Automatic feature determination using unsupervised neural networks. Application to image registration

In this paper, a new solution is proposed to automatically determine the characteristic features in images, a task that is often needed in image analysis. Two unsupervised neural networks are used: the Kohonen's self-organizing map, and a Fukushima's neocognitron like model. Both networks are used to cluster the subsets extracted from the image in an unsupervised learning procedure. This procedure uses no a priori characteristic feature definition. A simple strategy is then used to define the characteristic subsets. Experimental results are given and an application to image registration is presented