ApiaryActiveLive
Try: pause · settings · learn · wipe
← Community / Reading Room
AP
craft · 24 min read

A Photo Glossary Should Remember What the Sign Actually Says

A small exhibit has three versions of a label. One is a photographed sign whose lower corner is scuffed. The second is a typed catalog note. The third is a…

A small exhibit has three versions of a label. One is a photographed sign whose lower corner is scuffed. The second is a typed catalog note. The third is a tidy string returned by an image-to-text assistant. It would be easy to pick the neatest spelling, copy it into a glossary, and forget that the three records aren't interchangeable. The sign was real evidence; the tidy output was a guess about it.

A durable glossary stores what is actually visible, what a reviewer thinks it means, and what people can type to find it. Those are different jobs. The word that works best as a search alias may be the wrong choice for a quotation. The readable translation may not be a valid source for a missing character. A better workflow borrows AI for suggestions and leaves authority in the evidence.

This is a guide to documenting terms and indexing images, not translating workplace instructions, creating a place-name authority, or using a label to identify equipment. All examples are synthetic or built from ordinary illustrative vocabulary. Photograph only material you are permitted to capture and index. A public-facing glossary doesn't need private labels, visitor faces, barcodes, or address details as its source material.

Start with a decision about the collection

Decide whether your source is a set of signs in an exhibit, captions in a historical brochure, labels in a sampling kit display, or some other bounded collection. The source you chose defines what “seen” means. If one photograph shows an object's temporary exhibition label, it is evidence of what the label displayed during that exhibition, not evidence of how the original owner described the object a century earlier.

Write the glossary's purpose in plain language: “Help volunteers find which display label used a particular term, see it as photographed, and locate an approved explanation.” That sentence gives you three useful acceptance tests. Can volunteers find it under a common input? Can they see the line and its surroundings? Does the summary distinguish an editorial interpretation from the recorded string?

Avoid quietly changing scope to “list everything named in the collection.” If an item name can appear on labels in two rooms, search should find both. One occurrence might use a dialectal form or display spelling, and another might use a translation supplied by a later curator. The catalog does not need to rank one as better to connect the result to both sources. If a group later adopts a preferred term, give it a dated authority record, not a silent rewrite of old occurrences.

Preserve the minimum usable image evidence

Start with one permitted image showing the entire relevant object or display. Get close enough for a readable crop while leaving enough room to establish where the crop came from. If a reflective surface defeats legibility, try another allowed angle or lighting condition rather than asking software to fill the glare. Do not physically rearrange, clean, or wet a source unless the collection rules separately authorize it.

The copy used for OCR may need transformation. Record how to reproduce or identify that derivative: original image ID, bounding rectangle or crop reference, rotation, contrast adjustments, and capture or edit date if relevant. A clean-looking high-contrast derivative can suppress faint marks, so it never replaces the image from which it was made. In any review that depends on a tiny character, open the original alongside the derivative.

Maybe a physical label is mounted near visitors or captures private inventory information. A new picture of the bare label at an authorized time can be easier to govern than blurring a crowd scene. If the source is not appropriate to retain or display, don't catalog it by promising to fix its visibility later. Keep file names generic and publish only approved image crops, not original metadata with hidden location or identifying tags attached.

Give every term three separate text fields

A source occurrence is a record tied to a region of an image. It should have a transcribed string or a state explaining why there is no complete transcription. A concept record is editorial: a gloss, translation, scope note, or linked authoritative entry. A search alias is an index entry with a documented reason for matching that concept or occurrence. A normalized string is not automatically any of these.

Here is an imagined exhibition example:

FieldExample
Occurrence IDLABEL-017-02
Image and regionIMG-017, second printed line
Language at sourceFrench, checked by reviewer
Source textÉtiquette
Editorial glossLabel (noun), for the display glossary
Search aliasesetiquette, label
Review statusVerified as visible; gloss reviewed separately
NotesCase and accent preserved in source field

The unaccented alias might be convenient for someone using a keyboard that does not readily produce É. It may also match an unrelated word in another language. Present it as a search route, not an equivalent transcription or guaranteed translation. If several concepts match, show the alternatives and enough context for the visitor to choose. Suppressing ambiguity at lookup often creates the false impression that the glossary endorsed a single meaning.

