Mark Kalman Isaac Keslassy Daniel Wang 12/6/00 Compound Coding Mark Kalman Isaac Keslassy Daniel Wang 12/6/00
Outline Motivating Compound Coding Classification Algorithms Work from the Literature Spatial DCT-Based Methods Wavelets Our DCT Codebook Method Comparison Smoothing An Example
Motivating Compound Coding At 1bit/pel: Median graphics PSNR = 31dB Median text PSNR = 19.5 dB Different properties => different coding
Past Work (1) Spatial Domain Algorithms Chen (1990) : block variance Bones et al (1990) : edges (Sobel filter) DCT-Based Algorithms Chaddha et al (1995) : DCT-18 Absolute-Sum Konstantinides (2000) : DCT Bit Rate
Past Work (2) Wavelet-based Algorithms c2 goodness of fit to Laplacian similar to MSE “discreteness” of distribution concentration of data in peaks
Our DCT Codebook Scheme DCT coefficient values are differently distributed in text and graphic regions Each distribution is given by a codebook histogram graphics text <Do we need a slide or sth for the methodology before next slide?> Distribution for DCT coefficient (2,1)
The Algorithms Compared Talk About DCT Horizontal
Smoothing Classification Algorithm Non-linear Smoothing
An Example Uniform Coding .83 bits/pel Compound Coding .81 bits/pel
Conclusion Text and graphics have different characteristics Many methods of classifying blocks Spatial, DCT, Wavelets, DCT Codebooks Smoothing/post-processing Our methods compare favorably