Video super-resolution reconstruction based on correlation learning and spatio-temporal nonlocal similarity

A novel video super-resolution reconstruction algorithm based on correlation learning and spatio-temporal nonlocal similarity is proposed in this paper. Objective high-resolution (HR) estimates of low-resolution (LR) video frames can be obtained by learning LR-HR correlation mapping and fusing the spatio-temporal nonlocal similarity information between video frames. First, the LR-HR correlation mapping between LR and HR patches is established based on semi-coupled dictionary learning. With the aim of improving algorithm efficiency while guaranteeing super-resolution quality, LR-HR correlation mapping is performed only for the salient object region, and then an improved visual saliency-based nonlocal fuzzy registration scheme using the pseudo-Zernike moment feature and structural similarity is proposed for spatio-temporal similarity matching and fusion. Visual saliency and self-adaptive regional correlation evaluation strategies are used in spatio-temporal similarity matching to improve algorithm efficiency further. Experimental results demonstrate that the proposed algorithm achieves competitive super-resolution quality compared to other state-of-the-art algorithms in terms of both subjective and objective evaluations.