Automatic Recognition of Isolated Handwritten Digits by Hidden Markov Models
Abstract
Hidden Markov models have proven to be effective in the field of speech recognition. Their application to offline handwriting recognition is a subject still open to research, the goal of which today is to decrease the error rate. HMMs are stochastic models that resist noise, variation, and the elasticity of shape, which is a fundamental requirement in handwriting. The general objective of our work is to apply the Hidden Markov model to the recognition of isolated handwritten digits. We present in this thesis a system for recognizing isolated handwritten digits based on discrete HMMs of Left-Right and ergodic topology, by applying a vector quantization on the features extracted from the images of the digits by the K-means algorithm. Our approach is validated and evaluated on the MNIST corpus that we have used in its entirety. The experiments we have carried out have as arguments the topology of the HMMs, the number of states per model, the size of the Codebook, as well as the number of iterations.