IEICE Information and Communication Technology Forum
Analysis of coupled dictionary learning for super-resolving video of mixed resolution
Aleksandar Latic, Adrian Munteanu, Aleksandra Pizurica,
PDF download (1.1MB)
In this paper, we analyse the performance of coupled dictionary learning for video super-resolution. We make anextension of 2D coupled dictionary learning to a 3D scheme, where 3D atoms model temporal evolution of spatially collocated patches and require no explicit motion estimation. Our analysis shows the influence of different design parameters, such as the effect of a periodic dictionary re-training within a mixed resolution framework, dictionary size and the sparsity of reconstruction, as well as the choice between uni-directional and bi-directional reconstruction strategies. We believe that these results will be helpful in understanding better the potentials and current limitations of coupled dictionary learning in video super-resolution.