Enhanced Deep Residual Networks for Single Image Super-Resolution

The paper develops enhanced deep networks for single-image super-resolution, building on deep convolutional neural networks and residual learning. The core EDSR model improves on conventional residual networks by removing unnecessary modules and then expanding the model size while stabilizing the training procedure. The authors additionally propose MDSR, a multi-scale architecture and training method that can reconstruct high-resolution images at different upscaling factors within a single model.

These optimizations let EDSR and MDSR exceed the performance of contemporary state-of-the-art super-resolution methods on standard benchmark datasets. The approach proved its strength by winning the NTIRE2017 Super-Resolution Challenge, demonstrating that carefully simplified and scaled residual architectures can deliver leading super-resolution quality.