Image interpolation based on median-type filters is investigated.
We study this problem for two decimation schemes that are most
commonly used in practice: (1) quincunx lattice and (2) rectangular lattice.
For each decimation lattice, we propose a number of interpolation
structures that are based on some median or finite impulse response
(FIR)-median hybrid filters. Extensive simulations show that the proposed
median-based interpolators outperform significantly as compared
with linear schemes. To further verify the effectiveness of these medianbased
interpolators, we compare them with multilayer neural networks
designed through a training process, and still show a preference for the
median-based interpolators. As an application, they are used in an interpolative
Joint Photographic Experts Group (JPEG) coding scheme to
suppress blocking artifacts at very low bit rates.