Add searchable text to clear English handprint on a scan

Printed text recognition is close to solved. Handwriting is not, and the gap is not a matter of engine quality — it is a property of the input. Knowing which handwriting is realistically recognisable saves a lot of disappointment.

Why handwriting is a harder problem

Printed characters come from a font. Every lowercase a on a page is identical to every other one, which gives a classifier a stable target.

Handwriting has none of that stability:

• Every instance differs. Your own letters vary between the start and end of a page, let alone between people.
• Characters connect. In cursive, there is no clean boundary between letters, so the segmentation step that printed recognition relies on has nothing to segment.
• Baselines drift. Unruled paper produces lines that wander and slope.
• Spacing is ambiguous. The gap between words and the gap between letters are frequently the same size.
• Shapes overlap. A hurried u and v, a and o, r and n, 1 and 7, and 0 and O are genuinely ambiguous even to a human reader without context.

That is why the handprint recognition tool is scoped to clear, separated English handprint. It is a beta capability with an honest boundary rather than a promise to read anything.

What works and what does not

Input	Realistic outcome
Block capitals in form boxes	Good; this is the best case
Neat separated printing on ruled lines	Usable with review
Mixed printing and cursive notes	Patchy; expect to correct a lot
Fast joined-up cursive	Poor
Shorthand, doctor's notes, marginalia	Not a realistic target
Non-English scripts	Out of scope for this tool

The determining factor is separation. If the letters do not touch each other, recognition has something to work with. If they flow together, it does not.

Prepare the page first

The same preparation that helps printed recognition helps more here, because the signal is weaker to begin with:

1. Capture at good resolution. Aim for the equivalent of 300 dpi. Pen strokes are thin and low resolution turns them into ambiguous grey.
2. Correct the perspective. Use the document scanning tool if the source is a phone photo, so lines are straight before recognition tries to follow them.
3. Increase contrast. Pencil and light blue ink are much harder than black ballpoint; pushing contrast can rescue a faint page.
4. Remove the background. Ruled lines, grid squares, and coloured form backgrounds all compete with the writing.
5. Crop to the writing. Excluding printed form labels stops the recogniser from mixing printed and handwritten content.

The original page is preserved

An important design point: recognition adds a text layer behind the existing page image rather than replacing the page with the recognised text.

That means a misread never destroys anything. The scan still shows exactly what was written, your eyes remain the authority, and the recognised text is an index laid underneath. If the recognition is 80 percent accurate, you have a document that is 80 percent searchable and 100 percent readable — strictly better than the image you started with.

This is why the sensible framing is searchable archive, not transcript. Finding which of 400 scanned pages mentions a surname is enormously valuable even if the recognised text would be embarrassing to publish.

Review before you rely on it

Run the text extraction tool over the finished file to read the recognised layer as plain text. Errors that hide under the image become obvious immediately.

Pay particular attention to:

• Digits. Amounts, dates, phone numbers, and reference codes are where errors cost the most and where handwriting is most ambiguous.
• Names. Proper nouns have no dictionary support, so the recogniser cannot fall back on language context.
• Negations and single-letter answers. A misread Y or N in a form field inverts the meaning of the record.
• Anything you plan to act on. Read it against the image before using it.

Used within those limits — clear handprint, good capture, preserved original, reviewed output — handwriting recognition turns a box of scanned notes from an unsearchable pile into something you can actually query. Used outside them, it produces confident nonsense, which is exactly why the review step is not optional.