CISD: Continuous Implicit Signed Distance Networks for Arbitrary-Scale Super-Resolution

In this study, we tackle the challenge of fixed-scale constraints prevalent in single-image super-resolution (SISR) by introducing a novel continuous image representation technique utilizing signed distance functions. Drawing inspiration from implicit neural representations, we design a Continuous Implicit Signed Distance network (CISD) for arbitrary-scale image superresolution. The CISD network adeptly reconstructs continuous visual signals through the estimation of signed distances to the RGB color manifold, using image coordinates and predicted colors as inputs. Further augmenting our method, we incorporate latent codes to encode distinctive image features, allowing a shared decoder function to efficiently adapt to diverse images, which markedly boosts the model’s practicality and versatility. The qualitative and quantitative results demonstrate that CISD achieve favorable results and outperform the prior works in arbitrary super-resolution tasks.