Texture synthesis works well on homogeneous image

From left to right : Input Image, Texture synthesis.

brick homogeneous input rightarrow brick homogeneous textsynth output

Texture synthesis works hard on non-homogeneous image

From left to right : Input Image, Texture synthesis.

brick non-homogeneous input rightarrow brick non-homogeneous texsynth output

Image Restoration using Multiresolution Texture Synthesis and Image Inpainting with frequency domain decomposition algorithm (this paper)

From left to right : Input Image, Our method on homogeneous image.

brick homogeneous input rightarrow brick homogeneous output

From left to right : Input Image, Our method on non-homogeneous image.

brick non-homogeneous input rightarrow brick non-homogeneous output

Other examples

Posters: Left input, Right result.

posters input posters output

Tables: Left input, Right result.

tables input tables output

Ropeway: Left original, Middle masked input, Right result. (Picture: Copyright (C) 2002 Goshima Kazuhiro)

rope original rope masked input rope output

For further understanding

I can write this paper because there are these works.

I tried to complete these links, but it is not possible, as you know. Then, if you know more other related links, please inform me.

My short comments are just memorandums for me.


FAQ


Copyright (C) 2003-2005 YAMAUCHI Hitoshi
Most recent update : :