A Two-state Markov Chain Model of Degraded Document Images

We propose a two-state Markov chain model of degraded document images. The model generates random and burst noise to simulate isolated pixel reversal as well as blurring of a larger document region. In the Random state, the probability of pixel inversion is low compared to that in the Burst state. However, the model remains in the Random state for a much longer period of time. Validation of the model has been done using the statistical methodology of Kanungo et al [9]. To estimate the parameters efficiently, we use Genetic Algorithm (GA) to search through the parameter space in which the model parameter values are encoded into a concatenated bit string to form the chromosomes. We also show how the accuracy of an optical character recognition system with dictionary search varies with two derived parameters of the proposed noise model.