Performance Study of a Regularization-Based Deformable Handwritten Recognition Approach

摘要

This study clarifies the accuracy performance of a deformable handwritten recognition approach (DHRA) for digit characters. The deformable approach consists of regularization-based displacement computation, coarse-to-fine strategy, distance measurement and k-nearest neighborhood method. We focus on several conditions for investigating the accuracy and the sensitivity, that is, the definition of averaging area in regularization process, regularization parameters and the number of k for k-nearest neighborhood method. According to the simulation results, it was shown that the proposed method has the error rate of 0.42% for MNIST handwritten digit database, resulting in the top-group of the performances reported until now.