Abstract
-
The paper compares several subband analysis structures and filters for image compression using a generic subband quantisation and encoding method. The authors show that filter banks designed on the basis of an image correlation model outperform other filter banks. The DCT, lapped orthogonal transform (LOT), a cosine modulated filter bank, and an octave band orthogonal filter bank (discrete wavelet transform-DWT) perform in a similar manner. Wavelet filters used to implement the DWT are designed to maximise a coding gain metric. These filters significantly outperform filters designed using other criteria. Daubechies wavelet filters exhibit a coding gain close to these optimum filters and perform similarly for image compression. Some interesting properties relating filter time-width and step response ringing to the location of zeros are observed. Finally a modified DWT structure is proposed and shown to give an improvement over the DWT for some images. �� 1994 IEEE.