Bengali Handwritten Isolated Compound Characters Recognition by Applying Transfer Learned Deep Convolutional Neural Network

Optical character recognition (OCR) has been an area of interest for researchers for decades. Many researchers have contributed largely to the development of OCR for script-specific recognition. Handwritten characters recognition has been a big part of this field. Although some literatures have received well-established outcomes, others still haven’t experienced remarkable outcomes yet. Despite being the fifth most spoken language of the world by 228 million people, Bengali has not yet received a remarkable contribution to handwritten characters recognition. Some researchers have offered some promising results for basic Bengali handwritten characters recognition but very few researches offered the recognition of Bengali compound handwritten characters. These few researches have applied support vector machine classifier and some deep neural network classifiers for classification but the outcomes were not much satisfactory. In this research, we considered 171 Bengali compound handwritten characters and apply a modified ResNet-50 model for recognition. We achieved an overall accuracy of 96.15% which outperformed the previous best result of 90.33% by a remarkable margin.