Multi-scale Recurrent Network for Video Defocus Deblurring

Abstract Video stands as the paramount medium for humans to perceive the world, and within this realm, video defocus deblurring (VDD) emerges as a distinct blurring phenomenon, distinct from both single image defocus deblurring (SIDD) and motion-induced video blurring. Despite its significance, VDD remains an underexplored territory in academic research. In this study, we introduce a novel neural network architecture, grounded in a multi-scale recurrent framework, tailored for VDD. This architecture meticulously propagates information features through a recurrent neural structure, employing a cascaded approach. With a view to minimize computational overhead, we incorporate prior feature fusion techniques within the latent space. Additionally, we introduce a novel metric to quantitatively evaluate the clarity of defocused blur. Our model emerges as the preeminent performer on the synthesized, customized Defocus Video dataset.