Lightweight MIMO-WNet for single image deblurring

Single image deblurring, aiming at recovering a latent sharp image from a blurry image, is a highly ill-posed task as there exist infinite feasible solutions. One successful practice of the existing popular approaches is to extensively stack deep networks to regress the complicated relationships between the sharp and blurry image pairs, which results in inevitably high computational costs. In this work, we present a synergistic framework named Lightweight MIMO-WNet that optimally balances the performance and the computational costs. Our main proposal consists of a MIMO-WNet architecture that attempts to learn the complicated blurry-sharp relationships via balancing the spatial details and the high-level contextualized information, and a multiple information refining block (MIRB) that reduces the parameters while deepening the network. At each layer, the original ResBlock is replaced with the MIRB that divides the input into multiple parts and refines the information hierarchically. The experimental results on three benchmarks demonstrate that Lightweight MIMO-WNet obtains a better trade-off between the performance and the model complexity.