The Performance of Image Difference Metrics for Rendered HDR Images

Abstract

HDR is a field in image processing that has received a lot of attention in the later years. Techniques for capturing, tone map back to viewable data has been proposed. Many different ideas have been pursuited, some with a background in the Human Visual System (HVS), but the same problem with determining the quality of these reproductions still exist. In low dynamic range imaging, the solution to the problem has been either to do a visual inspection and compare the reproduction against an original, but as this is a labour intensive and time consuming and highly subjective process, and the need for automated measures which can predict quality has resulted in different image difference metrics. As for comparison of HDR and LDR, this is no trivial task. Currently, no method of automated comparison has been deemed a viable solution due to the difference in in dynamic range. In this master thesis, we present a novel framework extending on recent research which enables us to compare HDR and LDR content, and from this using standard image difference metrics to evaluate the quality of these. These measures are tested against data from a perceptual experiment to verify the stability and quality of the framework. Initial results indicate that the proposed framework enables us to evaluate the quality of such reproductions on the tested scenes, but that some problems are still unsolved.