LSR plus plus : An Efficient and Tiny Model for Image Super-Resolution

A single-image lightweight super-resolution (LSR) method was presented in our earlier work. LSR predicts the residual image between the interpolated low-resolution (ILR) and high-resolution (HR) images using a self-supervised framework. A further enhanced version of LSR, called LSR++, is proposed in this work. The improvement of LSR++ over LSR primarily lies in a new patch alignment scheme, where the spatially geometrical property of patches is exploited. Patches can be categorized into two main classes through alignment. Then, one regressor is needed for each class. Compared with LSR, LSR++ lowers the computational complexity from 3.83K to 1.71K FLOPs/pixel and the model size from 770K to 200K parameters while preserving close visual performance measured by PSNR/SSIM. LSR++ offers an attractive super-resolution method for mobile applications.