Automating The Analysis of Mammograms: A supervised learning approach to sensor movement

This paper investigates the problem of automating the analysis of mammograms and reports on the use of a method, named Constructive Image Partitioning in conjunction with a space-variant sensor. The algorithm iteratively partitions the input image at multiple resolutions concurrently. It isolates visual entities in the image by generating closed contours on the basis of light intensity data. The visual entities can then be analysed for its relevance to the recognition problem at hand before it is either discarded as visual noise or incorporated into the representation of the image. A method of learning foveation points to guide the movement of the spacevariant sensor is described. The significance of this method is that the partitioning will emerge in a manner that will not discard visual details as a result of a priori assumptions and is thus relevant to the pursuit of general purpose artificial vision. The search process operates concurrently on each resolution, wherein features extracted at each resolution may be used by the search process at