Efficient image coding with QMF pyramids
Eero P. Simoncelli, Edward H. Adelson, Rajesh Hingorani · Annual Meeting Optical Society of America · 1987
Quadrature mirror filters have been widely used in the subband coding of speech signals and have recently been applied to image coding as well.1 By using cascaded filtering and decimation we have derived QMF pyramids that are rapidly computed and afford very good data compression. High-pass kernels with even symmetry become possible when using a staggered sampling grid and can be combined separably for 2-D processing. Nonseparable 2-D kernels with interesting properties have also been derived. The good frequency localization permits data compression that takes advantage of scene statistics and the sensitivity of the human visual system. The good spatial localization of the kernels avoids ringing artifacts. The transform achieves localization without blocks and thus completely avoids block artifacts. The result is highly efficient coding that gives visually pleasing images.