A convolution of input data S.sub.0 and a Gabor function G.sub..alpha.,.omega. is implemented by a processing unit (20) to obtain a wavelet coefficient T.sub.g (2.sup.0,b) of a 0th stage. The input data S.sub.0 is applied to a low-pass filter (30) and is subjected to Gaussian-type filtering so as to be converted into input data S.sub.1 for a 1st stage. The frequency of the high-frequency components of this input data is halved by this filtering. A convolution of this input data S.sub.1 and a Gabor function G.sub..alpha. /3,2.omega./3 is implemented by a processing unit (21) to obtain a wavelet coefficient T.sub.g (2.sup.1,b) of the 1st stage. In the convolution of the 1st stage, the sampling interval is twice that of the 0th stage. Thenceforth, and in similar fashion, the sampling interval in processing operations of subsequent stages is doubled from one stage to the next. As a result, there is no increase in the amount of processing in each stage and a coefficient T.sub.g (2.sup.j,b) can be obtained with identical precision in each stage.