BWordDeepNet: a novel deep learning architecture for the recognition of online handwritten Bangla words

Online handwritten word recognition (OHR) in low-resource languages such as Bangla is still an open problem. Although the need and importance of OHR are increasing nowadays, research works on word-level recognition are few (specifically for Bangla script), and there is a lot of room for improving recognition performance. In the current work, we employed different Recurrent Neural Network (RNN) architectures such as Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BLSTM), Gated Recurrent Unit (GRU), and Bidirectional Gated Recurrent Unit (BGRU) for the recognition of online handwritten Bangla words written in an unconstrained domain. One of the challenges includes the variable number of strokes used to write words. This study aims to develop a segmentation-free recognition module where the features from constituent strokes of the word sample are fed to the developed RNN architectures. Sequential and dynamic information obtained from the strokes is considered as the features for the current experiment. The customized architecture of BLSTM known as BWordDeepNet (Bangla Word Deep-learning Network) provides the best performance with 98.35% correct recognition accuracy on the dataset having 7992 online handwritten Bangla word samples. Additionally, the model achieves a numerical gain of 8.08% compared to the Bangla word recognition work mentioned in [38] that was performed on the same word dataset containing 5550 word samples. We have also compared the performance of our proposed model with state-of-the-art techniques used for the same purpose.