Fine-grained neural architecture search for image super-resolution

Heewon Kim, Seokil Hong, Bohyung Han, Heesoo Myeong, Kyoung Mu Lee

Abstract: Highlights•A new search space that allows for channel-level architectural changes.•A new algorithm that automatically finds multiple heterogeneous activation functions.•The state-of-the-art lightweight and efficient networks for image super-resolution.•Efficiency improvement of the ClassSR framework using multiple searched networks.