Learning Reduced-Resolution and Super-Resolution Networks in Synch
share
Recent studies have shown that deep convolutional neural networks achieve the excellent performance on image super-resolution. However, CNN-based methods restore the super-resolution results depending on interpolations a lot. In this paper, we present an end-to-end network (Reduced & Super-Resolution Network, RSRNet) for reduced-resolution and super-resolution without any cheap interpolation. 1) The reduced-resolution network is trained without supervision, which preserves more effective information and brings better visual effect in the LR image. 2) The super-resolution network learns a sub-pixel residual with dense connection to accelerate the convergence and improve the performance. For image super-resolution, we evaluate the proposed network on four benchmark datasets and set a new state of the art. Moreover, the reduced-resolution network in RSRNet is applicable to generate photo-realistic low-resolution image and improve image compression by using existing image codecs.
READ FULL TEXT