Subband patches: A new technique for wavelet-based still image and video coding

In this paper, the third type of wavelet decomposition is proposed that allows us to combine the plain hierarchical structure of Mallat pyramid and the adaptivity of wavelet packets. The proposed approach assumes local adaptive subband decomposition, i.e. after pyramid wavelet decomposition, each subband is segmented to find blocks that are profitable for the secondary transform in the terms of rate-distortion cost. It was suggested to encode groups of wavelet coefficients that have strong 1-D correlation (e.g., contours) using methods based on spatial-orientation trees, as rate-distortion performance of the explicit shape coding methods is low enough. The overall objective performance of the implemented encoder is estimated by rate-distortion curve. Experiments reveal 0.1-0.2 dB improvement for the same bit-rate over the coder based on space-frequency quantization.