Iterative identification framework for robust hand-written digit recognition under extremely noisy conditions

Abstract

A new classification framework is proposed for noise invariant hand-written digit recognition, which is based on the Turbo decoding technique and the Viterbi algorithm. Specifically, labeled training digit images are transformed into a two-dimensionally correlated Markov Chain Model (MCM). In order to increase the discriminant function of MCMs, a novel sequence learning algorithm is proposed to obtain Sequence Maps and improved MCMs for each digit class, minimizing entropy of MCMs within individual digit classes. The target image is accordingly transformed by Sequence Maps and explored by improved MCMs in the horizontal and vertical directions iteratively to calculate the likelihood with respect to each digit class. The effectiveness of the proposed approach is verified through extensive experiments, showing that our classification algorithm can significantly enhance the accuracy of hand-written digit recognition even under extremely noisy conditions.