EFFICIENT CONVOLUTIONAL NEURAL NETWORK FOR IMAGE CLASSIFICATION

TAMANG, YOGENDRA

EFFICIENT CONVOLUTIONAL NEURAL NETWORK FOR IMAGE CLASSIFICATION

Publisher

Pulchowk Campus

Abstract

Problem of object recognition and image classification cannot be only solved by image processing techniques but can also by machine learning technology. This thesis is about classifying images using machine learning techniques. Artificial Neural Networks are be deployed to work with this problem. One of deep learning architectures, Convolutional Neural Network (CNN) is specifically used to learn the features of images. To design such CNN, different parameters are taken into account. These parameters may be filter size, number of convolution layers, Dropout layers etc. By careful choosing and study of these parameters, somehow efficient architecture is designed. On this thesis, Different neural network architectures for CIFAR-10 and MNIST dataset are developed. These networks are different in terms of number of hidden layers, filter sizes and other measures. They are trained on Graphical Processing Unit (GPU). Dropout technique is used for reducing over-fitting issues. Networks are then fed with testing data and the accuracy of these architectures are calculated. The research is run by continuous evaluation of accuracy of different architectures. Finally results are compared and analysed to find out best architecture thus giving a way to design efficient architecture for image classification.

Description

Problem of object recognition and image classification cannot be only solved by image processing techniques but can also by machine learning technology.