Visual Shape Analysis by Wavelets

Recent complexity analysis shows that the problem of visual recognition is an intractable problem, as is the training of neural networks. We propose a representation scheme which greatly simplifies recognition. The approach relies on the wavelet analysis, a recent tool developed for local spacefrequency analysis of a broad class of functions. Based on two assumptions, ( i ) visuai objecrs are localized in the space domain, and (ii) the recognition system is robust, we conclude that for the purpose of recognition every object is locdized in the frequency domuin as well. This fact, combined with the properties of wavelet transform, allows us to keep the same amount of information for objects of any size. We further analyze the consequences of spatial and frequency localization of the objects, and a new framework for visual object representation and recognition arises from this analysis.