An incomplete occurrence should not be filled from the glossary's usual spelling. A training record could show “étiq[unreadable]” with the bracket annotation stored as a transcription convention or a separate uncertainty span. A downstream display could read “partial transcription” rather than visually pretending the bracketed words were on the sign. If a second authorized photograph reveals the missing characters, attach that evidence and replace only the uncertain segment with the newly observed text. Keep a revision note detailing what changed.

Make OCR compete with another reading

OCR can save tedious keystrokes and can prioritize candidate regions for human review. It isn't a second witness if the reviewer reads only its output. When a term will anchor a published entry, have someone record what they can see before exposing the model's candidate. A truly independent reading helps catch a neat suggestion that would anchor the reviewer on an invented accent or convenient spelling.

Differences to pay attention to include accents, letters that resemble other scripts, punctuation that changes segmentation, leading or trailing glyphs, text copied from adjacent columns, and line breaks that combine two labels into one. Some tools present an estimated confidence. That value can help order a queue for a particular tool configuration, but it is not proof of correctness and doesn't compare uniformly across scripts or photo conditions. An obviously wrong string with a high score stays wrong.

The Tesseract guide to improving output quality discusses image quality and page segmentation choices that affect recognition. Its engine-specific examples help explain why cropping a region or changing orientation can affect a result. Do not infer from one successful preprocessing recipe that every form of noise is removable without throwing away evidence. Keep the source view available and evaluate your own image conditions and language coverage.

Imagine a blurry synthetic sign fragment with a likely lower-case é. The engine prints “e” with confidence. A reviewer sees a short mark but can't distinguish a diacritic from a stray impression. The appropriate state is unresolved at that character, not “accent added because the dictionary says so.” If the gloss identifies a known term and you have an authoritative catalog page, cite that in an editorial note. It supports the gloss, not a claim about what pixels were visible.

Keep canonical encoding separate from forgiving search

Two visually equivalent strings can be encoded differently. An accented letter may appear as a single code point or as a letter followed by a combining mark. Exact comparisons on bytes or code points can therefore miss obvious matches. Unicode Standard Annex #15 specifies normalization forms and distinguishes canonical from compatibility equivalence. A comparison or index can use a chosen normalization strategy without replacing the stored evidence.

Decide explicitly which behavior is intended for each field. A source transcription may be preserved as captured or stored under a documented canonical encoding with the original image retained. Your search key might normalize to NFC for canonically equivalent matching and also index individually approved, accent-insensitive aliases. Accent removal is not the same as NFC. Compatibility normalization can collapse distinctions that are meaningful for a label, so don't silently apply it to the displayed source string.

Run a few checks using invented examples. Paste a visually identical composed and decomposed accented word, query both, and verify that they return the same occurrence under canonical comparison. Inspect what gets displayed. Then test an accent-free spelling that could have two meanings and verify that the interface offers both if they are approved aliases. For systems with automatic case handling, try language-specific cases relevant to the collection rather than assuming the same case operation fits every language. Keep the reason for each search expansion auditable even if it is set by a collection-wide rule.

As collections grow, organize your alias records so “spelling on a photographed label” cannot be mistaken for “a curated translation” on export. Otherwise a search optimizer's convenience substitution can turn into a supposed historical fact after someone reuses your dataset.

Set language where it can help readers

One label may contain two languages; a brand name may not tell you which language surrounds it. Record language or script only to the quality you can support. If the label uses a script you recognize but the language cannot be distinguished responsibly, keep the language unknown pending review. Don't fill an uncertain tag by assuming neighboring labels were cast on the same template.

For a web glossary, W3C's guide to declaring language in HTML explains how to declare a page's default language and mark a span in a different language. The application here is straightforward: mark the source term's language on the term when known, and let the surrounding gloss retain the language of the page. Do not place the language qualifier only in a search index where it cannot help the presentation layer.

If you display right-to-left and left-to-right strings in one line, test the layout of punctuated terms in the actual destination, including text copied into search. The rendered order and the underlying code-point sequence are not something you should infer solely from a screenshot. Include representative native-reader feedback when your collection requires conventions you don't personally know. A result page that visually mangles a correct term is still a broken glossary for its intended readers.

