Fourier-wavelet techniques in image searching

This paper reports on a set of techniques useful for searching collections of images for the set of images that match a given template according to a set of criteria. The techniques in this paper compare a filtered form of the template to a similarly filtered form of the image where the filtering is used both to reduce the amount of data and to enhance the features that are being compared. The main idea of the search is to use multiresolution wavelet representations in which coarse subbands are used for an initial search, and then progressively finer subbands are used to search images that pass the initial discrimination test. The wavelet representations are compared by using Fourier transforms on the wavelets in a way that provides several different comparison criteria. The method has an additional advantage that the Fourier transforms of coarse wavelets can be used within the computation of the Fourier transforms of finer wavelets to reduce the cost of the refinement computation. A key feature of the search algorithm is that it uses masking to reduce the effects of image occlusions.