Introduction

Abstract

This book is concerned with developing algorithms for some important computer vision problems, especially at a low-level using artificial neural networks. The task of low-level vision is to recover physical properties of visible three-dimensional surfaces from two-dimensional images. One module of low-level vision, for instance, extracts depth information from two eyes, making binocular images, or from one eye over a period of time, making a sequence of monocular images Low-level processes also provide motion information about objects over a sequence of two-dimensional images, the optical flow, for motion detection and representation. As intelligent interpretation of an image requires knowledge about the objects that appear in the scene, learning, representation, and use of prior knowledge must be coordinated. Certainly the human brain is very good at performing vision tasks, but today’s computers are not. This is because of the massive amount of two-dimensional array data that needs to be analyzed and the lack of learning or self-organizing capabilities of most modern day computers. From a mathematical point of view, low-level vision problems are ill-posed according to Hadamard [Had02, Mor84]. An efficient method for solving an ill-posed problem using artificial networks is the Tikhonov regularization technique. The idea of the regularization technique is to narrow the admissible solution region by introducing suitable a priori knowledge or stablizing the solution by means of some auxiliary non-negative functional [Mor84].