ABSTRACT

A computer can read and interpret intelligible handwritten input from sources like paper, photos, and other devices, known as Handwriting recognition (HWR). Besides, handwritten recognition is an interesting challenge in machine learning and deep learning. Because several strategies and approaches have been followed already to solve this challenge, machine learning and deep learning provided the best results.Handwritten digit recognition is a part of HWR. It is getting popular day by day because many applications could be made using this system like OCR, postal code recognition, license plate recognition, bank checks recognition, etc. Besides, the importance of recognizing the Bangla digit from the document is increasing. But the works available in Bangla handwritten digit recognition are very few. Similarly, none of them are robust, and some of them are overfitted. Therefore, we need to make some improvements to this system considering its importance. This paper explores the presentation of transfer learning with the help of some best-in-class profound CNN strategies for the acknowledgment of manually written Bangla digits. It considers two deep CNN architectures, such as Mobile Net and Residual Network (Reset) based on performance and accuracy. This model was trained and tested with the CMATERdb dataset. The study suggests that transfer learning provides 97% accurate results, where traditional CNN provides 86-92%.

Keywords: - BangleHandwritten Digit, CNN, MobileNet, ResNet50, Transfer learning