This article describes a classifier of patterns based on a pre-processing system, located at the input of a recognition system using a Hopfield neural net, which recognises pattern transformed by translation, rotation and scaling. After a detailed description of components forming the chain of the pre-processing system, we present some results obtained by supplying the system input with handwritten characters deformed from rotation, scaling, and translation. The patterns gotten out of the the pre-processing system itself. Besides the well known problems deriving from the scarce memorisation ability of the Hopfield net, is faced by a strategy that foresees the subdivision of the training patterns in groups minimally correlated.
展开▼