Accurate Image Super-Resolution Using Very Deep Convolutional Networks

The paper proposes a highly accurate method for single-image super-resolution using a very deep convolutional network inspired by the VGG-net used for ImageNet classification. The final model uses 20 weight layers, and the authors find that increasing network depth significantly improves accuracy, because cascading small filters many times in a deep structure exploits contextual information over large image regions in an efficient way.

Because convergence speed becomes a critical issue when training very deep networks, the authors introduce a simple yet effective training procedure: they learn only the residuals and use extremely high learning rates, about 10^4 times higher than SRCNN, made stable by adjustable gradient clipping. The resulting method performs better than existing approaches in accuracy, with visual improvements that are easily noticeable, demonstrating the value of depth and residual learning for super-resolution.