torn p utati on of waveletli ke - based features

Given a gray scale or color image, the focus of attention is usually on a particular area. One way to accomplish this is to segment the image into various regions based on texture. This selective segmentation process has applications to both medical and satellite image processing. In this paper, we detail an approach to this selective segmentation process based on features extracted from a set of analog-computed pseudowavelet coefficients. The silicon retina of Mead1 displays many features of human retinal processing . Two key features of interest here are the optical input to the silicon retina chip and the center surround response of each node in the retina that is reminiscent of a Morletlike mother wavelet. This center surround response, which is similar to the difference of Gaussians, is accomplished by differencing the input at a pixel with a local average computed via an analog resistive grid. In this paper, we propose an optoelectronic system for performing resistive grid-based wavelet transformations. We include an example where the multiresolution wavelet coefficients2 have been further processed to produce local variance and vanance of variance features , which can be used to perform texture discrimination. The twin advantages of this hybrid approach are the near real-time processing to obtain wavelet coefficients and/or texture features and the nonlinear datadriven nature of the effective wavelets produced by a resistive grid composed of nonlinear components.