Developing Bangla Handwritten Numerals, Basic and Compound Characters Recognition System Using a Deep Convolutional Neural Network
Risal Shahriar Shefin, Emrana Kabir Hashi · 2021
Bangla handwritten character recognition is a crucial task in the field of digitization, building OCR, etc. There are numerous characters in the Bangla language. They have their own styles and shapes. Some of them have so many similarities too. The handwriting style is also a big factor in character recognition. There is a lot of versatility in handwriting style between people. So, recognizing handwritten characters is a very difficult job. But few works were focused on recognizing all types (basic, compound, and numerals) of Bangla individual characters. A multi-branch Deep CNN based model with data augmentation has been proposed in this paper to perform the task. The data augmentation is performed by modifying the positions of the characters in the images to make the recognition system position-independent. Multi-branch is used to explore and process multiple types of features from the same source together. The datasets used to train, validate and test the model are the CMATERdb dataset and the Ekush dataset. Different datasets have been used for this experiment to analyze the results more precisely. The model has obtained 97.82% validation accuracy and 97.3% testing accuracy on the CMATERdb dataset (231 character class) and 96.18% validation accuracy and 96.13% testing accuracy on the Ekush dataset (122 character class). From the results, it can be observed that the model achieved pretty high accuracy on different datasets. On recognizing various Bangla characters from mouse input in real-time, it took only 0.66 ms on average to recognize them. Handwritten character recognition has many real-life applications as everything is going to be in digital form. Because of high classification accuracy, our proposed model can be used and extended to do more complex tasks in the corresponding fields.