Optimized Sequential CNN Architecture for Robust Handwritten Digit Recognition

Abstract

Optical character recognition (OCR) system is difficult to train because traditional handwritten digit recognition (HDR) algorithms have relied significantly on preexisting data and specific characteristics. Convolutional neural networks are the primary deep learning methods utilized in HDR research and have recently shown ground-breaking results. It is challenging to identify and categorize numbers efficiently due to the existence of distortion, haziness, light, size as well as people's varied writing styles. Numerous techniques have been put out to address these constraints, including CNNs for HDR; however, these techniques suffer from low accuracy and low computing performance. There are so many pre-trained CNN modes like: LeNet-5, Alex-Net, VGG-16, ResNet-50 etc. However, we are proposing a customized sequential model of CNN that offers great computational speed and accuracy for HDR. Our proposed model of classifier uses MNIST dataset with appropriate settings for training and testing to recognize handwritten digits. The technique mentioned above effectively transmits the accuracy of 99.14%, which is higher than previously suggested methods. Additionally, the suggested system considerably reduces the computing time needed for testing and training of the HDR system that enabling the proposed model more efficient.