Offline Bangla Handwritten Text Recognition: A Comprehensive Study of Various Deep Learning Approaches

Abstract

Offline Handwritten Text Recognition (HTR) is a technique for translating handwritten images into digitally editable text format. Due to the presence of cursive letters, punctuation marks, and compound characters, it is more complex to recognize Bangla handwritten text. Over the years, several approaches to the optical model of the HTR system have been developed, including Hidden Markov Model (HMM) or deep learning techniques such as Convolutional Recurrent Neural Networks (CRNN), and current state-of-the-art Gated-CNN based architectures. Despite this, there are relatively limited works available for Bangla word recognition. In this paper, we introduce an end-to-end system for Bangla word recognition. We used a variety of popular pre-trained CNN architectures, including Xception, MobileNet, and DenseNet, followed by recurrent units such as LSTM or GRU. Furthermore, we experimented with Puigcerver’s CRNN based and Flor’s Gated-CNN based optical model architectulimited works available in Bangla.res. Flor architecture provided the highest recognition rate in our experiment, with a CER of 12.83% and a WER of 36.01%.