Handwritten digits parameterization for HMM based recognition

Abstract

Handwriting classification or recognition methods based on neural networks (NN) have been extensively studied and they are now well known. This process, which parameterises the geometric structure of the digits as a previous stage to their recognition by the neural network, has the inconvenience of ignoring the sequential character of handwriting. The method proposed explores the improvement introduced in a handwritten recognition system when it incorporates the sequential information of handwriting and the hidden Markov model (HMM) is used as a classifier. The handwritten off-line classifier proposed acquire the handwritten characters by a scanner and after their parameterisation (include noise filtering, binarization, thinning and vectorisation) as a sequence is recognised by the HMM classifier, which provides a good probabilistic representation of sequences having large variations. Different parameterisation techniques are introduced and compared.