Count disagreement in the terms you can fix

A straightforward review register uses states such as awaiting first read, candidate generated, discrepancy, verified, partial, and not for release. It doesn't need to award points for how many labels were “done.” Record whether the unresolved problem is reading the source, translating the term, deciding language, or choosing an alias. These issues have different fixes and different owners.

A second photo may resolve legibility while leaving translation unchanged. A dictionary may validate a gloss while leaving the scuff-mark ambiguity intact. A collection manager may decide that an image can't be shown to the public even after the transcription becomes readable. Keeping an individual state for each layer prevents one good answer from turning the entire record green.

Picture a review day with eleven fictional occurrences: seven readable on first pass, two unreadable without retakes, one readable but needing translation, and one photograph with a recognizable visitor who cannot be shown in the planned crop. Your next action is not “ask the model to try harder on four rows.” It's obtain two better source images, queue a language expert once, and replace one photograph through an appropriate process. Report the seven that cleared all mandatory gates and identify each blocker without putting any protected image into the report.

Write down what a correction may change

A practical correction form asks for the occurrence ID, the disputed text field, what the reviewer sees, and the new evidence. An instruction like “change this to the proper spelling” is not specific enough. It could be a correction to the photographed transcription, a change to the preferred editorial heading, or a new match term. Preserve the old record's meaning until you know which change is being proposed.

For example, an exhibit volunteer might report that a new photograph reveals an accent in a label's second line. Compare the two approved captures and correct the source entry only if the mark is now identifiable. Keep the editorial gloss unless its meaning changed. A different volunteer might ask to index a plural spelling because local visitors use it at the search kiosk. That is an alias request; it should not cause any changes to the old photograph's transcription.

A correction sometimes exposes that two sources were merged into one entry by mistake. In that case, don't change the single source field and throw out the conflicting evidence. Split the occurrences, give each its own region and review state, and let them point to the same concept only if the connection has been assessed. Keep the earlier merged ID as a redirect or archive pointer if it appears in existing references. That way, fixing the glossary doesn't quietly break the route readers used to see it last week.

Avoid requiring volunteers to upload a new photo when an already approved crop supports the review. An image reference plus a precise region is often easier for a curator to verify than a copy with a different size and no context. Write the accepted result back to the appropriate layer and retest the public lookup.

Set a boundary for the unit you transcribe

One photograph can contain three signboards, two languages, a painted border and a reflection. The item you preserve isn't necessarily the entire file treated as one text record. It could be one labeled panel, one block of a multilingual sign, or a term that recurs elsewhere. Define the unit in a sentence a second editor can apply. A glossary record might keep one term as displayed in one specified region, linked to the full sign and several neighboring regions for context. A complete sign transcription is another task. Combining them in a single text entry can obscure how the order, spelling and source were decided.

Consider a fictional ferry exhibit with a category name on a printed label, a translation on a sticker below it and a chalk arrow with directions next to both. The photograph holds all three. A glossary editor working on the printed label should draw the target boundary around it, preserve the sticker and chalk arrow as context, and say why they are not tokens of the term. A subsequent collection record may index the sticker as its own displayed text. Neither OCR nor a convenient rectangle around all recognizable letters should decide how many items appear in the record. The visual evidence tells you where multiple elements are; the collection rule tells you which ones count for this use.

Record line order when it matters. Some signs place two columns side by side, with matching terms in each column; others combine a title across both with a list below. Reading across in an automated extraction may pair elements that don't belong together. Preserve a panel or region identifier such as upper title, left block, row one, rather than relying on text position in a flattened dump. If the reading order is ambiguous in a particular view, say that and let a specialist resolve it. A structured field can support the uncertainty rather than forcing the nearest plausible reading into a definitive order.

Keep an explicit field for relations you did not verify. Two forms on neighboring panels may be translations, transliterations, separate terms or different names for different objects. A reader might be able to verify spelling in each independently without determining the link between them. Treat that link as a separate editorial claim. A tidy bilingual interface should not make an accidental association look like museum interpretation. If a public layout presents the signs alongside an explanatory translation, label which is on the source and which has been supplied later.

Keep editorial choices separate from questions of how to encode them

A readable transcription and its exact storage representation pose different problems. A typographic inspector might notice an accent looks hard to separate from background noise. A database maintainer may instead be looking at characters that display identically under a common font despite having different underlying sequences. The tempting solution is to say one is the “real” problem and dismiss the other as detail. They require distinct decisions. Record which visual mark is supported by the photo and which digital normal form is used for consistency. If one changes, the other need not.

