Learning Object Representations by Clustering Banana Wavelet Responses
Gabriele Peters, Norbert Kr · 1997
For object recognition systems it is essential to have an internal representation of the object to be recognized. We introduce a system which learns such a representation from training images of an object class. Every image is preprocessed with banana wavelets. The result is a description of local areas of an image in terms of curved lines. These features are the input to a clustering algorithm which learns to seperate features specific for an object class from features generated accidentally by variations in background or illumination. This leads to a representation of an object class which can be visualized in form of a line drawing. The representation is sparse, for the most part free of redundancies and independent of varying backgrounds and illuminations in the training images. It comprises representative features only, and has already been utilized successfully for object recognition tasks. 1 Introduction The human visual system posseses the remarkable property that it can reco...