Visual shape analysis by wavelets

The complexity analysis shows that the problem of visual recognition is an intractable problem, as is the training of neural networks for this purpose. The authors propose a representation scheme which greatly simplifies recognition. The approach relies on wavelet analysis, a recently developed tool for local space-frequency analysis on a broad class of functions. Based on two assumptions, that visual objects are localized in the space domain and that the recognition system is robust, it is concluded that for the purpose of recognition every object is localized in the frequency domain as well. This fact, combined with the properties of the wavelet transform, make it possible to keep the same amount of information for objects of any size. The consequences of spatial and frequency localization of the objects are analyzed, and a framework for visual object representation and recognition is derived from this analysis