Automated Construction of Identification Procedures for Objects Belonging to Several Classes

摘要

Automated approaches to the choice and tuning of image analysis algorithms for solving a particular problem are considered. A new automated method for the construction of near-optimal object identification procedures is described. The objects to be identified can belong to several classes. The identification is based on reference images of those objects and uses a learning sample of images. The construction of the desired procedure assumes that it is selected from a set of procedures detecting the given object. This set is formed by the reference image of the object and uses algorithms of certain predefined types. The selection procedure is based on a genetic algorithm described in the paper.