Handwritten Numerals Recognition by Employing a Transfer Learned Deep Convolution Neural Network for Diverse Literature

Having around 6,500 languages worldwide, handwritten numerals recognition has been a domain of research for decades now as numerals are a common phenomenon among all these diverse languages. Previously, researchers have contributed significantly to recognize the handwritten numerals of diverse literature. Different approaches have been discovered to be feasible for language-specific numerals recognition. However, finding a common architecture to recognize numerals have been a goal from the very beginning. But despite having many efforts, discovering a common architecture for high recognition of numerals of diverse literature has always been a challenging task to solve and not many contributions have been made in this regard. Therefore, in this research, we focused on seven benchmark datasets of six languages and proposed a modified DenseNet-201 architecture. Our proposed architecture achieved an overall accuracy of 99.04%, 99.33%, 98.83%, 99.50%, 99.83%, 99.54%, and 99.74% for Bengali (CMATERdb 3.1.1), Devanagari (CMATERdb 3.2.1), Arabic (CMATERdb 3.3.1), Telugu (CMATERdb 3.4.1), Nepali, ARDIS II, and ARDIS III datasets respectively which outperformed all notable previous works by a noteworthy margin.