SIGNAL AND IMAGE FEATURE EXTRACTION FROM LOCAL MAXIMA OF GENERALISED CORRELATION

摘要

A new method of wavelet packet analysis is presented where the wavelet packets are chosen from a manifold rather than a discrete grid. A generalisation of the wavelet transform is defined on this manifold by correlation of the wavelet packets with the signal or image, and a discrete subset of the wavelet packets is then chosen from local maxima in the modulus of this function as a form of signal or image feature extraction. We show that consideration of the geometry of the manifold aids the search for these local maxima. We also show that the resulting wavelet characterisation is the best local approximation to the signal or image and represents signal and image components with the greatest signal to noise ratio, and is thus useful to surveillance and detection.