NEW: Alignment module

[carterbox]

Purpose

Adds Fourier-based shift operator and Fourier-based alignment solver to use for alignment problems.

Approach

Ported the shift operator and translation registration functions from nikitinvv/deformcg. These functions use multiplication and cross-correlation in the Fourier space to shift or align stacks of 2D images.

I also moved the FFT plan caching functionality into its own class because I wanted to use the same functionality for both the Propagation and Shift operators.

Pre-Merge Checklists

Submitter

• [x] Write a helpfully descriptive pull request title.
• [x] Organize changes into logically grouped commits with descriptive commit messages.
• [x] Document all new functions.
• [x] Write tests for new functions or explain why they are not needed.
• [x] Build the documentation successfully
• [x] Use yapf to format python code.

Reviewer

• [ ] Actually read all of the code.
• [ ] Run the new code yourself.
• [ ] Write a summary of the changes as you understand them.
• [ ] Thank the submitter.

[pep8speaks]

Hello @carterbox! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

There are currently no PEP 8 issues detected in this Pull Request. Cheers! :beers:

Comment last updated at 2020-06-23 17:26:55 UTC

[carterbox]

@nikitinvv I didn't do a comprehensive study of different plan shapes and sizes, but when I initially implemented the plan cache for the Propagation operator, using the cache was faster than recomputing the plan every time. CuPy will make a plan by default, but it doesn't yet have a built-in plan cache. However, I did not compare the cached-plan implementation with no plan. You can tell CuPy not to use a plan; I believe this means it does a series of 1D transforms. Do you think no plan is worth benchmarking?