Geometric Image Analysis Using Multiscale Statistical Features

he Artificial Visual System (AVS) (Figure 1) ~mputational framework for computer vision. [1] n AVS convolves an image with a series of D mtial filters, yielding a series of D filtered aages. While any set of filters can be applied in ds way, families of filters can be designed to .~oduce a compact, task-specific decomposition or a meral, task-independent decomposition of the aage. The filter set disperses information about .~age structure through the filtered images. The ~tputs of the filters at a particular location, =(x,y), define pattern (i n the sense ofstatistical lttern recognition), Px (Figure 2). This pattern m be treated as a point in a D-dimensional ature space; the convolution of each filter with Le original image computes one feature for every xel. Thus, we say that the filters map each pixel to a point in a D-dimensional feature space, so e can refer to pixels Ix, or to their corresponding ~ints Px. Mapping pixels into a feature space is a powerful idea. The tools of statistical pattern recognition2,3,6 (classification, clustering, multidimensional scaling, various transformations and warpings of space) can be applied to discover structure or to determine an appropriate labeling of each pixel. Decision methods are nonlinear functions that recombine the information dispersed by the filter set to yield the desired inferences. Recombination algorithms as simple as thresholding, absolute value, or location of relative extrema have proven useful in the past.