Feature Selection for DNN-HMM Based Mongolian Offline Handwriting Recognition

摘要

The purpose of this paper is to assess the performance of several popular features for handwriting on Mongolian offline handwriting recognition system (HWR). They have been classified into handcrafted and automatically learned features. The handcrafted features are distribution feature, concavity feature, local gradient histogram(LGH) feature and transforms feature. The automatically learned features are extracted by Restricted Boltzmann Machine (RBM) and autoencoder. In this paper, the handwriting recognition system is based on hybrid architectures of hidden Markov models (HMMs)-deep neural networks (DNN) which play state of art role on speech recognize (ASR) tasks. In order to performance comparison, several experiments based on different features extracted from MHW database were performed. The best system on word error rate is based on the LGH feature (5.90%), followed by the autoencoder feature (6.42%).