Inversion Methods for Computational Imaging (CTh5A)

Presider: Vidya Ganapati, Swarthmore College
Authentication Required

Access to this content is only available to registered attendees. Please login now to access the live content.

15:00 - 15:30
(UTC - 07:00)

Model Adaptation for Inverse Problems in Imaging (CTh5A.1)
Presenter: Gregory Ongie, Marquette University

Deep networks trained to solve inverse problems in imaging can be fragile to changes in the forward model at deployment. This talk focuses on novel algorithms and retraining procedures to resolve this issue.

Authors:Gregory Ongie, Marquette University

15:30 - 15:45
(UTC - 07:00)

Solving Inverse Problems using Self-Supervised Deep Neural Nets (CTh5A.2)
Presenter: Jiapeng Liu, Southern Methodist University

A modular framework combining the expressive power of generative models with physics-assisted learning is proposed to solve inverse problems. The process is iterative, unsupervised, and only requires knowledge of the physical/forward model.

Authors:Jiapeng Liu, Southern Methodist University / Muralidhar Madabhushi Balaji, Southern Methodist University / Chris Metzler, University of Maryland / M.Salman Asif, University of California / Prasanna Rangarajan, Southern Methodist University

  Paper

15:45 - 16:15
(UTC - 07:00)

Interpretable Neural Networks for Solving Dictionary Learning Problems in Neuroscience and Imaging (CTh5A.3)
Presenter: Demba Ba, Harvard Medical School

Abstract not available.

Authors:Demba Ba, Harvard Medical School

16:15 - 16:30
(UTC - 07:00)

Shift-Variant Deblurring for Rotationally Symmetric Systems (CTh5A.4)
Presenter: Amit Kohli, University of California, Berkeley

We present a fast image deblurring method for rotationally symmetric systems with spatially-varying aberrations. We calibrate our method with point spread function measurements along a line. Our method outperforms standard deconvolution on the UCLA Miniscope.

Authors:Amit Kohli, University of California, Berkeley / Anastasios Angelopoulos, University of California, Berkeley / Sixian You, University of California, Berkeley / Laura Waller, University of California, Berkeley

  Paper

16:30 - 16:45
(UTC - 07:00)

MultiWienerNet: Deep Learning for Fast Shift-Varying Deconvolution (CTh5A.5)
Presenter: Richard Shuai, University of California, Berkeley

We present a deep-learning method based on Wiener filters and U-Nets that performs image reconstruction in systems with spatially-varying aberrations. We train on simulated microscopy measurements and test on experimental data, demonstrating high resolution reconstructions.

Authors:Richard Shuai, University of California, Berkeley / Kyrollos Yanny, University of California, Berkeley / Kristina Monakhova, University of California, Berkeley / Laura Waller, University of California, Berkeley

  Paper

16:45 - 17:00
(UTC - 07:00)

Learning network for laser absorption imaging in flames using mid-fidelity simulations (CTh5A.6)
Presenter: Chuyu Wei, University of California Los Angeles

A deep neural network is trained using mid-fidelity reacting flow simulations to assist laser absorption imaging of species and temperature in flames with sparse view angles. The method is compared to linear tomography.

Authors:Chuyu Wei, University of California Los Angeles / Kevin Schwarm, University of California Los Angeles / Daniel Pineda, University of Texas at San Antonio / R. Mitchell Spearrin, University of California Los Angeles

  Paper