Model Adaptation for Inverse Problems in Imaging

Deep neural networks have been applied successfully to a wide variety of inverse problems arising in computational imaging. These are typically trained using forward model that describes the measurement process be inverted, which is often incorporated directly into network itself. However, these approaches sensitive changes model: if at test time varies (even slightly) from one was for, reconstruction performance can degrade substantially. Given solve an initial problem with known model, we propose two novel procedures adapt change even without full knowledge change. Our do not require access more labeled data (i.e., ground truth images). We show simple adaptation achieve empirical success problems, including deblurring, super-resolution, and undersampled image magnetic resonance