Texture Image Extrapolation for Compression

1
Texture Image Extrapolation for Compression

• Sung-Won Yoon SeongTaek Chung
• Stanford University
• December 4, 2000

2
Outline

• Motivation Recover the whole image using
  partial information of the image
• Methods
• non-parametric sampling
• estimation using subband decomposition of wavelet
  coefficients
• estimation by induced correlation

3
Non-parametric Sampling

• Based on the approach of Alexei A. Efros and
  Thomas K. Leung
•  Method

4
Results

• Big Hole
• Window size
• Filling order
• Scattered Holes
• Better performance
• More data needed

5
Estimation Using Subband Decomposition

• Spatial locality
• Similarity between subbands
• Estimate A from B by use of the mapping from C to
  B

6
Model

• One-to-four linear mapping
• A. Pentland B. Horowitz
• Mapping between pairs of subbands are similar
• Full search possible because repetitiveness of
  texture image

7
Results
Detail level 1 zeroed
Detail level 1 estimated
Original
8
Limitations

• Statistical differences in different subbands
• Assumption of propagation of mapping does not
  hold in general
• Very limited mapping information from lower
  subbands

9
Estimation by Induced Correlation

• System Model

10
Results
Estimated image 1 (PSNR 15.33dB)
Original
Interpolated image (PSNR 12.73dB)
Estimated image 2 (PSNR 14.73dB)
11
Conclusions

• Non-parametric sampling
• window size and computation load
• Estimation using subband decomposition
• lack of mapping similarity between pairs of
  subbands
• different statistical characteristics for
  different subbands
• Estimation by induced correlation
• optimal filter is hard to find