Feature Extraction with MLP and CNN in Writer Identification

Authors

  • Vinita Patil, Rajendra R Patil

Abstract

In this work, features from the handwritten documents are extracted using two methods, namely, multilayer perceptrons and convolutional neural networks. Features extracted from these two models were used to define the states of a hidden markov model. Performance of the models were tested on two datasets, namely, VTU-WRITER and IAM datasets. The VTU-WRITER dataset is a custom created dataset by collecting the handwritten documents exclusively for this research work. The performance of the two models namely, MLPHMM and CNNHMM are compared with the hidden markov chain model that has singular values as the features. Baum-Welch algorithm is used to determine parameters of HMM models.

Index Terms—Hidden markov models, Convolutional neural network, Writer identification, Multi layer perceptron etc.

Downloads

Published

2020-05-18

Issue

Section

Articles