Bengali Handwritten Compound Characters Recognition by Utilizing Transfer Learning

Being the fifth most spoken language of the world, Bengali is spoken by 265 million people globally. Therefore, the recognition of handwritten Bengali characters has been an area of interest for a decade now. Previously, researchers have contributed to the recognition of Bengali numerals and basic characters e.g. vowels and consonants. However, apart from basic Bengali characters, Bengali has a wide range of compound characters as well. Not many remarkable contributions have been yet contributed to this domain. With a focus on contributing to this field, in this research, we started with 171 Bengali compound characters. We applied a modified DenseNet-201 architecture for the recognition of the compound characters. After result analysis, we achieved an overall accuracy of 96.89% which outperformed all previous studies by a remarkable margin. Hence, we concluded that our design is more capable of accurate recognition of the Bengali compound characters.