Quantifying the intra and inter subband correlations in the zerotree-based wavelet image coders

Abstract

The basis of compression is redundancy removal. In digital image compression, discrete wavelet transform (DWT) is applied to remove the inter-pixel redundancies. Although the DWT is very powerful at removing the linear redundancy, there are still various correlations left in the DWT coefficients. These correlations can be modeled as within-subband clustering(Intra) and cross-subband similarity(Inter). The success of recent wavelet image coders can be mainly attributed to the innovative strategies for data organization and representation that exploit these Inter and Intra correlations one way or the other. In this paper, we try to quantify the performance loss if these correlations are removed. Experiments are performed on two best zero-tree coders: set partitioning in hierarchical trees (SPIHT) and set partitioned embedded block coder (SPECK). After recapitulate the data organization adopted in SPIHT and SPECK in a more general tree formation framework, a block or coefficient based pseudo-randomization is applied to remove the Inter, Intra correlation. Our experimental results indicate that the performance loss due to the removal of Intra correlation is much bigger than the performance loss due to the removal of Inter correlation. Therefore, the excellent PSNR performance of the well-known SPIHT algorithm should be much credited for the data structure that exploit the Intra correlation although only the Inter correlation is widely mentioned in the previous literature.