Signature verification for any signed document

Find the signature.
Lift it out.
Compare it with your records.

Chequedb checks whether a document is signed, lifts the signature out of the image, and compares it with the samples you already hold — cheques, forms, contracts, delivery notes. When it isn’t sure, a person decides, with the document, the sample and the reason in front of them.

  • Spot blank, missing and covered signatures before anything else.
  • Works on cheques, forms and other scanned documents — from a scanner or a phone.
  • Keeps the result, the sample version and the reviewer’s decision together.

Every document shown on this page is a synthetic illustration. No customer images are used.

How a signed document gets checked
Animated walkthrough of a signature being scanned, located, lifted out, compared with a sample on file, and recordedA synthetic cheque stands in for a signed document through five steps. The image is scanned, the signature area is found and outlined, the signature is lifted into a card, it is set beside an authorised sample, and a result badge confirms the comparison and record.EXAMPLE BANKSYNTHETIC CHEQUE · NOT NEGOTIABLECHEQUE NO. 000123DATEPAY TOExample recipientIllustration onlySPECIMENMEMOAUTHORISED SIGNATURESignature foundFROM THIS PAGEON FILEResult, sample version and reviewer decision kept as one record

A scan or a phone photo of the document is checked for sharpness and glare before anything else happens.

A drawn example on a synthetic cheque. The same five steps apply to any signed document — a form, a contract or a delivery note. No live model runs here and no similarity score is shown.

How it works

Detection, localisation and comparison, in plain words.

Five steps turn an image of a signed document into a decision you can explain later. The technical names for those steps live in the technical reference; the rest of this page uses ordinary language.

  1. 1

    Take the picture

    A scan or a phone photo. The image is checked for sharpness, glare and framing first.

  2. 2

    Find the signature

    A quick check of whether the document is signed at all, and roughly where the signature sits.

  3. 3

    Lift it out

    The signature is cut out of the image at full quality, with a note of exactly where it came from.

  4. 4

    Compare

    It is set beside the samples you already have on file for that account, customer or signer.

  5. 5

    Record the decision

    The result, the sample version and the reviewer’s reason are stored together.

Where it applies

Cheques, forms, contracts — anything with a signature.

Signature detection, localisation and comparison are not cheque-specific. If a document carries a handwritten signature and you hold authorised samples for it, the same five steps apply. Cheques are simply the case we know best.

Cheques and deposit slips

The drawer’s signature at the foot of the cheque, plus endorsements on the back where you need them.

Account-opening and loan forms

Applicant, co-applicant and guarantor signature blocks, often spread across several pages.

Contracts and agreements

Signature blocks on the final page, and initials in the margins of amended clauses.

Delivery and goods-received notes

Proof-of-delivery signatures that need checking against a known signatory before payment.

Insurance claim and settlement forms

Claimant and assessor signatures on documents that arrive by post, email or as phone photos.

Mandates and payroll forms

Standing-order mandates, letters of authority and payroll change forms with an authorised signatory.

What changes from one document type to the next

Where the signature is expected. A cheque has a defined signature block; a contract may have one at the end and initials in the margins. Expected locations are configured per document type.

How many signers. Some forms need one signature, some need two, and joint accounts need rules for both. The count is part of the configuration, not an afterthought.

Which samples apply. Authorised samples belong to a person or an account, not to a document. The same reference set can serve every document that person signs.

What the document’s own rules require. Cheque law, contract execution requirements and your internal policy all differ. The check supports those rules; it does not replace them.

Step 2 · Detection

Is it signed — and can we read it?

Before any comparison, two simpler questions come first: is there a signature, and is the image clear enough to use? There are three possible answers, and each one leads somewhere different.

Signed

A signature is there and looks usable. It gets lifted out of the image and compared with the samples you hold.

Next: compare it with the samples on file

Nothing there

The signature space is empty. Nothing gets scored — the document is marked as unsigned and sent to a person.

Next: flag it as unsigned and route it

Hard to read

A stamp, glare, blur or a clipped edge is in the way. A signature may well be there, but we say so instead of guessing.

Next: ask for a clearer image, or send it to review

Example: one cheque, three outcomesSynthetic illustrations
Synthetic cheque: signature thereA fictional cheque with a highlighted signature field at the lower right. The box encloses a drawn signature. EXAMPLE BANKSYNTHETIC CHEQUE · NOT NEGOTIABLECHEQUE NO. 000123DATEPAY TOExample recipientIllustration onlySPECIMENMEMOAUTHORISED SIGNATURE

There is a signature here

The outlined area contains a signature-like mark. It can be lifted out of the image and compared with the samples on file.

