Causality considerations for missing data reconstruction in image sequences

The 3D autoregressive (AR) model with a non-causal support region has been successfully employed in the reconstruction of texture and missing regions in image sequences. This paper discusses the causality considerations when selecting the reconstruction model. When a distorted area to be reconstructed is large, a substantial computational load reduction can be obtained by implementing a predictor with a purely causal AR support. A novel reconstruction scheme which employs a selective causal/anti-causal (S-C/AC) AR model is presented. Experimental results suggest that the S-C/AC scheme produces a good trade-off between computational and reconstruction performance.