A comparative study on wavelets and residuals in deep super resolution

Despite the advances in single-image super resolution using deep convolutional networks, the main problem remains unsolved: recovering fine texture details. Recent works in super resolution aim at modifying the training of neural networks to enable the recovery of these details. Among the different method proposed, wavelet decomposition are used as inputs to super resolution networks to provide structural information about the image. Residual connections may also link different network layers to help propagate high frequencies. We review and compare the usage of wavelets and residuals in training super resolution neural networks. We show that residual connections are key in improving the performance of deep super resolution networks. We also show that there is no statistically significant performance difference between spatial and wavelet inputs. Finally, we propose a new super resolution architecture that saves memory costs while still using residual connections, and performing comparably to the current state of the art.