Next step: check the image is sharp enough, then compare

Drawn examples explain the workflow; they are not live model predictions or performance measurements.

Finding a signature does not tell you who signed it. Printed names, stamps, initials and handwritten notes can all complicate the picture, so these cases belong in your own evaluation set — whatever document they appear on.

Step 3 · Localisation

Getting the signature
out cleanly.

Cutting the signature out of the document makes it easier to compare, and easier for a person to check. Engineers call this signature localisation. What matters in practice is that the crop is clean, correctly placed, and still tied to the page it came from.

Keep it tied to the document. A crop is easier to compare, but a reviewer needs the whole page to see nearby stamps, print and layout.

Don’t clip the flourish. A crop that cuts off the end of a stroke invents a difference that isn’t really there.

Expect more than one box. Joint accounts and some layouts carry several signature areas. Each expected area is handled separately.

See how signature regions fit document data extraction
01 / LOCATE THE REGION640 × 320 px
Synthetic cheque: signature thereA fictional cheque with a highlighted signature field at the lower right. The box encloses a drawn signature. Example box in a 640 by 320 image: x 365, y 174, width 222, height 90 pixels.EXAMPLE BANKSYNTHETIC CHEQUE · NOT NEGOTIABLECHEQUE NO. 000123DATEPAY TOExample recipientIllustration onlySPECIMENMEMOAUTHORISED SIGNATUREx:365 y:174
Illustrative source-image crop at the displayed signature coordinatesEXAMPLE BANKSYNTHETIC CHEQUE · NOT NEGOTIABLECHEQUE NO. 000123DATEPAY TOExample recipientIllustration onlySPECIMENMEMOAUTHORISED SIGNATURE

02 / RETAIN THE CROP

x: 365 · y: 174
width: 222 · height: 90

Example pixel coordinates, relative to the source image.

Synthetic localisation example on a cheque. The same step applies to any signed document: preserve the original image and the crop coordinates so a reviewer can return to the source.

Step 4 · Comparison

Does it match what
you already have on file?

Comparison looks at the shape of the writing, the spacing between strokes and where they sit — measured against the signed samples you already hold. It is a measurement, not a verdict.

Compare the visible geometry

Authorised reference

Synthetic reference signature

Reference sample

From this document

Synthetic candidate signature from the document, with small natural variations

Candidate sample

Similar shapes can still contain natural variation

Look at the opening stroke, proportions, spacing, and terminal flourish. Small changes are normal; evaluate against several authorised samples and the image quality.

Synthetic drawings. No live comparison is performed and no similarity score is implied. The step is identical whether the page is a cheque, a form or a contract.

Use more than one sample

Nobody signs the same way twice. Several authorised samples describe real variation far better than one perfect specimen.

A bad photo is not a bad signature

Glare, blur and low resolution are image problems. They are reported as image problems, not as a mismatch.

You set the bar

How close is close enough depends on your documents, your risk appetite and how much review capacity you actually have.

The awkward cases

The documents that need a human eye.

A useful pilot includes the difficult documents, not only the clean ones. These are the cases that decide whether signature checking helps your operation or gets in its way.

Glare and shadows

An overhead light or a phone-camera flash can wash out the strokes and make a genuine signature look faint.

Too small or too soft

Low-resolution scans lose the stroke detail a comparison depends on. Resolution is worth checking before anything else.

Something covering it

Stamps, annotations and printed text can sit over the signature area and hide part of the writing.

More than one signer

Joint accounts and multi-party forms need a sample set for each person, plus a rule for how many signatures are expected.

A signature that drifted

People’s signatures change over the years. Samples from a decade ago describe a different hand.

Initials and amendments

Initials beside an amendment are not a full signature, and should not be measured as one.

None of these are reasons to skip signature checking. They are the cases to put in front of a system before you trust it with the easy ones.

What a photo
can and cannot tell you.

An image of a document contains visible pixels. It can support analysis of shape, texture, overlap, proportions, and image quality. It does not record the original pen pressure, writing speed, or stroke timing.

