Fast image extraction by lifting wavelets

Abstract

Two kinds of fast image extraction methods using lifting wavelets are introduced, which were developed recently by us. The first method uses lifting biorthogonal wavelet filters proposed by Sweldens. The lifting filters include controllable free parameters. Our idea is to learn the free parameters so that lifting highpass components in horizontal direction are equal to those in vertical direction. The learned filters are applied to extract objects in a sequence of images. The second method employs lifting dyadic wavelet filters proposed recently by us. Wavelet transform obtained using these filters is shift-invariant. A learning method of free parameters contained in the dyadic wavelet filters is proposed to achieve fast object extraction. Since the accuracy of the extraction is very high, it can be used as a personal identification tool.