Investigation on performance Improvement of deep learning based Convolutional neural network Models using novel optimization Techniques for image recognition of Handwritten digits

Abstract

The scope of computers over every field makes them increasingly newlineintegrated into human life and everyday activities. The advancements over the newlinecomputers made the human use them intelligently and genuinely, which makes newlinethem integrate with the humans naturally. The recognition of complex digits using newlinea pre-dominant method is essential. Thus, the advantage of automatic handwritten newlinedigit recognition becomes obvious. Initially, this thesis provides an in-depth newlineanalysis of recognizing the Handwritten Digit Recognition (HDR) methods newlineadopted by various investigators. Recently, the role of Deep Learning (DL) in newlinevarious fields intends to offer the finest solutions with better accuracy. Therefore, newlinethis research uses deep learning approaches for HDR and promises to work newlineefficiently over real-time applications. This research aims to accurately identify newlinethe digits from the benchmark dataset over the thousands of handwritten images newlineand perform various experimentation using diverse algorithms. The investigation newlineare done to evaluate and analyze the efficiency of these algorithms. This research newlineprovides four major phases for HDR: pre-processing, feature extraction, newlineoptimization, and recognition. The Modified National Institute of Standards and newlineTechnology Dataset (MNIST) is used as the input data for providing nonsegregated newlinedigits for further processing in real-time applications. It is composed newlineof two diverse sources known as Special Database 1 (SD-1) and Special Database newline3 (SD-3) which contain the binary images of handwritten digits. newlineIn this first approach, a hybrid Mini-Batch and Stochastic Hessian-Free newlineoptimization (MBSHF) is proposed to attain faster convergence and better newlineaccuracy for predicting the handwritten digits. The anticipated model adopts an newlineiterative minimization algorithm to attain faster convergence with added newlineparameters. The second-order approximation is applied to attain the finest newlineperformance and solve the quadratic equations for defending storage requirements newlineand computational cost. The convex approximation is formulated as an outcome, newlineand performance is measured using a standard MNIST dataset newline