Virtual Completion of Facial Image in Ancient Murals

Restore damaged face images is an important issue for ancient murals conservation and research. Since the structure and style of facial image varies with identity, dynasties, location and artists, it is a time consuming and challenging job for researchers to achieve objective and precise restoration. In this paper, we propose a novel facial image completion method based on a semantic learning framework. By few user interactions, our approach can fill missing regions with a reconstructed face which is based on candidate images with high semantic similarity. Since most of the ancient murals contain deteriorations, to archive visual pleasant result, we adopt a PCA based region blending method to produce a seamless completion. Moreover, for the criteria of mural restoration are evaluated in the view of both scientific and art, our completion process can provide the researcher an opportunity to preview a diverse set of the final results virtually before they actually start their job. The real applications in Dunhuang mural show the efficiency and effectiveness of our approach. Furthermore, these experiments suggested the potential applications for restoration facial image in ancient line drawing.