Generator Surgery for Compressed Sensing

Published in NeurIPS 2020 Workshop Deep Inverse, 2020

Recent work has explored the use of generator networks with low latent dimension as signal priors for image recovery in compressed sensing. However, the recovery performance of such models is limited by high representation error. We introduce a method to reduce the representation error of such generator signal priors by cutting one or more initial blocks at test time and optimizing over the resulting higher-dimensional latent space. Experiments demonstrate significantly improved recovery for a variety of architectures. This approach also works well for out-of-training-distribution images and is competitive with other state-of-the-art methods. Our experiments show that test-time architectural modifications can greatly improve the recovery quality of generator signal priors for compressed sensing.

Article Link


  title={Generator Surgery for Compressed Sensing},
  author={Jung Yeon Park and Niklas Smedemark-Margulies and Max Daniels and Rose Yu and Jan-Willem van de Meent and Paul Hand},
  booktitle={NeurIPS 2020 Workshop on Deep Learning and Inverse Problems},