Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Image Classification using Functional Analysis and Neural Networks

Wendsjö, Albert LU (2021) STAN40 20211
Department of Statistics
Abstract
In this thesis a number of methods for image classification are investigated. The main goal is to explore an approach based on functional data analysis and compare it with Neural Networks. The main idea is to view each image as a function and then capture the most important variation of each image class through Functional Principal Component Analysis. For comparing the methods the cifar10 dataset is used, which consist of 60000 32x32 rgb images with 10 different classes. Results indicate that the functional approach works, yielding an accuracy of 31%, where a standard Neural Network had an accuracy of 29%. A Neural Network trained on the Principal Component scores performed at 34\%. However the Convolutional Neural Networks still... (More)
In this thesis a number of methods for image classification are investigated. The main goal is to explore an approach based on functional data analysis and compare it with Neural Networks. The main idea is to view each image as a function and then capture the most important variation of each image class through Functional Principal Component Analysis. For comparing the methods the cifar10 dataset is used, which consist of 60000 32x32 rgb images with 10 different classes. Results indicate that the functional approach works, yielding an accuracy of 31%, where a standard Neural Network had an accuracy of 29%. A Neural Network trained on the Principal Component scores performed at 34\%. However the Convolutional Neural Networks still outperformed all other methods with an accuracy of 45% when using 1D convolutions and 66% when using 2D convolutions. Lastly some possible possible improvements of the functional approach is discussed and areas for future research. (Less)
Please use this url to cite or link to this publication:
author
Wendsjö, Albert LU
supervisor
organization
course
STAN40 20211
year
type
H1 - Master's Degree (One Year)
subject
keywords
image classification, splines, functional data analysis, eigenfunctions, neural networks
language
English
id
9054846
date added to LUP
2021-08-06 08:11:35
date last changed
2021-08-06 08:11:35
@misc{9054846,
  abstract     = {{In this thesis a number of methods for image classification are investigated. The main goal is to explore an approach based on functional data analysis and compare it with Neural Networks. The main idea is to view each image as a function and then capture the most important variation of each image class through Functional Principal Component Analysis. For comparing the methods the cifar10 dataset is used, which consist of 60000 32x32 rgb images with 10 different classes. Results indicate that the functional approach works, yielding an accuracy of 31%, where a standard Neural Network had an accuracy of 29%. A Neural Network trained on the Principal Component scores performed at 34\%. However the Convolutional Neural Networks still outperformed all other methods with an accuracy of 45% when using 1D convolutions and 66% when using 2D convolutions. Lastly some possible possible improvements of the functional approach is discussed and areas for future research.}},
  author       = {{Wendsjö, Albert}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Image Classification using Functional Analysis and Neural Networks}},
  year         = {{2021}},
}