IEICE Information and Communication Technology Forum
A two stage patch-based Markov random field approach to structure-aware image inpainting
Attila Fesus, Tijana Ruzic, Aleksandra Pizurica,
PDF download (10.3MB)
We introduce a two stage patch-based structure aware inpainting method within a Bayesian framework. An original conditional Markov random field (MRF) model is introduced to encode spatial clustering of image patches with missing parts. Within this model an anisotropic MRF encodes spreading of given classes of image content in locally dominant directions. We infer posterior probabilities for each patch to belong to different classes, and use these to select plausible candidate replacement patches. The second stage solves an optimization problem with multiple candidates for each damaged patch, as in recent global inpainting methods. The results demonstrate a clear advantage of our structure-aware candidate selection.