Do it yourself dynamic single image super resolution network via ODE

Abstract

Single Image Super Resolution (SISR) aims at characterizing fine-grain information given a low-resolution image. Recent progress shows that SISR can be viewed as a dynamic process that can be modeled using Ordinary Differential Equations (ODEs). As a result, ODE inspired neural network shows superior performance with limited number of parameters, as well as interpretability for network structure. However, the current ODE based approach restricts the neural network structure to a static single-branch residual network, while dynamic structures can adaptively adjust their parameters(or even structures) best suitable for each test image and lead to better SISR performance. To take advantage of ODE and dynamic network structures in both, we introduce the Implicit Runge–Kutta scheme to construct an ODE-inspired multi-branch residual module that serves as a basic module, which is helpful to capture information at different scales. Then, an attention module is applied on the weights of the Implicit Runge–Kutta scheme to obtain a new dynamic network module, which is equivalent to encourage different branch to jointly attend different positions to obtain the best performance. Experiments demonstrate that our approach outperforms state-of-the-art ODE-inspired methods with less or comparable number of parameters.