Home

Awesome

Equivariant Imaging: Learning Beyond the Range Space

arXiv GitHub Stars

Equivariant Imaging: Learning Beyond the Range Space

Dongdong Chen, Julián Tachella, Mike E. Davies.

The University of Edinburgh

In ICCV 2021 (oral)

flexible flexible Figure: Learning to image from only measurements. Training an imaging network through just measurement consistency (MC) does not significantly improve the reconstruction over the simple pseudo-inverse (<img src="https://render.githubusercontent.com/render/math?math=A^{\dagger}y">). However, by enforcing invariance in the reconstructed image set, equivariant imaging (EI) performs almost as well as a fully supervised network. Top: sparse view CT reconstruction, Bottom: pixel inpainting. PSNR is shown in top right corner of the images.

EI is a new self-supervised, end-to-end and physics-based learning framework for inverse problems with theoretical guarantees which leverages simple but fundamental priors about natural signals: symmetry and low-dimensionality.

Get quickly started

Overview

The problem: Imaging systems capture noisy measurements <img src="https://render.githubusercontent.com/render/math?math=y\in R^m"> of a signal <img src="https://render.githubusercontent.com/render/math?math=x\in R^n"> through a linear operator <img src="https://render.githubusercontent.com/render/math?math=A\in R^{m\times n}: y"><img src="https://render.githubusercontent.com/render/math?math==Ax"> + <img src="https://render.githubusercontent.com/render/math?math=\epsilon">. We aim to learn the reconstruction function <img src="https://render.githubusercontent.com/render/math?math=f(y)=x"> where

The challenge:

The motivation:

We assume the signal set <img src="https://render.githubusercontent.com/render/math?math=\mathcal{X}"> has a low-dimensional structure and is invariant to a groups of transformations <img src="https://render.githubusercontent.com/render/math?math=T_g"> (orthgonal matrix, e.g. shift, rotation, scaling, reflection, etc.) related to a group <img src="https://render.githubusercontent.com/render/math?math=\mathcal{G}">, such that <img src="https://render.githubusercontent.com/render/math?math=T_gx\in \mathcal{X}"> and the sets <img src="https://render.githubusercontent.com/render/math?math=T_g\mathcal{X}"> and <img src="https://render.githubusercontent.com/render/math?math=\mathcal{X}"> are the same. For example,

Key observations:

overview Figure: Learning with and without equivariance in a toy 1D signal inpainting task. The signal set consists of different scaling of a triangular signal. On the left, the dataset does not enjoy any invariance, and hence it is not possible to learn the data distribution in the nullspace of <img src="https://render.githubusercontent.com/render/math?math=A">. In this case, the network can inpaint the signal in an arbitrary way (in green), while achieving zero data consistency loss. On the right, the dataset is shift invariant. The range space of <img src="https://render.githubusercontent.com/render/math?math=A^{\top}"> is shifted via the transformations <img src="https://render.githubusercontent.com/render/math?math=T_g">, and the network inpaints the signal correctly.

Equivariant Imaging: to learn <img src="https://render.githubusercontent.com/render/math?math=f"> by using only measurements <img src="https://render.githubusercontent.com/render/math?math=\{y_i\}">, all you need is to:

  1. define a transformation group <img src="https://render.githubusercontent.com/render/math?math={T_g}"> based on the certain invariances to the signal set.
  2. define a neural reconstruction function <img src="https://render.githubusercontent.com/render/math?math=f_\theta: y\rightarrow x">, e.g. <img src="https://render.githubusercontent.com/render/math?math=f_\theta=G_\theta \circ A^{\dagger}"> where <img src="https://render.githubusercontent.com/render/math?math=A^{\dagger}\in R^m\rightarrow R^n"> is the (approximated) pseudo-inverse of <img src="https://render.githubusercontent.com/render/math?math=A"> and <img src="https://render.githubusercontent.com/render/math?math=G_\theta: R^n\rightarrow R^n"> is a UNet-like neural net.
  1. calculate <img src="https://render.githubusercontent.com/render/math?math=x^{(1)}=f_\theta(y)"> as the estimation of <img src="https://render.githubusercontent.com/render/math?math=x">.
  2. calculate <img src="https://render.githubusercontent.com/render/math?math=x^{(2)}=T_gx^{(1)}"> by transforming <img src="https://render.githubusercontent.com/render/math?math=x^{(1)}">.
  3. calculate <img src="https://render.githubusercontent.com/render/math?math=x^{(3)}=f_\theta(Ax^{(2)})"> by reconstructing <img src="https://render.githubusercontent.com/render/math?math=x^{(2)}"> from its measurement <img src="https://render.githubusercontent.com/render/math?math=Ax^{(2)}">.

flowchart

Requirements

All used packages are listed in the Anaconda environment.yml file. You can create an environment and run

conda env create -f environment.yml

Test

We provide the trained models used in the paper which can be downloaded at Google Drive. Please put the downloaded folder 'ckp' in the root path. Then evaluate the trained models by running

python3 demo_test_inpainting.py

and

python3 demo_test_ct.py

Train

To train EI for a given inverse problem (inpainting or CT), run

python3 demo_train.py --task 'inpainting'

or run a bash script to train the models for both CT and inpainting tasks.

bash train_paper_bash.sh

Train your models

To train your EI models on your dataset for a specific inverse problem (e.g. inpainting), run

python3 demo_train.py --h

Citation

@inproceedings{chen2021equivariant,
    title     = {Equivariant Imaging: Learning Beyond the Range Space},
    author    = {Chen, Dongdong and Tachella, Juli{\'a}n and Davies, Mike E},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2021},
    pages     = {4379-4388}}