Equivariant Imaging: Unsupervised Learning with Symmetry for Scientific Imaging

Math+ML+X Seminar Series

Organizer：

Angelica Aviles-Rivero

Speaker：

Dongdong Chen (Heriot-Watt University)

Time：

Thur., 16:00, May 29, 2025

Online：Voov (Tencent): 957-841-954

Title：

Equivariant Imaging: Unsupervised Learning with Symmetry for Scientific Imaging

Abstract:

Deep networks provide state-of-the-art performance in many inverse imaging problems, ranging from medical imaging to computational photography. In several scientific imaging problems, we typically only have access to incomplete and noisy measurements of the underlying signals, which complicates most learning-based strategies that typically require pairs of signals and associated measurements for training. Learning only from incomplete measurements is generally impossible, as the incomplete observations do not contain information outside the range of the forward sensing operator. In this talk, I will present our recent work on Equivariant Imaging (EI), which overcomes this limitation by exploiting the equivariances present in natural signals. Our proposed learning strategy is fully unsupervised and performs as well as fully supervised methods. Experiments demonstrate the potential of this framework on inverse problems including sparse-view X-ray computed tomography, accelerated MRI, and image inpainting. Finally, I will briefly introduce our DeepInverse library - a free and award-winning software for inverse imaging problems.