Rotation- and scale-independent pattern recognation through optimization

This paper presents a two-dimensional pattern recognition paradigm, which is independent of rotation and scale. It uses a Hopfield neural network with a configuration similar to that employed for the traveling salesman problem (TSP). Enhancements to the original Hopfield design are included which are based on the eigenvalues of the connection matrix and on modifications suggested by simulations. The goal of the method is to determine the underlying linear transformation between a test image and an observed input pattern by minimizing the energy of a Hopfield network. This procedure is easy to implement and is well suited for sparse binary images, e.g. skeletalized representations or the output from vertex detection algorithms. In this work the effects of noise are studied both analytically and through Monte Carlo techniques yielding some useful performance guidelines and greatly increasing the utility of the method. Experiments show that this technique performs successfully on a variety of rotated and scaled images and is robust against additive noise.