A novel enhancement for hierarchical image coding

Hierarchical image coding usually codes a down-sampled version of an original image and then the difference between the original image and a reconstructed version that is interpolated from the down-sampled layer. In this paper, we demonstrate, for the first time, that when the bit-rate used to code the residual layer falls into a critical region (which covers almost all typical bit-rates used in practice), it often happens that all pixels in the down-sampled layer would be deteriorated if the corresponding coded residuals are added into them. To avoid this problem, we first propose a “naive” solution: no coded residuals will be added back into the down-sampled layer; whereas coded residuals will be added only into the interpolated pixels. Then, we propose to apply a constrained quantization technique during the coding of the residual layer so that all residual pixels at the interpolated positions will end up with an improved quality. To verify its effectiveness, we conduct extensive tests to show that the gap between the hierarchical coding scheme and its single-level counterpart (which is typically around 2–3 dB in the 2-level hierarchy) will be filled up by a rather big percentage.

► We identify, for the first time, a problem of quality drop in the traditional hierarchical image coding scheme.
► We develop two solutions to this problem.
► One solution applies a highly constrained quantization technique.
► A remarkable improvement over the traditional hierarchical coding scheme has been achieved.