Pattern classification with incremental class learning and Hidden Markov models

Abstract

Incremental class learning - Hidden Markov models (ICL-HMM) system combines two different approaches adopted in pattern recognition area to form a new, robust solution. Our system is composed of two parts - ICL feature extractor and HMM sequence recognizer. The former, ICL, is an artificial neural network capable of incrementally learning to recognize features of patterns from a narrow window sliding over them. HMMs simulate systems that transfer from one hidden state to another. In every state the system generates some observations. In our system we train one HMM for every class of patterns by presenting to it the sequences of observations generated by ICL for patterns belonging to its class. In the testing phase, every HMM checks how well it models the sequence of observations generated for an unknown pattern. We present promising results of applying ICL-HMM system to printed Latin character recognition task.