BAYESIAN LEARNING METHOD IN THE DISCRETE CELLULAR NEURAL NETWORKS APPLIED TO IMAGE PROCESSING

Tóm tắt

This paper focuses on training the coefficients of the discrete cellular neural networks using Bayesian learning. In this method, the prior distribution, which is assumed to be a Gauss distribution, is determined based on prior information and the posterior distribution is then calculated using Bayes' theorem. A Markov Chain Monte Carlo method, specifically the Metropolis-Hastings algorithm, is used for generating random samples corresponding to the posterior distribution, thereby helping to estimate the coefficients of the network. We have modified the Metropolis-Hastings algorithm to reduce the coefficient estimation time. Some image processing experiments are implemented with estimated coefficients. In more details, we target the size of the network to be trained. Using the training data obtained from the distillation technique, we found that training a smaller network size using the method described above for image processing also gives equivalent results as compared to a larger network size. This can reduce the training time, resulting in smaller training costs and thus increasing the training efficiency of the discrete cellular neural networks.