Sketched Equivariant Imaging Regularization and Deep Internal Learning for Inverse Problems

  • 2025-06-06 18:52:23
  • Guixian Xu, Jinglai Li, Junqi Tang
  • 0

Abstract

Equivariant Imaging (EI) regularization has become the de-facto technique forunsupervised training of deep imaging networks, without any need ofground-truth data. Observing that the EI-based unsupervised training paradigmcurrently has significant computational redundancy leading to inefficiency inhigh-dimensional applications, we propose a sketched EI regularization whichleverages the randomized sketching techniques for acceleration. We apply oursketched EI regularization to develop an accelerated deep internal learningframework, which can be efficiently applied for test-time network adaptation.Additionally, for network adaptation tasks, we propose a parameter-efficientapproach to accelerate both EI and Sketched-EI via optimizing only thenormalization layers. Our numerical study on X-ray CT and multicoil magneticresonance image reconstruction tasks demonstrate that our approach can achievesignificant computational acceleration over standard EI counterpart insingle-input setting and network adaptation at test time.

 

Quick Read (beta)

loading the full paper ...