Evaluation of digital inpainting quality in the context of artwork restoration

Abstract

Improved digital image inpainting algorithms could provide substantial support for future artwork restoration. However, currently, there is an acknowledged lack of quantitative metrics for image inpainting evaluation. In this paper the performance of eight inpainting algorithms is first evaluated by means of a psychophysical experiment. The ranking of the algorithms thus obtained confirms that exemplar based methods generally outperform PDE based methods. Two novel inpainting quality metrics, proposed in this paper, eight general image quality metrics and four inpainting-specific metrics are then evaluated by validation against the perceptual data. Results show that no metric can adequately predict inpainting quality over the entire image database, and that the performance of the metrics is image-dependent. © 2012 Springer-Verlag.