Texture Synthesis by Patch-Based Sampling

1
Texture Synthesis by Patch-Based Sampling

• Real-Time Texture Synthesis By Patch-Based
  Sampling, LinLiang, Ce-Liu, Yingqing Xu, Baining
  Guo, Heung-Yeng Shum, 2001.
• Presented by Maryia Kazakevich
• Dec.3, 2004

2
Outline

• Background to Texture Synthesis
• Patch-based sampling
• Algorithm outline
• Some details
• Summary of the method
• Results

Maryia Kazakevich
3
Background

• Texture mapping simple tiling might be
  unacceptable

• Want output texture to appear perceptually
  similar to input
• Several techniques possible
• Similar pixel-based sampling
• Neighbourhood match

4
Patch-Based Sampling

• Select next patch from the input image based on
  the boundary zone match

5
Outline of Algorithm

• Randomly select first texture patch.
• Choose the next patch from the input image to
  paste in the output.
• Repeat previous step until output image is fully
  covered.
• Blend pixels in the boundary zones.

6
Algorithm in detail

• To select the next patch
• Construct a set of possible texture patches with
  boundary zone matching the output
• If the set is empty, chose the patch with minimum
  d(EBk, Ekout)
• Otherwise randomly select a patch from yb as next
  texture patch to paste into the output image

7
Algorithm in detail (constructing yb)
Boundary zone matching
A number of pixels in the boundary zone Pi
i-th pixel of the boundary zone EBk, Ekout
boundary zone of the next patch and output
image e relative matching error between
boundary zones
8
Boundary zone blending

• To smooth the transition between adjacent
  patches,
• blend them using feathering
• d(x) distance to the border pixel
• w monotonic function (used w(d(x)) d(x))
• I original intensity
• C output intensity

9
Summary

• Paste random patch in one of the corners.
• Select next best suited patch and paste it in the
  next spot in the output image.
• Blend the overlap regions between patches.

10
Input parameters

• Patch size how well local characteristics of
  input texture are captured
• Larger boundary zone width stronger statistical
  constraint, more natural transition between
  boundaries
• Relative matching error controls similarity of
  overlapping boundary zones

11
Influence of the parameters

• Input 32x32, Output 64x64
• Checkers
• Patch size 16x16 8x8
• boundary 2 8
• Relative matching error 0.6 0.2
• Dots
• Patch size 16x16 6x6
• boundary 2 10
• Relative matching error 0.2 0.2

12
My interface

• Not shown
• Input fields for the output texture size
• Buttons for selecting a different input picture
  and for synthesizing

13
Results

• Patch size16x16, boundary 2
• Relative matching error 0.2
• Input 32x32, Output 64x64
• Input 32x32, Output 84x84