In conventional 1D filter banks used for reversible subband coding of
gray-level still images, the high-band signals are extrapolative or in
terpolative prediction error signals, the former using one side of the
pixels alone and the latter using two sides of the pixels but with le
vels four times that of the input signals. This paper proposes a rever
sible subband coding method which has a high compression efficiency by
reducing the entropy of the high-band signal. This method employs 1D
filter banks which use interpolative prediction error signals having l
evels twice that of the input signal as the high-band signals. This pa
per also describes the design of nonseparable filter banks which predi
ct a pixel in question by using four pixels around the point. The comp
ression efficiency of the proposed method has been compared with the c
onventional method by using real images. The results show the followin
g: the entropy of the proposed method is smaller than the conventional
method by 0.1 to 0.5 bits/pel; some samples processed by the proposed
method show a higher compression efficiency than those in the reversi
ble predictive coding; the separable filter banks have a higher compre
ssion efficiency than the nonseparable filter banks (both based on the
proposed method); and the reduced images using the proposed separable
method with a lowpass filter have many taps to suppress aliasing.