Analysis of Deep Learning Techniques for Handwritten Digit Recognition

Abstract

The huge variations in culture, community and language have paved the path for a massive diversification in the handwriting of humans. Each one of us tends to write in a different pattern. Character or digit recognition finds humongous applications in the recent days especially in the processing of bank statements, sorting of postal mails and many more. Although many classification models exist in literature that successfully classifies the handwritten digits, yet the problem that is still persisting is which one can be termed as an optimal classification model with higher accuracy and lower computational complexity depending upon the circumstances. In this paper, the Machine Learning classification models involving the likes of K-Nearest Neighbor (KNN), Support Vector Machine (SVM) and XGBOOST were compared with that of Deep Learning models like Artificial Neural Network (ANN) and Convolutional Neural Network (CNN). The comparison clearly portrays how convolution operation on images plays a vital role to outperform rest of the classification models. Then the paper compares the CNN models on the basis of two different sets of Loss functions and Optimizers to delineate their role in enhancement of accuracy of the model. The only limitation lies in the fact that in spite of being a handwritten recognizing model this model only recognizes digits in the range of 0–9.