Abstract

We have previously described¹ the use of steerable filters, in which a small set of oriented basis filters are used to analyze images in a uniform and efficient manner. Filters of arbitrary orientation can be synthesized as linear combinations of the basis filters. We now describe a multiscale pyramid decomposition using steerable filters. The steerable pyramid breaks an image down into a set of bandpassed subimages; the subimages at a given level of the pyramid form a steerable set. The steerable pyramid is overcomplete and therefore nonorthogonal. However, it is self-inverting, meaning that the analysis functions used to build the pyramid are the same as the synthesis functions used to reconstruct. We have implemented image noise cleaning within this representation. The pyramid is used to analyze local image orientation. Neighborhoods with strong orientation are likely to represent image structure, while neighborhoods with little or no orientation are likely to represent noise. The pyramid coefficients are modified according to the likelihood that they represent image or noise. The pyramid transform is then inverted, producing a cleaned image.

© 1991 Optical Society of America