Digital restoration of historical photographic slides

This thesis aims to develop an automated metric for assessing colour cast severity, develop a new colour restoration algorithm, and provide a framework for systematic evaluation of algorithm performance using both human subjects and the newly developed automated metric.;The results of this study form a basis for the hypothesis that the context in which restored images are shown affects human perception of colour quality. This hypothesis is validated experimentally and the results indicate that if restored images are mixed with undamaged ones, their quality is assessed differently than when they are shown alongside their original damaged versions.;A model of assessing colour cast severity is built using image colour features. Multiple sets of faded slides, their restored versions, and undamaged digital photographs, are assessed by human observers. Lab space-based statistics are calculated from these images and are used to develop a regression model for automatically assessing colour cast severity.;A new colour restoration algorithm is developed from one of the nine algorithms tested earlier and is evaluated by both human participants and the new model.;It evaluates nine existing white-balancing algorithms using faded slides and synthetic images that simulate the appearance of faded slides. This is done by human subjects and, for the synthetic images, using several numerical metrics based on the RMSE and distance in CIE Lab colour space. Results indicate that none of the examined algorithms perform an acceptable restoration of real faded slides. Furthermore, RMSE metrics are not indicative of the human judgement of colour quality.;This work allows for automated measurement of colour cast presence in faded slides before and after restoration. The new method for restoring colour does work better than the existing methods to which it was compared. Results show, however, that assessment by human observers still remains the most accurate means of evaluating colour restoration algorithms. Model development, the evaluation methodology, and results are fully described to ensure replicability and aid future improvements.