Efficient Module Based Single Image Super Resolution for Multiple Problems

Example based single image super resolution (SR) is a fundamental task in computer vision. It is challenging, but recently, there have been significant performance improvements using deep learning approaches. In this article, we propose efficient module based single image SR networks (EMBSR) and tackle multiple SR problems in NTIRE 2018 SR challenge by recycling trained networks. Our proposed EMBSR allowed us to reduce training time with effectively deeper networks, to use modular ensemble for improved performance, and to separate subproblems for better performance. We also proposed EDSR-PP, an improved version of previous ESDR by incorporating pyramid pooling so that global as well as local context information can be utilized. Lastly, we proposed a novel denoising / deblurring residual convolutional network (DnResNet) using residual block and batch normalization. Our proposed EMBSR with DnResNet and EDSR-PP demonstrated that multiple SR problems can be tackled efficiently and effectively by winning the 2nd place for Track 2 (× 4 SR with mild adverse condition) and the 3rd place for Track 3 (×4 SR with difficult adverse condition). Our proposed method with EDSR-PP also achieved the ninth place for Track 1 (×8 SR) with the fastest run time among top nine teams.