SCN: Self-Calibration Network for fast and accurate image super-resolution

Deep and wide Convolutional neural networks (CNNs) achieve remarkable performance on image super -resolution (SR), while it is hard to be directly applied on edge devices due to the strict requirements on data transmission and processing. Therefore, numerous lightweight SR methods have been proposed. However, these methods often come with compromises on the reconstruction performance or speed due to the weak reconstruction supervision provided by shallow networks. Aiming to alleviate the above limitation, we introduce a self-calibration network (SCN) equipped with a hierarchical self-calibration module for fast and accurate image SR. Specially, we design a lightweight internal calibration module (ICM) to enable the network to economically focus on the key information by collaboratively refining the weights in both channel and spatial dimensions; and an efficient external calibration module (ECM) to focus on calibrating the prominent error of knowledge learned from the network on Low-Resolution (LR) data for saving computing resources. We experimentally demonstrate the synergy among the two modules guides to a superior SR performance against the state-of-the-art lightweight methods while requiring much fewer parameters and operations.