KANO: Interpretable Super-Resolution with Kolmogorov-Arnold Theorem

This paper introduces KANO, a novel interpretable operator for single-image super-resolution (SR) based on the Kolmogorov-Arnold theorem. It addresses the limitations of existing black-box deep learning approaches by providing a transparent and structured representation of the image degradation process. The use of B-spline functions to approximate spectral curves allows for capturing key spectral characteristics and endowing SR results with physical interpretability. The comparative study between MLPs and KANs offers valuable insights into handling complex degradation mechanisms.