PEI%20YEAN%20LEE

1
TEXTURE SYNTHESIS

• PEI YEAN LEE

2
What is texture?

• Images containing repeating patterns
• Local stationary

3
What is texture synthesis?

• An alternative way to create textures
• Construction of large regions of texture from
  small example images.

Texture Synthesis
Input
Result
4
Goal of texture synthesis ?

• Given a texture sample
• Find synthesize a new texture that,
  when perceived by a human observer,
  appears to be generated by the same
  underlying process.

5
Application 1 Computer Graphics

• Make things look real
• Rendering life-like animations

6
Application 2 Image Processing

• Image compression
• Image restoration and editing

7
Application 3 Computer Vision

• To verify texture models for various tasks such
  as texture segmentation, recognition and
  Classification.

8
Some definitions

• Image pyramid
• A collection of images of reduced resolutions of
  the original 11 image 12n
• Gaussian pyramid
• Consists of a set of low-pass filtered versions
  of the image
• Pg. 161 (Fig 7.17)

9
Some definitions

• Laplacian pyramid
• Consists of a set of band-pass filtered versions
  of the image
• Pg. 198 (Fig. 9.8)

10
Approach 1 Physical simulation

• Advantages
• produce texture directly on 3D meshes, thus avoid
  texture mapping distortion problem
• Disadvantages
• Applicable only to small texture class

11
Approach 2 Probability sampling

• Zhu, Wu Mumford (1998)
• Markov Random Field (MRF)
• Gibbs Sampling
• Advantages
• Good approx. for wide range of textures
• Disadvantages
• Computationally expensive

12
Approach 3 Feature matching

• Model textures as a set of features and generate
  new images by matching the features in an example
  feature.
• Advantages
• More efficient than MRF

13
Approach 3 Feature matching

• Heeger Bergen (1995)
• model textures by matching marginal histograms of
  image pyramid
• Advantages
• Works well for highly stochastic textures
• Disadvantages
• Fails on more structured textures patterns such
  as bricks.

14
Approach 3 Feature matching

• De Bonet (1997)
• Synthesizes new images by randomizing an input
  texture sample while preserving cross-scale
  dependencies
• Advantages
• Works better on structured textures
• Disadvantages
• Can produce boundary artifacts if the input
  texture is not tileable.

15
Approach 3 Feature matching

• Simoncelli Portilla (1998)
• Generate textures by matching the joint
  statistics of the image pyramids
• Advantages
• Can capture global textural structures
• Disadvantages
• Fails to preserve local patterns

16
Web demo

• http//graphics.stanford.edu/projects/texture/