Joint Back Projection and Residual Networks for Efficient Image Super-Resolution

Benefiting from the great power of graphic processor units, researchers can now come up with more and more sophisticated and complex deep learning structures to solve computer vision problems in various fields with excellent results. However, real-time performance is the bottleneck for deep learning in some applications, like image super-resolution. In this paper, we propose an image super-resolution making use of both the advantages of Back Projection and Residual Networks (BPRN). It generalizes the residual networks as a hierarchical back projection process. We use both convolution and deconvolution to down- and up-sample images to feedback the residues for super-resolution. Furthermore, we come up with a Lighter BPRN (L-BPRN) model to achieve similar state-of-the-art PSNR but fewer network parameters. The testing process is much faster and also accurate for image super-resolution with different scaling factors. Compared with recent deep learning based image super-resolution approaches, experimental results show that our proposed methods can achieve the state-of-the-art PSNR and SSIM performance as well as fast realization.