The Unicode normalization specification is a technical source for understanding equivalence and different normalization processes. It doesn't tell you whether a photographed sign intended a mark that isn't visible in a blurred image. For a synthetic example, a term in two electronic indexes might look like café on screen. A comparison of canonically normalized forms can help retrieve those versions consistently, if that's appropriate to the project. It should never be used to infer that a real photograph definitely shows the accent, or to require a transliteration that is better classified as a search alias. Store the visual interpretation and note any evidence limitations separately from the storage policy.

Compatibility-oriented transformations can remove distinctions some projects want to preserve. A researcher looking for a historical type form or a layout convention may care about precisely what the display showed, while a visitor search can be forgiving. Don't apply compatibility changes to the authoritative text because search performed better after them. Instead use a deliberately chosen derived form and test it. Document what the process changes, including whether case conversion, punctuation reduction, or space folding are deliberate operations in that collection. A convincing display can be legible and accessible while keeping attribution for the original form intact.

A small test pack makes this concrete. Choose one unambiguous accented spelling, a near lookalike that is a different term for local readers, a faded patch where the visual record does not resolve either, and a reasonable unaccented search alias. Verify that none of the stored source fields changes just because you ran the search transformations. Verify that the alias finds the intended public entry if that is the policy, and that potential collisions are handled visibly rather than redirecting to whichever record was imported first. The unresolved photo should stay unresolved even if a popular alias would make one reading more convenient to index. This is a measurement of your system's behavior, not of the photographed term itself.

Compare recognizer output without making the output a vote

Readers often think confidence values make a decision comparable even across tools or runs. An OCR result, especially a persuasive one from a clear synthetic test image, doesn't automatically establish a universal error probability. Different outputs may share the same wrong assumption because they consume the same low-quality crop or depend on similar training examples. The Tesseract documentation on image quality helps explain why preprocessing and segmentation choices can affect output. It is a prompt to preserve how the output was produced rather than a guarantee that an alternate preprocessing choice will recover missing print.

Keep a table with each run's output, input version, settings that materially affected what the tool saw, and the corresponding image evidence slot. If the first run ignored a lower block because it read the image as one line, another output might help you inspect that particular region. Two runs both on the same tight crop don't independently prove the absent lower block wasn't there. Likewise a better-looking word in one result may reflect a useful language setting, or a dictionary making a reasonable guess. Ask which visible strokes support the result, not whether the algorithm provides a smoother sentence after your change.

Suppose a fictional exhibit uses a short word with a vertical stroke near the edge. A wide photo shows a thin upright reflection crossing some letters. An OCR run includes a different early letter from the one recommended by a second run on a cropped version. If the reflection still hides the distinguishing marks from both, no vote should settle the original. One note can say “two plausible readings, distinction obscured in this frame.” If a second verified image from another angle makes the mark clear, use it under its own source identification rather than editing the earlier photo or entering a winner based on confidence alone. Both files can continue to support an interesting account of the project's decision.

A useful quality count also separates regions from tokens. An extraction may produce a correct term on a large, clear title and miss a small second column entirely, making its list of emitted words look almost perfect when evaluated against itself. Compare the output against a prelisted set of expected readable regions in the test pack. Give omitted regions a known status, and distinguish regions that humans could not read from ones OCR failed to find. A results table with counts for regions located, exact source forms verified, unresolved regions, and alias-only differences offers more actionable feedback than an alluring single accuracy number.

Run a publishing preview in the right reading order

An editor's worksheet can show one string per row without revealing what the published page will do with mixed directions, punctuation or focus navigation. Open the local or otherwise authorized preview and test representative entries with the layout that readers actually receive. Observe whether the displayed labels and their annotation remain distinguishable, whether each region comes in the intended order, and whether a footnote lands next to the right block. Such tests can be done with synthetic entries and do not require sharing an unpublished sign photograph with another service. A browser's rendering can help locate the bug but can't verify what was intended in an unseen source.

