Adaptive 2D-AR framework for texture completion

Texture extrapolation techniques enable to fill large holes of missing information. Many applications can be targeted such as image and video coding, channel block losses, object removal, filling of 3D disocclusions etc. For more than two decades, many approaches have been developed, even though each contains pros and cons which force to choose the best compromise for the targeted application. In this paper, we propose to continue exploring and improving a popular parametric completion method using the autoregressive (AR) model. In this framework, the training area is automatically optimized. A consistency criterion also enables to assess and regularize the model. Moreover, a post-processing step enables to remove the remaining seam artefacts. A comparison with the state-of-the-art is provided for both subjective quality and complexity which remains a major constraint for texture completion.