Image inpainting is the process of correcting undesirable changes to an image in an unobtrusive way.
The existing literature in this research field describes predominantly techniques designed to correct narrow missing regions, which thus often produce undesirable results when the damaged region is large.
This paper presents a novel exemplar-based image inpainting technique for automatic filling-in missing region of an image. Our solution offers two major improvements compared to existing techniques. Patches for filling in missing regions are identified using an appearance space vector, which not only encodes pixel colours, but also colour gradients, feature distances and other measures for computing image similarity. In order to speed up the search for a matching patch we use a Principal Component Analysis to reduce the size of a feature vector used for patch comparison.
The second major improvement is the technique used combine patches filling in a missing region. In order to avoid visible seams we use a Poisson-guided interpolation to blend patches.
Our evaluation and comparison with existing techniques demonstrates significantly improved performance for inpainting missing image regions. |
Cite as: Nguyen, H. M., Wunsche, B. C., Delmas, P. and Lutteroth, C. (2014). Poisson Blended Exemplar-based Texture Completion. In Proc. Thirty-Seventh Australasian Computer Science Conference (ACSC 2014) Auckland, New Zealand. CRPIT, 147. Thomas, B. and Parry, D. Eds., ACS. 99-104 |
(from crpit.com)
(local if available)
|