The W3C guidance on declaring language in HTML gives useful distinctions for language information in ordinary documents. It doesn't resolve every issue of writing direction, fonts or how content should be transcribed. Your preview should treat those as separate checks rather than assuming that one attribute corrects them all. Language annotations should match the relevant spans when verified and useful. Don't assign a further language based solely on recognition of a script shared by multiple languages. It is possible to get the correct script and still mislabel the language, or to have the right language context with incorrectly ordered columns.

Give the preview a realistically ambiguous fixture. A fictional label holds a short heading, an adjacent date with punctuation, one line in a different language, and an italic editorial comment beneath it. Test whether readers can use keyboard navigation to reach the source image and the description in an understandable sequence, whether the date and comment appear in the right grouping, and whether entry names are legible in the narrow layout. A search result snippet should show the actual field responsible for the match when the discrepancy matters and link to the checked entry. If the link is correctly usable but the snippet promotes an alias as the source spelling, treat that as an editorial bug rather than letting technical navigation success obscure it.

One more test belongs outside the polished preview. Open a plain-text or accessible alternate form. Can someone recover which text is a literal transcription, which is a translation, and which is an uncertain suggestion without relying on color or visual placement? Add terse labels wherever a format change would otherwise collapse the distinctions. A published glossary can accept simple presentation tradeoffs while keeping text evidence inspectable. The smallest useful record is one in which the visual source and its interpretation retain their roles after exports and previews, not just while the project editor is explaining the interface aloud.

Preserve a correction path for source revisions and later witness images

Signs can change in the world while the glossary remains stable. A return visit may find a new panel in precisely the same place with different wording. Do not overwrite the earlier entry as though a new spelling in that location proves what the old image contained. Create a second appearance event and keep the dates attached to both versions. If the new image instead resolves a detail on the same documented panel, indicate why it is appropriate to use it as corroboration rather than treating all similarly framed images as a view of one immutable object. Record changes without pretending the object has no physical history.

An ordinary correction entry names the old reading or status, the reason it was recorded that way, the new result, and the evidence or method supporting it. A separate alias update might show that a reader can now find the item with a common alternate spelling; it needn't change displayed transcription. A presentation correction may update a language annotation without affecting either search or textual evidence. Make each kind of edit identifiable. The process lets a reviewer reproduce what the evidence currently supports without being asked to check the entire collection again whenever a small part changes.

For a synthetic handoff, imagine an annual exhibit's label was reshot more clearly. In the new image the right side includes a diacritic the old photo doesn't resolve; the publisher wants to “fix” the old entry. The right decision may depend on whether the new label is the same object or a reprint, whether its relevant surface changed, and which kind of record the editor intended to preserve. If that cannot be established, show the two observations with different resolution descriptions. The singular term page may treat one spelling as likely for the collection without recasting the actual earlier image as clear. A display that says exactly where confidence improved is more useful than a history stripped of its uncertainty.

Take care that review copies do not leak unrelated photo material. Use narrowly scoped crops or synthetic fixtures where practical, preserve source provenance and follow the collection's access and reuse rules. An outsider may require enough context to understand the term without receiving the raw surroundings or personal information incidental to the picture. Keep internal provenance pointers separate from public evidence pointers and test both according to their audience. The purpose of preservation is accountability, not an automatic duty to expose everything inside the photographer's first frame.

Finish by asking an editor to answer three questions from the record alone. What could the photographer see at the specific panel? What did the editorial team decide to present, with what uncertainty? What other forms might readers reasonably search for? If these answers disagree, the disagreement may be the reason to keep the record, not a glitch to clean up. A multilingual glossary is especially good when it answers all three clearly enough to improve next year's reading rather than repeating the original guess in a more professional-looking font.

Include a test query whose best result is a choice

Create a completely fictional exhibit with two recorded spellings and a third visually similar but unrelated term. An unaccented search might reasonably return two candidates rather than sending the reader to whichever entry an import created first. Define acceptable behavior before implementing the test: for instance, exact source text takes precedence, the other form remains findable, and any relevant language filter applies only when the language is actually established. The result should make it possible to understand which field matched, rather than implying all variants are attestations from the same sign.

Now include a query with no supported answer. Use a plausible-looking sequence generated by a recognizer for an intentionally blurred sample, but do not mark the sequence accepted in the index. A successful result may show no match and retain a route back to the browse list; there is no obligation to present an unsupported guess as a source term to avoid a blank page. If one alternate spelling matches a distinct term, check that result selection remains under the reader's control. Good alias handling entails finding a reasonable item without substituting it for a different claim.

