Image Registration

Revision as of 15:11, 15 May 2020 by David (talk | contribs)
\( \newcommand{\P}[]{\unicode{xB6}} \newcommand{\AA}[]{\unicode{x212B}} \newcommand{\empty}[]{\emptyset} \newcommand{\O}[]{\emptyset} \newcommand{\Alpha}[]{Α} \newcommand{\Beta}[]{Β} \newcommand{\Epsilon}[]{Ε} \newcommand{\Iota}[]{Ι} \newcommand{\Kappa}[]{Κ} \newcommand{\Rho}[]{Ρ} \newcommand{\Tau}[]{Τ} \newcommand{\Zeta}[]{Ζ} \newcommand{\Mu}[]{\unicode{x039C}} \newcommand{\Chi}[]{Χ} \newcommand{\Eta}[]{\unicode{x0397}} \newcommand{\Nu}[]{\unicode{x039D}} \newcommand{\Omicron}[]{\unicode{x039F}} \DeclareMathOperator{\sgn}{sgn} \def\oiint{\mathop{\vcenter{\mathchoice{\huge\unicode{x222F}\,}{\unicode{x222F}}{\unicode{x222F}}{\unicode{x222F}}}\,}\nolimits} \def\oiiint{\mathop{\vcenter{\mathchoice{\huge\unicode{x2230}\,}{\unicode{x2230}}{\unicode{x2230}}{\unicode{x2230}}}\,}\nolimits} \)

Image registration is recovering an affine transformation (rotation + translation) between two images.

Image registration can be performed in the frequency domain with the Fourier transform or in the spatial domain with a log-polar transformation.

Problem Statement

We are given two images \(\displaystyle I_1\) and \(\displaystyle I_2\).
Let \(\displaystyle (x,y)\) be uv coordinates within the image.
We want to find a rotation and translation from \(\displaystyle (x,y)\) to \(\displaystyle (x',y')\) such that \(\displaystyle I_1(x,y) = I_2(x', y')\).
This is represented as:
\( \begin{align} x' &= a_1 x + a_2 y + a_3\\ y' &= a_4 x + a_5 y + a_6 \end{align} \)
This can also be written as:
\(\displaystyle \begin{pmatrix} x' \\ y' \\ 1 \end{pmatrix} = \begin{pmatrix} a_1 & a_2 & a_3\\ a_4 & a_5 & a_6\\ 0 & 0 & 1 \end{pmatrix} \begin{pmatrix} x \\ y \\ 1 \end{pmatrix} \)

Log-Polar Transformation

This is copied from Wolberg and Zokai[1].

The log-polar transformation is defined as follows:
\( \begin{align} b &= \log(r) = \log\left(\sqrt{(x-x_c)^2 + (y-y_c)^2}\right)\\ a &= \operatorname{arctan2}(y-y_c, x-x_c) \end{align} \)
where \(\displaystyle (x_c, y_c)\) is the center of the image and \(\displaystyle r\) is the distance from the center of the image.

Here a rotation in Cartesian coordinates \(\displaystyle (x, y)\) around the center \((x_c, y_c)\) corresponds to a shift in \(a\) in log-polar coordinates.

A scale change (i.e. enlarge or stretch) is a shift in log-space:
\( \lambda r \mapsto \log(\lambda r) = \log(\lambda) + \log(r) \)

These translations can be found using Wikipedia: Cross-correlation.

Algorithm

For each resolution from coarse to fine, do the following:

  1. Crop central region \(\displaystyle I_1'\) from \(\displaystyle I_1\)
  2. Compute the low-polar transformation \(\displaystyle I_{1p}'\)
  3. For all positions \((x,y)\)
    1. Crop region \(I_{2p}'\)
    2. Compute \(I_{2p}'\)
    3. Cross-correlate \(I_{1p}'\) and \(I_{2p}'\) to get \((dx, dy)\)
    4. If max correlation, save \((x, y)\) and \((dx, dy)\)
  4. Scale = \(dx\)
  5. Rotation = \(dy\)
  6. Translation = \(x, y\)

References

  1. George Wolberg, and Siavash Zokai (2000). Robust Image Registration Using Log-Polar Transform DOI: 10.1109/ICIP.2000.901003 URL: https://home.cis.rit.edu/~cnspci/references/wolberg2000.pdf