Abstract

The pattern spectrum was proposed to represent morphological size distribution of images. A nobel definition of pattern spectrum, called multiresolution pattern spectrum (MPS), is proposed in this paper. The most important significance of MPS is the capablity to separate spectra of small-sized details from those of large-sized rough draft of meaningful figures. In case of morphological filtering, small-sized figures are regarded as noises and to be removed, while large-sized figures are regarded as noiseless images and to be preserved. Thus MPS is much more meaningful measure of the design of morphological filters than the conventional one. The application of MPS to the optimization of a filter by unsupervised learning is also presented.