Zero-shot inverse imaging with auto-tuned hyperparameters and untrained generative priors

초록

Inverse imaging aims to reconstruct underlying images from noisy and corrupted observations. Conventional model-based methods heavily rely on prior information and hyperparameter tuning, while recent deep learning-based methods inherently require large training datasets. This study presents a robust zero-shot inverse imaging approach based on untrained generative priors. Our method combines a constrained model-based reconstruction with an untrained generative model as a prior, enabling simultaneous network training and target image esti mation through an iterative process. The key contributions of this study include: (1) proposing a new method for simultaneous image reconstruction and prior learning, thereby eliminating the need for training datasets; (2) introducing automated hyperparameter tuning for both image reconstruction and network training; and (3) proposing an ensemble of networks with automated outlier removal, ensuring enhanced robustness of inverse imaging against training failures. Our method is evaluated across various inverse imaging tasks, consistently out performing existing methods without requiring large training datasets or manual hyperparameter tuning. The code is available at https://github.com/pyomin/zerotune.

키워드