Bangla Handwritten Digit Recognition using RNN-CNN Hybrid Approach

Abstract

The automatic recognition of handwritten English material has seen a lot of progress. However, research on automatic Bangla handwriting numerals recognition is far behind. Even the most effective recognizers now in use do not produce an adequate performance for real-world applications. This paper suggested a strategy based on deep neural networks. In this paper, we have used the BanglaLekha-Isolated handwriting dataset along with ResNet50 and DensNet201 models for benchmarking process. Then we proposed two new models one is a Gated Recurrent Unit (GRU) based and another one is a Hybrid of Convolutional Neural Network (CNN) and Convolutional Long Short-term Memory (ConvLSTM). As for our proposed GRU model it performs closely to the DensNet201 and REsNet50 models while requiring very few parameters compared to these two models. On the other hand, our proposed Hybrid ConvLSTM model outperforms both of the aforementioned benchmarking models. Finally, we have developed a new Bangla Handwriting Numerical dataset containing a total of seven thousand training, one thousand validation, and two thousand test images. Our proposed best-performing model (Hybrid ConvLSTM) achieves 98.84% accuracy in the test data of our dataset while the GRU model gained 91.33% test accuracy without any help of image preprocessing steps.