Please use this identifier to cite or link to this item:
Title: Deep learning for anomaly detection in computational imaging
Authors: Du, Xinglin
Keywords: Engineering::Computer science and engineering::Computing methodologies::Pattern recognition
Engineering::Electrical and electronic engineering
Issue Date: 2021
Publisher: Nanyang Technological University
Source: Du, X. (2021). Deep learning for anomaly detection in computational imaging. Master's thesis, Nanyang Technological University, Singapore.
Abstract: One of the applications of deep learning is anomaly detection. In this thesis, supervised and semi-supervised deep learning anomaly detection are compared. For supervised method, three methods are used: multilayer perceptron, convolutional neural network and transfer learning. Multilayer perceptron and convolution neural network are compared in MNIST and Fashion-MNIST dataset, which comes out that convolution layers are more suitable for image input. Transfer learning is used in marble surface dataset to avoid data imbalance problems. VGG 16 and Dense 201 pre-trained model are used and VGG 16 is the better one. For semi-supervised method, autoencoder(AE) is introduced. Since the idea of AE is to compare the difference between input and output, firstly in MNIST, two criteria are discussed: Euclidean distance and cosine similarity. Based on results, cosine similarity can have a better result. Then in Fashion-MNIST dataset, fully connected layers AE and convolution AE are compared, and convolution AE leads to a better performance. For marble surface dataset, conventional AE and MemAE are compared. Based on the result, MemAE can inhibit the generalization ability and have a better result both in theory and in practice.
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:EEE Theses

Files in This Item:
File Description SizeFormat 
Deep Learning for Anomaly Detection_amended.pdf
  Restricted Access
3.72 MBAdobe PDFView/Open

Page view(s)

Updated on May 14, 2022


Updated on May 14, 2022

Google ScholarTM


Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.