A Comparison Study of Recurrent Neural Networks in Recognition of Handwritten Odia Numerals

Abstract

Character recognition in handwritten data is a challenging work as the writing style varies from person to person. Choosing a method for handwritten numeral recognition is also having importance as the result depends on the method used in the recognition model. Deep learning-based recurrent neural network (RNN) is being used for prediction in time series data, generation of text lines, and other sequential data processing. But in this work, the pixel values of image are used as time stamp dependent input to the recurrent networks. We have applied simple RNN, long short term memory (LSTM), and gated recurrent Unit (GRU) for recognizing Odia handwritten numerals. A comparison study is also provided to understand the effect of vanishing gradient in RNN and how this drawback of RNN has been overcome by LSTM and GRU cells. The Adam optimizer is used in each proposed method. The accuracy values obtained in RNN, LSTM, and GRU are 50.04%, 88.81%, and 86.24%, respectively. KeywordsCharacter recognitionOdia handwritten numeralsDeep learningRNNLSTMGRUAdam optimization