Using a genetic algorithm to adapt 1D nonlinear matched sieves for pattern classification in images

Many methods have been developed to recognise objects in a scene, most involving some pre-processing step to extract local information, such as edges, from the image of the scene. Usually these features are compared with some targets and a classification decision is made. A popular pre-processing step is to decompose the image into a set of scale related features. This type of pre-processing is useful because it can be arranged to remove details that would confuse a simple pattern recogniser. Conventionally, these scale related decompositions use linear processes (like wavelets or the difference-of-Gaussians). However, there is no need for such a decomposition to be based on linear operations and, in this paper, we concentrate on one, the recursive median sieve decomposition, that is nonlinear