phase-recovery is one README and a licence file

The repository is one Markdown file and a licence

The top level of this repository has two entries. A licence file and a readme.

There is no code, no test directory, no continuous integration configuration, no issue template and no contributing file. The metadata records no primary language, which is consistent with a repository whose entire content is prose. There are no releases, so there is nothing to install and no tag to pin.

What the readme is, is a linked bibliography. Its first substantive line says the resources were released from a digital object identifier, and that identifier is registered as the repository's homepage. So the list was published alongside a journal article, and the repository is the long-lived copy of what would otherwise be a supplementary document.

That also explains the shape. The page defines its subject first, in one sentence, as calculating the phase of a light field from its amplitude or intensity measurements, and immediately notes that the term covers several distinct bodies of work, naming holography and interferometry, the transport of intensity equation, optimisation-based phase retrieval, wavefront sensing and deep learning approaches. Everything after that sentence is a filing system for papers and people into those five families.

The first line is a commented link definition holding an email address

The very first line of the file is not prose. It is a Markdown link reference definition, which is a syntax that produces no visible output, with its target replaced by an empty bracket and the whole line hidden behind a comment marker.

Used as a carrier for a comment, it holds a person's name and an email address. That is a common convention in a file like this because the readme is a single flat document with nowhere else to record who maintains it, but it does mean the address is in the plain text of the file rather than in a contact page or in commit metadata.

The same document also carries a contributors graphic and a back-to-top link, both of which are decorative additions to what is otherwise a table of contents followed by a list. The graphic is a link to the repository's own contributor graph, which is the standard badge and nothing more.

None of this is a fault. It is worth knowing because the file is long enough that a reader looking for the maintainer's contact will find it on line one, formatted as though it were broken syntax.

The deep learning taxonomy has sixteen leaves under three stages

The largest part of the classification is the deep learning side, and it is organised by where the network sits in the reconstruction pipeline rather than by architecture.

Before processing has four subcategories: pixel super-resolution, noise reduction, hologram generation and autofocusing.

Inside processing has five, and these are given as named strategies with acronyms. Two are network-only, one driven by the dataset and one driven by physics. The other three connect the two in different directions: physics connected to network, network inside physics, and physics inside network. That is a useful distinction rather than a cosmetic one, because the three place the training objective in different places relative to the forward model, and it is the distinction most papers in this area are actually arguing about.

After processing has four more: noise reduction again, resolution enhancement, aberration correction and phase unwrapping. Note that noise reduction appears on both sides of the pipeline, which is accurate rather than a duplicate, since reducing noise before reconstruction and cleaning the result afterwards are different problems.

A fourth branch is not a pipeline stage at all but a task list: segmentation, classification and imaging modality transformation.

So the deep learning side is sixteen leaf categories under four headings, plus the conventional side, and the whole scheme is deep enough that finding a paper means first deciding which of twenty leaves it belongs to.

The conventional side splits optimisation by geometry and then by convexity

The non-deep-learning half of the list is organised more conventionally, with three measurement families first.

Holography and interferometry is one. The transport of intensity equation, which infers phase from intensity derivatives rather than from interference, is the second. Wavefront sensing is the third.

Optimisation-based retrieval is the fourth, and it is the only one of the four that is subdivided. Under it come six entries. One is alternating projection, the classic iterative method. Three are variants of it distinguished by which axis or direction the intensity information is gathered in, described as axial, radial and angular multi-intensity. The last two are non-convex and convex optimisation.

So the optimisation branch splits first by technique family and then by whether the objective is convex, with the multi-intensity variants sitting inside the alternating projection entry rather than beside it. A reader who has never heard of the multi-intensity variants would not guess from the outline that they are a sub-branch rather than siblings of alternating projection, because the outline shows them at the same indentation level under the optimisation heading with only the AP acronym marking the parent.

Above all of this sit the non-paper sections: research groups by region, companies, workshops and courses, review and tutorial papers split the same two ways as the main list, plus books and dissertations.

The format rule is broken inside the section it governs

The contribution guidelines are short and reasonable: edit the raw file in Markdown, avoid typos, do not add what is already there, make new entries match the format of existing ones, and note the order new additions go in.

The groups section is where that last rule applies, and it is where the file is least consistent with itself.

Each entry is a person, in bold, then an affiliation in parentheses, then a line of keywords. The keyword label alternates between two spellings across entries in the same list, sometimes with the two words joined and sometimes spaced. Affiliations vary in capitalisation for the same institution and for different institutions in the same country. Names carry a trailing space before the opening parenthesis in every entry, which is at least consistent.

One entry lists two people jointly with a single link and a single affiliation, where every other entry names one person. And the section header instructs the reader to search with the keyboard rather than scan, giving fifteen keywords to search for, which is an admission that the alphabetical-by-surname ordering is not enough to find a field.

The listing that was visible ends mid-line in the middle of a keyword phrase, so the file continues beyond what any reader can see at once.

Fifteen keywords stand in for a search index

The groups section opens with a search instruction and a keyword list, and the keyword list is doing the job a structured index would do in software.

The fifteen terms span the field: the core phrases, the measurement families, the related instrumentation terms, and two application areas. They mix the names of techniques with the names of subfields, and they include at least one term that appears in the search list but not in the paper classification, which suggests the two lists were written at different times.

The same section points outward in two places. It notes that more groups exist in a companion repository covering computational imaging more broadly, which is the honest way to handle a list that cannot be complete, and it links contributor submissions through the pull request route rather than a form.

The regional structure is four headings. Asia is the one with entries in the visible portion, and its size relative to the other three is not something the outline reveals, because the outline lists all four headings at the same level with no counts.

So the practical way to use this part of the file is to open it, press the search key, and type a technique. The alphabetical ordering by surname is for the reader browsing, and the keyword list is for the reader who knows what they want.

The contribution rule leaves the sort order open

The last contributing guideline asks you to note the order of new additions, and offers two choices without saying which to use.

Elsewhere in the same file the answer is given, but only for one section: the research groups are explicitly ordered alphabetically by surname. Papers, companies, workshops, reviews, books and dissertations have no stated order, so a contributor following the guideline literally can pick either.

That matters more than it sounds in a flat file this size. There is no per-section index, no table of contents entry with a count, and no script. A list that is half chronological and half alphabetical with no marker distinguishing the two is one that becomes unreadable at a few hundred entries rather than at a few thousand.

The other guideline that would need tooling is the one about not duplicating what is already there. With a single flat Markdown file and a keyword search as the only lookup, the check is manual, and it is the check most likely to be skipped.

The repository's own state suggests the guidelines are read at least sometimes. The default branch last moved on 2026-06-11, and there are no releases to compare against, so the only way to see what changed is to read the commit log.

Editorial conclusion

phase-recovery is worth bookmarking if you are new to the field, because the taxonomy is the useful artefact here and the distinction it draws, between where a network sits relative to the physics, is the kind of organising idea that saves a graduate student a year of reading. Two things to know before you rely on it. It is a list, not a maintained library, so nothing is versioned, nothing is tested and the two years before the last commit are not documented. And the contribution rule asks you to match the existing format while the existing format is visibly inconsistent, so a careful new entry will still make the file slightly larger than it needs to be.

Frequently asked questions

Official sources

1. Issues
2. kqwang/phase-recovery on GitHub
3. License: MIT
4. Project website
5. README