That is why signature checking works best as one part of the decision. For cheque fraud detection, the signature result is combined with the other transaction checks, and the final call is recorded with a name against it.

Step 5 · Review

Nothing is approved by a score alone.

Anything unclear, unmatched or simply unusual goes to a person, with everything they need on one screen. That is the part most teams care about most, because it is the part they have to defend later.

What the reviewer sees

The whole document, the signature crop, the sample it was compared with, and the reason it was sent for review.

What gets recorded

The result, the sample version, the reviewer’s action, the reason and the time — stored with the image it came from.

What never happens

A missing signature, a blurred image or a failed check is never quietly turned into a pass.

Technical reference

The same five steps, in system terms.

For engineering, risk and operations teams. Everything above is the operational summary; the definitions, data handling and measurement detail for a build or a procurement review are below.

Detection, localisation, comparison

Detection establishes whether a signature candidate is present in the document image, and whether the region is usable. It is a presence check before any comparison.

Localisation returns the region of the source image that contains the candidate — the bounding box, the image dimensions it refers to, and any rotation, resizing or deskew applied first.

Comparison evaluates the candidate crop against an authorised reference set and returns quality flags and comparison signals. Detection on its own does not establish who signed a document.

Coordinates, crops and image handling

Retain the box, image dimensions, and any rotation or resizing applied before extraction so a crop always maps back to its source image.

Preserve the full page alongside the crop. A crop helps comparison; the full image lets a reviewer inspect nearby stamps, text and layout.

Avoid cutting off a flourish or including unrelated handwriting, and handle multiple signature fields explicitly rather than assuming one box per document.

Reference sets, thresholds and calibration

Compare against authorised samples for the correct person, account or signer, and keep the version and provenance of each reference set.

Thresholds belong to you. Calibrate them on labelled data from your own capture channels, and treat a quality failure differently from a comparison result.

A similarity score alone is neither proof of authenticity nor a calibrated probability, so record the threshold and reference version alongside every result.

Failure handling

Missing signature, unavailable reference, unusable image, and processing failure must not silently become a successful match.

Each condition is returned as its own outcome with a reason, so the workflow can route it to recapture or to a reviewer instead of producing a misleading score.

Retain the failures in your evaluation set as well. Reviewing only the clean successful examples hides exactly the cases that matter.

Integration

Evidence that travels with the result.

Use the Chequedb API integration to connect image processing with your deposit, operations or review system. Agree the exact request and response contract for your deployment.

Input. Document image or signature crop, document type, transaction context, and the authorised reference source when comparison is required.

Processing evidence. Presence outcome, region location, image quality, and comparison signals where available.

Review record. Reference version, rule outcome, reviewer action, reason, and time — linked back to the source image.

Failure handling. Missing signatures, unavailable references, unusable images and processing failures each return their own outcome — never a quiet pass.

For extraction endpoints and the broader data contract, start with the bank check OCR API. Request a sample evaluation to establish the signature-specific fields and thresholds.

An example signature evidence record

Illustrative structure · not an API contract

{
  "example_only": true,
  "document": "cheque",
  "image": { "width": 640, "height": 320 },
  "signature": {
    "presence": "candidate_present",
    "box_px": {
      "x": 365, "y": 174,
      "width": 222, "height": 90
    }
  },
  "comparison": {
    "reference_set": "example-reference-v1",
    "similarity": null,
    "status": "not_evaluated"
  },
  "review": { "status": "pending" }
}

The box matches the synthetic localisation diagram above. Similarity is left empty because no model has evaluated the drawing. Field names and response formats depend on the agreed integration.

Evaluation · Measure each step separately

A useful pilot includes the difficult documents.

Test genuine variation, missing signatures, poor scans, overlapping stamps, and known mismatches across your own layouts and document types. Agree the acceptance criteria before deciding what to automate.

Measures for evaluating signature detection, localisation, comparison, and review
StageMeasureQuestion to answer
DetectionPrecision and recallAre signatures found, and are other marks incorrectly flagged?
LocalisationBox overlap and crop completenessDoes the region contain the signature without cutting off useful strokes?
ComparisonFalse acceptance and false rejectionAt the chosen threshold, which mismatches pass and which genuine samples fail?
OperationsUnusable-image rate and manual-review rateHow much requires recapture or manual work, and can the team handle it?

Report performance by document type, capture channel, layout, and image quality. Keep the evaluation set separate from calibration data, and record thresholds and reference versions with the results.

