Tree representation and implicit tree matching for a coarse to fine image matching algorithm

In medical imaging arises the problem of watching two images of the same objects but after movements or slight deformations. We present a new method for a primitive global transformation and an improvement of a recent matching strategy which makes it more robust. The strategy consists of two steps. We consider the grey level function (modulo a normalization) as a probability density function. First, we apply a density based clustering method in order to obtain a tree which classifies the points on which the grey level function is defined. Secondly, we use the identification of the hierarchical representations of the two images to guide the image matching. The general transformation invariance properties of the representations permit to extract invariant image points. But in addition, we design a new robust coarse to fine identification of the trees which applies an implicit error measure in a prediction - correction scheme using thin plate splines to interpolate the transformation function in a finer way at each step. Therefore, we will find the correspondence between invariant points even if these have locally moved. The method's results for matching and motion analysis on real images will be discussed.