Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.763277
Title: Representation learning for anomaly detection in computer vision
Author: Andrews, Jerone Theodore Alexander
ISNI:       0000 0004 7661 0374
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2018
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
This thesis is a collection of three engineering-based research contributions, aiming to detect anomalous images, without a priori knowledge of the anomaly class. However, devising discriminative data representations in such settings is patently problematic. Obviating the need for explicit prior domain knowledge, this work roots itself in representation learning, using deep convolutional neural networks, charged with solving pseudo tasks. To begin, we investigate unsupervised auto-associative sparse dictionary learning to infer a set of basic elements. Significantly, we show that these elements are not unique to the training data and can be utilised for the faithful reconstruction of anomalous images. Furthermore, we highlight that encoded representations do not always improve upon those in raw pixel space. Moving away from reconstructive-based approaches, in our second contribution, we propose a novel deep distance metric learning approach, generating freely available supervisory signals that exist within visual data. Importantly, we demonstrate that the learnt appearance features can be effectively combined with generic pretrained image representations. Finally, premised on the notion that learning to recognise one kind of object assists with identifying another, we explore supervised inductive transfer learning. Representations are induced by learning to discriminate between different sub-concepts of the normal data, using fine-grained semantic labels. By forming a distribution over the sub-concepts of the normal class, we are able to detect previously unseen samples that deviate from the overarching concept. Notably, we show that current out-of-distribution detectors which utilise the maximum softmax probability, as an anomaly score, are incapable of illuminating the similarity of a novel sample to a universal concept of normality.
Supervisor: Griffin, L. D. ; Nelson, J. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.763277  DOI: Not available
Share: