English
Read the words on signs, packs and printed sheets
Text that is meant to be looked at behaves differently from text meant to be read. A sign puts four sizes of lettering on one board and expects you to take them in a particular order; a food pack hides the useful part — allergens, a use-by date, a batch code — in four-point type on a curved surface; a newspaper page runs seven columns under a headline that belongs to only one of them. Evidano keeps that hierarchy instead of flattening it into a list of words.
Sample pages
Real pages from public collections, shown beside their transcriptions. Pages whose transcription is still being checked are marked.
Library of Congress, Farm Security Administration collection, December 1935 · Source · Public domain · Photograph by Walker Evans
Transcription being verified
The Library of Congress catalogue record describes the view but does not transcribe the signs; the smaller ones at the far end of the block are below the resolution of the negative.
What the model is told to watch for: Capture stylized capitalization, deliberate line wrapping, prices, units, trademarks, logos containing text, and symbols embedded within product or directional information. Distinguish decorative glyphs from letters and preserve the reading hierarchy created by size, orientation, panels, columns, and foreground or background placement.
US National Archives and Records Administration, between 1906 and 1946 · Source · Public domain
Transcription being verified
The National Archives catalogue gives a scope note rather than a transcription. The label is a useful test of whether every panel is picked up, including the printer’s imprint along the bottom edge.
What the model is told to watch for: Capture stylized capitalization, deliberate line wrapping, prices, units, trademarks, logos containing text, and symbols embedded within product or directional information. Distinguish decorative glyphs from letters and preserve the reading hierarchy created by size, orientation, panels, columns, and foreground or background placement.
Source · CC BY-SA 4.0 · Photograph by Wikimedia Commons user Stefanik
Transcription being verified
Retail packaging is not published anywhere as text; this pack is included because the same panel holds two languages, a date code and several logos with words inside them.
What the model is told to watch for: Capture stylized capitalization, deliberate line wrapping, prices, units, trademarks, logos containing text, and symbols embedded within product or directional information. Distinguish decorative glyphs from letters and preserve the reading hierarchy created by size, orientation, panels, columns, and foreground or background placement.
Source · Public domain · United States Post Office Department, printed in the River News Herald
Transcription being verified
No digitised text of this issue is available. The form is a printed specimen, so no real payee, amount or serial number appears on it.
What the model is told to watch for: Capture stylized capitalization, deliberate line wrapping, prices, units, trademarks, logos containing text, and symbols embedded within product or directional information. Distinguish decorative glyphs from letters and preserve the reading hierarchy created by size, orientation, panels, columns, and foreground or background placement.
What counts as a sign or a label
This covers fascias and hanging signs, street name plates, wayfinding and regulatory signs, notices taped inside a window, menus and chalkboards, posters and handbills, tickets and vouchers, and every kind of packaging from a tin label to a blister pack. It also covers ordinary printed sheets — a newspaper page, a leaflet, a set of instructions — where the interest is the layout as much as the words. What unites them is that the design carries meaning: what is biggest is the name, what is in the red band is the claim, what is in the box at the bottom is the legal small print.
The typography is deliberately awkward. Display lettering is drawn rather than typeset, condensed or expanded to fill a board, set in capitals, arched round a curve, reversed out of a dark panel or given a shadow. Numbers arrive as prices with a currency sign, as weights with units and the estimated-quantity mark ℮, as percentages in a flash, and as date codes that mean nothing outside the factory that printed them. Logos contain words that are part of the brand — and symbols that are not words at all, such as the recycling triangle, the wheelie bin or a country flag. On packaging two languages often sit on the same panel, alternating line by line.
Then the photograph gets in the way. Signs are read at an angle from the pavement, so lettering foreshortens and the far end of a fascia is half the height of the near end; glass and gloss paint throw the sky back at the camera; a hanging sign is cut off by the edge of the frame or by a passing lorry. Packaging curves away, so the ingredients list bends out of focus, and a chiller cabinet adds its own reflections. Street scenes contain a dozen signs at different distances, most of them too small to read, and it matters that a transcript says which ones those are.
Why ordinary OCR struggles here
Size and position are the reading order
A board that reads SUN COFFEE SHOP, then ORIGINAL FRENCH DRIP, then COFFEE in the largest letters, then OPEN DAY AND NIGHT is one message in four bands. Sorted by size, or read strictly top to bottom without grouping, it turns into disconnected phrases that no longer name a business.
Small print is where the facts are
Allergens, storage instructions, net weight, a use-by date and a batch code are set at a fraction of the size of the brand name, often in low contrast over a tint. That is exactly the text people need out of a photograph of a pack, and exactly the text that a downscaled snapshot loses first.
Letters that are not letters, and letters inside pictures
A logo may be a monogram, a device or a word drawn as a picture; a quality mark carries text inside a roundel; a symbol such as ℮ or ♻ sits in the line like a character. Deciding what is a word, what is a mark to be named and what is ornament is a judgement about design, not about shapes.
Angles, glare and everything cut off
Photographs of signage are taken from below and to one side, through glass, in sunlight. Perspective narrows one end of the text, reflections erase a line, and the frame slices words at the edge. A transcript that silently completes a half-visible word is worse than one that marks it as cut off.
Who works with this material
Retail and packaging teams
Artwork proofs, competitor packs and shelf photographs all need their copy in text: claims, weights, allergen statements and the words inside every mark. Getting them out of a photograph is faster than retyping and keeps the wording exactly as printed.
Accessibility and translation projects
Signs and labels are where people get stuck in an unfamiliar place or an unfamiliar language. Reliable text is the raw material for a spoken description or a translation, and it has to include the small print, not just the headline.
Surveyors, inspectors and asset teams
Site photographs carry plant plates, safety notices, test labels and meter readings. Pulling the model number, the serial and the inspection date off a plate photographed at arm’s length in bad light is the whole job.
Historians of the built environment and advertising
Street photographs are evidence of who traded where and what they charged. Transcribing the fascias, painted walls and window bills in a set of photographs turns a picture collection into something searchable by name and by trade.
Getting the best transcription
Frame the whole sign, then go closer
Take one photograph that includes every edge of the board or panel, then a second of the small print from close up and square on. The wide shot fixes the layout and the close shot gives the characters; together they do what neither can alone.
Ask for the hierarchy, not just the words
Request the text grouped by panel or band, with a note of relative size and colour, and with any logo reported as a mark plus the words inside it. That is what lets you rebuild the sign, or compare two packs, instead of holding a bag of phrases.
Name the fields you actually need
On packaging, say that you want the product name, net quantity, use-by date, batch code, allergens and country of origin as separate fields. Those are the parts that get checked, and asking for them by name is what stops them being buried in a paragraph of marketing copy.
Deal with curve, glare and perspective at the camera
Turn the pack so the panel faces you, move a step to the side to lose the reflection, and photograph a fascia from as close to square on as the pavement allows. A minute spent moving beats any amount of processing afterwards.
Frequently asked questions
- Are prices and units kept exactly as printed?
- Yes — the currency sign, the decimal point, the unit and any qualifier such as per kg or ℮ are transcribed as they appear. Nothing is converted between currencies or unit systems unless you ask for it separately.
- What happens to logos and symbols?
- Words inside a logo are transcribed and the mark itself can be named — a recycling triangle, a quality seal, a flag — so that a pack description is complete. Purely decorative devices are described briefly rather than turned into letters.
- Can it read a sign photographed at an angle?
- Usually, within reason. Moderate perspective is handled; an extreme angle, where the far end of the lettering is a few pixels high, is not, and in that case the unreadable part is reported as unreadable rather than guessed at.
- How is a bilingual label handled?
- The two languages can be separated and labelled, so that the Welsh and the English, or the English and the French, come back as parallel fields rather than as alternating lines of one confused paragraph.
- Will the small print on packaging come out?
- It depends entirely on the photograph. Allergen and storage text is often smaller than two millimetres, so it needs a close, well-lit, square-on shot; from a general picture of the shelf it will be marked as present but illegible.
- Can several signs in one street photograph be listed separately?
- They can, each with its own text and a note of where it sits in the frame, which is more useful than a single block. Signs too distant or too foreshortened to read are listed as such rather than being invented from context.
Related scripts and pages
Get the text off a sign or a pack
Upload a photograph of a shopfront, a notice, a label or a printed sheet, choose Signs, Labels and Printed Materials, and name the fields you need pulled out.