Attention-based similarity

A similarity measure is described that does not require the prior specification of features or the need for training sets of representative data. Instead large numbers of features are generated as part of the similarity calculation and the extent to which features can be found to be common to pairs of patterns determines the measure of their similarity. Emphasis is given to salient image regions in this process and it is shown that the parameters of invariant transforms may be extracted from the statistics of matching features and used to focus the similarity calculation. Some results are shown on MPEG-7 shape data and discussed in the paper.