Image recognition using a neural network
Ken-Chung Ho, Bin Chang Chieu · 2003
A new type of feedforward neural network for recognition of MRF (Markov random field) images is presented. The proposed forward and backward networks are essentially generalizations of the forward and backward procedures in backpropagation training for general feedforward networks. Due to the feedforward structure of the networks, they are recurrent for homogeneous MRF images and easy to implement. Because of the use of the maximum-likelihood criterion, this approach always performs well if all classes of images are equally likely. Basically, the proposed approach takes advantage of the feedforward neural networks and, by the joint probability, solves two basic problems in MRF modeling: how to measure a Gibbs distribution and how to estimate the Gibbs parameters from clean and noisy MRF samples.>