Handwritten Character Recognition Using Convolutional Neural Network

Abstract

This paper applies machine learning technique for handwritten character recognition. Handwritten recognition can recognize the handwritten character in different styles efficiently on the computer or any electronic device by receiving the input from a touch screen, electronic pen, scanner, images, and paper documents and convert them into digital form. It has various authorities such as reading advice for bank cheques, recognizing letters and digits from form applications. The existing method from machine learning like Convolutional Neural Network (CNN) based handwritten character recognition utilized to solve this problem. The goal of this study is to create the model that can identify and determine the handwritten character from the EMNIST with better accuracy. The EMNIST dataset that consists of English alphabets and digits is used to train the neural network. EMNIST datasets are downloaded from the Kaggle. This paper aims to develop on algorithm based on convolution neural network for handwritten character recognition. The objectives of this study are to develop an optimization technique that can increase the accuracy of handwritten character recognition and analyze the performance of the proposed algorithm with the test data set. It illustrates the use of convolution neural networks for generating a system that can recognize handwritten characters. In the created model, each character is described by binary values that act as input to a pure feature extraction system. The output is fed to the neural network model. The CNN approach is applied to complete the task of classifying words directly and character segmentation. CNN also act as a classifier to train the EMNIST dataset. This project used Python with TensorFlow libraries to develop a simulation of federated learning algorithms on the model. The project has made use of OpenCV to perform image processing and predicted image. The regularization parameters such as Dropout to prevent overfitting. The proposed work enhances the recognition rate of characters with accuracy of 88.7%.