MR-Net: Multiresolution sinusoidal neural networks

Hallison Paz, Daniel Perazzo, Tiago Novello, Guilherme G. Schardong, Luiz Schirmer, Vinícius da Silva, Daniel Yukimura, Fabio Chagas, Hélio Lopes, Luiz Velho

Abstract: Highlights•Framework to train deep networks to represent signals in multiresolution.•Network initialization is used to control the frequencies learned by the model.•MR-Net provides a continuous representation of signals spatially and in scale.•MR-Net represents images with higher or comparable PSNR than other architectures.•Applications in texture magnification and minification, and antialiasing.