Signature detection & verification: your questions

What the plain version means, and what the system actually returns.

Does this work on documents other than cheques?

Yes. Signature detection, localisation and comparison need two things: an image of a document, and authorised samples of the signature you expect to find. That applies to deposit slips, account-opening and loan forms, contracts, delivery notes, insurance claim forms and payroll mandates just as much as to cheques. What changes is where the signature is expected, how many signers there should be, and which reference set applies — all of which are configured per document type.

Does this tell us whether a signature is genuine?

No, and it is worth being precise about that. Signature checking answers narrower questions: is the document signed, is the image clear enough to read the signature, and how does the visible writing compare with the samples you hold? A cashier, a bank or a reviewer decides what that means for the payment or the agreement. Treat the result as one signal inside a review process, not as proof of who signed.

Can we use phone photos, or do we need a scanner?

Both work, with different limits. Scanners give even lighting and a predictable resolution. Phone photos are convenient, but glare, shadows and angle matter more, so image-quality checks do more of the work. The honest answer is that capture quality decides how much the comparison can support, which is why quality is reported separately from the result.

What happens when a signature is missing or the image is poor?

Those are reported as their own outcomes rather than as a mismatch. A blank field has nothing to compare, and a blurred or covered signature cannot support a dependable comparison. Both go to a person, with the reason attached, instead of returning a score that quietly means nothing.

What is the difference between signature detection, localisation, and verification?

In plain terms: detection answers "is it signed?", localisation answers "where is the signature?", and comparison asks "how does it look next to the samples on file?" Verification is the combination of those signals with image quality, configured thresholds and review policy. Detecting a mark on its own does not establish who signed it.

Can I integrate signature detection and comparison through an API?

Signature processing can be included in a Chequedb cheque API integration. Define the source-image format, signature-region output, authorised reference source, quality flags, comparison signals, and review handling for your deployment. Request the current API contract and a sample evaluation before integration. The example record on this page is illustrative, not an endpoint specification.

How does signature localisation work on a scanned document?

Localisation identifies a candidate signature region within the document layout. Keep its bounding box and the original image dimensions with the crop. Pixel coordinates must refer to the correct image, especially if the image was resized, rotated or deskewed. Chequedb treats signature regions as part of document field extraction. The diagrams on this page illustrate the relationship between a source image, a region, and its crop.

Can we keep document images and reference signatures on-premise?

Chequedb offers on-premise processing. Define where images and reference sets are stored, which services can access them, retention periods, and the review audit trail as part of the deployment. Confirm the configuration and integration scope during your evaluation.

How many reference signatures should we hold per signer?

More than one, and ideally from the same channels you see in production. A single idealised specimen describes one version of a hand; several authorised samples describe natural variation and reduce false alarms. Keep the version and provenance of each sample so a reviewer can see exactly what the document was compared against.

Can a scanned signature reveal pen pressure or writing speed?

A static scan contains pixels, not the original pen pressure, velocity, or stroke timing. Those dynamic signals require a device that records the signing process. Evaluation of scanned documents should focus on observable image features and reference comparison.

How are two handwritten signatures compared?

Image-based comparison looks at visible shape, proportions, spacing, stroke geometry, alignment, and texture. Compare with several authorised reference samples to account for natural variation. Interpret similarity alongside image quality and thresholds calibrated on representative labelled data. A similarity score is not automatically a probability of authenticity.

Can signature detection distinguish a blank field from a stamp or handwritten text?

A useful presence check separates a visible signature candidate, an empty field, and an unclear or obstructed region. Stamps, printed lines, initials, and handwritten notes can complicate detection. Evaluate these cases on your own document layouts and route uncertain images for review instead of treating any mark as a valid signature.

What accuracy should we expect from automated signature verification?

Performance depends on document layouts, scan quality, the reference set, natural variation, fraud types, and thresholds. Test representative labelled samples and measure detection precision and recall, localisation overlap, false acceptance, false rejection, unusable-image rate, and manual-review rate separately. The drawings on this page are explanatory examples, not benchmark results.

From illustration to evaluation

See what your document images can support.

Bring representative layouts — cheques, forms or both — together with your reference-signature process and the review outcomes you need. Define the detection, localisation, and comparison scope with the Chequedb team.

Plan a signature evaluation

Discuss sample data, API scope, and deployment.