Deep Autoresolution Networks

Gabriel Pereyra, Christian Szegedy

Abstract:

Despite the success of very deep convolutional neural networks, they currently operate
at very low resolutions relative to modern cameras. Visual attention mechanisms
address this by allowing models to access higher resolutions only when
necessary. However, in certain cases, this higher resolution isn’t available. We
show that autoresolution networks, which learn correspondences between lowresolution
and high-resolution images, learn representations that improve lowresolution
classification - without needing labeled high-resolution images.

Conflicts: google.com, usc.edu