Wavelet-based image coding involves discardingrnsubband information in an organized fashionrnto compress an image to a given rate, i.e. some kindrnof rate control. In a practical coding system, this taskrnrequires knowledge of the relationship between subbandrnquantization step-size and compressed rate. At high rates,rnthis behavior can be characterized by modeling eachrnsubband as a Gaussian random process. At low rates,rnsuch methods break down since 1) quantized coefficientsrnno longer resemble Gaussian distributions, and 2) mostrnof the coefficients are zero, but the positions of non-zerorncoefficients are spatially dependent. These dependenciesrncan be handled by extending models to characterize morernlocalized behavior. Based on this observation, a modelbasedrnrate-control technique is presented. Rate estimatesrnare generated by considering each subband a collectionrnof quantized random processes. It is demonstrated that atrnlow rates (0.5 bpp and below), these methods accuratelyrnpredict the rate of a compressed image within five percentrnof the rate achieved using a state-of-the-art coder.
展开▼
