| dc.description.abstract | This thesis describes the process of developing and evaluating a design science research artefact for anonymising faces in images. The artefact is developed for use in RRI-research in the ViSmedia research group at the University of Bergen. A facial anonymisation algorithm is defined as a system which can provide facial deidentification, while retaining face realism and preserving image quality. Research from psychology and artificial intelligence serve as the theoretical background for the development, where human face perception, computer vision and machine learning are topics of interest. Face morphing techniques are used to manipulate faces according to the facial anonymisation definition. The thesis aims to answer three questions. First, it is asked if it possible to to create an artefact capable of automatically anonymising faces using an average face morphing approach. Secondly, if the artefact can be used to investigate the views on normative practices in manipulation and usage of face photographies in the news media. Finally, it is investigated if there are other possible use cases for the artefact or its technology. Two prototypes are proposed for solving the anonymisation problem. Based on findings from a preliminary development prototype, both a featural and a holistic anonymisation prototype is developed and evaluated. The evaluations are a combination of performance The featural prototype performs poorly on de-identification, while the holistic prototype shows more promise. Both are decent at providing realistic faces and in image quality preservation. The facial anonymisation algorithm is argued to have potential as an investigative tool, and several potential use cases are proposed. The artefact in its current state is but a proof of concept and its limitations and challenges are described. Several proposals for future development are also provided in order to produce a more complete solution | en_US |