Consider a second variant where a reader proposes an alias after a scheduled exhibit changes. Their report may be accurate about how people search today while giving no information about the meaning or spelling in the older label's image. Record their input as a proposed findability improvement. If useful and allowed under the collection's policy, add it to the appropriate alias review and test the result. Do not add it to a verified witness field. This separation is possible even in a spreadsheet with no search engine beyond an ordinary filter: use explicit columns and explain what they mean.

Read the test result aloud as if you were providing support to someone unfamiliar with the project. Could they name the observed form, the entry they landed on, and the field that matched the original query? If they cannot tell an editorial hint from photographed text, the fixture has identified a defect even when the links are technically fine. Record the condition and test data with the result, without collecting unnecessary visitor information. Next year's editor should be able to add new cases without needing to discover what rules were in force during the first demo.

A small pass sheet can capture query class, expected entries, displayed explanation and whether navigation led back to the correct documented source. Keep one case in which two reasonable answers are shown rather than hidden behind a winner. Readers can handle a choice. They are worse off when the search interface resolves a question the evidence still leaves open.

Test the public view as intentionally as the data

Open a result as a visitor. Can they tell whether the term is an exact occurrence or an alias match? Can they get to an approved source region and enough context to see how the source was used? Is an uncertain label visibly different from a verified one without requiring the visitor to distinguish two colors? Can they understand why two source spellings appear side by side?

A label that looks beautiful but requires opening an image just to read the text fails a different requirement from a database that strips diacritics. Offer selectable transcription and a meaningful reference to the image. Provide appropriate alternative text for images, based on their role and surrounding text. Avoid duplicating every string in a way that makes someone using assistive technology hear the same text three times. For a diagram or label with meaningful layout, a short note can explain what is next to what.

When you export, check one representative entry in the actual file a collaborator receives. A private image path may not open for the collaborator. A missing language or review column may make a rough gloss look like a verified translation. Search indexing, authoring, and distribution often have different permissions; choose a public export that contains only what the intended audience can see. Preserve the restricted original separately when your organization allows it.

Start with five candidate occurrences, not a thousand automated records. Include one clean label, one partial label, one term whose convenient alias is ambiguous, one label with text in different languages, and one source that shouldn't be released. Try to make a reader fail at least once and use the failure to improve the record shape. The win is a searchable term that leads back to real evidence, with no need to pretend that an assistant, a dictionary, and a photograph all saw the same thing.

AI disclosure: This draft was produced with AI assistance; examples are illustrative, not records of actual work.

Frequently asked
What is A Photo Glossary Should Remember What the Sign Actually Says about?
A small exhibit has three versions of a label. One is a photographed sign whose lower corner is scuffed. The second is a typed catalog note. The third is a…
What should you know about start with a decision about the collection?
Decide whether your source is a set of signs in an exhibit, captions in a historical brochure, labels in a sampling kit display, or some other bounded collection. The source you chose defines what “seen” means. If one photograph shows an object's temporary exhibition label, it is evidence of what the label displayed…
What should you know about preserve the minimum usable image evidence?
Start with one permitted image showing the entire relevant object or display. Get close enough for a readable crop while leaving enough room to establish where the crop came from. If a reflective surface defeats legibility, try another allowed angle or lighting condition rather than asking software to fill the glare.…
What should you know about give every term three separate text fields?
A source occurrence is a record tied to a region of an image. It should have a transcribed string or a state explaining why there is no complete transcription. A concept record is editorial: a gloss, translation, scope note, or linked authoritative entry. A search alias is an index entry with a documented reason for…
What should you know about make OCR compete with another reading?
OCR can save tedious keystrokes and can prioritize candidate regions for human review. It isn't a second witness if the reviewer reads only its output. When a term will anchor a published entry, have someone record what they can see before exposing the model's candidate. A truly independent reading helps catch a neat…
References & sources
  1. Apiary Reading Room — Open, cited knowledge base — funded to keep bee & practical research free.
From the Apiary Reading Room. Opinion & editorial — not financial advice. We don't overclaim.
More from the Reading Room