Use this URL to cite or link to this record in EThOS:
Title: Visual recognition of human rights violations
Author: Kalliatakis, Grigorios
ISNI:       0000 0004 8502 3155
Awarding Body: University of Essex
Current Institution: University of Essex
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
This thesis is concerned with the automation of human rights violation recognition in images. Solving this problem is extremely beneficial to human rights organisations and investigators, who are often interested in identifying and documenting potential violations of human rights within images. It will allow them to avoid the overwhelming task of analysing large volumes of images manually. However, visual recognition of human rights violations is challenging and previously unattempted. Through the use of computer vision, the notion of visual recognition of human rights violations is forged in this thesis, whilst this area is addressed by strongly considering the constraints related to the usability and flexibility of a real practice. Firstly, image datasets of human rights violations which are suitable for training and testing modern visual representations, such as convolutional neural networks (CNNs) are introduced for the first time ever. Secondly, we develop and apply transfer learning models specific to the human rights violation recognition problem. Various fusion methods are proposed for performing an equivalence and complementarity analysis of object-centric and scene-centric deep image representations for the task of human rights violation recognition. Additionally, a web demo for predicting human rights violations that may be used directly by human rights advocates and analysts is developed. Next, the problem of recognising displaced people from still images is considered. To solve this, a novel mechanism centred around the level of control each person feels of the situation is developed. By leveraging this mechanism, typical image classification turns into a uniform framework that infers potential displaced people from images. Finally, a human-centric approach for recognising rich information about two emotional states is proposed. The derived global emotional traits are harnessed alongside a data-driven CNN classifier to efficiently infer two of the most widespread modern abuses against human rights, child labour and displaced populations.
Supervisor: Not available Sponsor: ESRC
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available