Image-based product authentication
The premise sounds too good: photograph a suspect pack, get a verdict, change nothing about your packaging. It genuinely works — but only if the system is built on the right idea about what genuine means, and only if the photograph is good enough to measure. Both of those deserve explaining, because they are where implementations succeed or fail.
Image-based product authentication compares a photograph of a suspect pack against a statistical model of genuine variation learned from photographs of known-good units. It measures geometry, colour and print structure against tolerance bands rather than against a single reference image, because legitimate printing varies between units. It requires no code, label or packaging change and works on stock already in the market.
Key takeaways
- Genuine is a distribution, not an image. Systems that compare against one golden sample reject genuine product constantly.
- 20–50 genuine units per SKU across several batches is the realistic reference requirement.
- Capture quality is the dominant failure mode in the field — more verifications fail on blur and glare than on anything analytical.
- It catches redrawn and reprinted fakes very well, and uniform-scale exact copies poorly. Know which you have.
Why one golden image does not work
Offset and rotogravure printing vary legitimately from unit to unit, and the numbers are larger than people expect:
| Source of variation | Typical legitimate range |
|---|---|
| Colour-to-colour registration drift | ±0.1–0.2 mm |
| Ink density drift | ΔE 2–4 |
| Substrate shade between reels | Visible to the eye |
| Print-to-die-cut drift | ±1–2 mm |
Compare a genuine pack against a single golden image and any of those will look like a discrepancy. The correct approach is to photograph a set of genuine units and treat the spread across them as the definition of acceptable — the standard deviation across genuine references is the legitimate print variation. The system then asks whether a suspect sits inside that envelope, and by how many standard deviations it sits outside.
What the system actually measures
Rectification
The suspect photograph is matched to a reference frame using scale-invariant feature matching and a homography, so a photo taken at an angle can be compared against one taken flat. Without this step, perspective alone would fail every verification.
Geometry against learned bands
A set of stable anchor points is detected on the artwork, and the pairwise ratios between them are compared against tolerance bands learned from the genuine set. Ratios are used rather than absolute distances because they survive changes in camera distance.
Colour against learned ranges
The pack is divided into a grid and each cell's colour is compared, in a perceptually uniform colour space, against the range observed across genuine samples. This catches the ink shifts that redrawn artwork almost always carries.
Print structure where resolution allows
Halftone frequency and screen angle are properties of the press. When the capture resolves them, this is powerful evidence; when it does not, the layer should degrade gracefully rather than guess.
Model-based comparison and adjudication
A vision model extracts specific features from both images, compares them element by element, and then adjudicates with the measurements in front of it. The output that matters is not the score but the list of concrete differences.
Explainability is the deliverable
A brand-protection team cannot act on "score 0.24". They can act on "header band: genuine muted olive-gold, suspect deep cobalt blue" and "guilloche: genuine pale delicate fine lines, suspect darker and coarser". Those are reviewable by a human, defensible to a distributor, and usable in a legal notice.
When evaluating a system, ask to see the evidence output for a real counterfeit, not the dashboard. If the answer is a number and a red banner, it will not survive its first argument with a large distributor who insists the stock is genuine.
Capture is the real constraint
In deployment, far more verifications fail on photography than on analysis. What helps:
- Live quality gates — sharpness, glare percentage and steadiness measured on the video stream before the shutter is available at all.
- Manual capture, never automatic. Let the gates tell the user when the frame is good and let the user press the button. Auto-shutter fires on frames the user has not composed.
- A fixed shot list. Full face, logo close-up, label and batch panel. Comparing like with like matters more than capturing more.
- Untouched uploads. Store what the camera produced. Re-encoding destroys the fine structure the forensic layers rely on.
- An honest INCONCLUSIVE verdict when the gates were not met, rather than a guess.
Where it works well, and where it does not
| Situation | Suitability |
|---|---|
| Printed cartons, flat panels, offset or gravure | Strong |
| Redrawn or reprinted counterfeit artwork | Strong — this is the core case |
| Labels on bottles, cans, curved surfaces | Weak without cylindrical unwarping |
| Mostly flat-colour or repeating-pattern artwork | Weak — too few stable anchors |
| Exact artwork copy at uniform scale | Weak on geometry; caught by colour and model comparison instead |
| Transparent or highly reflective packaging | Difficult — glare dominates |
| Images forwarded through messaging apps | Triage only — re-encoding destroys the evidence |
Any vendor unwilling to name their weak categories has not deployed enough to know them.
What deployment looks like
- Pick two or three high-risk SKUs. Not the catalogue.
- Photograph 20–50 genuine units per SKU across several production batches, in controlled conditions.
- Run the first fifty field verifications as calibration, not production. Thresholds tuned on anything other than your own packs will need moving.
- Give field officers the shot list and a clip-on macro lens if print structure matters for your category.
- Review every SUSPECT by hand for the first month and feed the disagreements back into the thresholds.
- Only then expand, by SKU rather than by geography.
Treating the first phase as calibration is the single strongest predictor of whether the system is trusted a year later. The combined design with serialisation is in the anti-counterfeit software guide.
Frequently asked questions
How can software tell a fake from a photograph?
By comparing measurements of the suspect image — geometry between stable anchor points, colour in a perceptually uniform space, and print structure where resolution allows — against tolerance bands learned from photographs of genuine units, then having a vision model compare specific features and adjudicate with those measurements in front of it.
Why do you need many genuine samples rather than one?
Because printing legitimately varies between units: registration drifts ±0.1–0.2mm, ink density moves ΔE 2–4, substrate shade differs between reels and print-to-die-cut drifts ±1–2mm. Comparing against a single golden image treats all of that as discrepancy and rejects genuine product. The spread across genuine samples defines what acceptable means.
How many photographs of genuine product are needed?
Typically 20 to 50 units per SKU, spanning several production batches so the model captures batch-to-batch variation rather than just one press run. Too few samples produce a tolerance band so tight that ordinary genuine product falls outside it.
What kinds of counterfeit does image authentication catch best?
Redrawn and reprinted artwork, which is the overwhelming majority of counterfeit packaging. Those fakes carry ink shifts, typography substitutions and geometry differences that are measurable even when they look convincing to the eye. It is weakest against exact artwork copies reproduced at uniform scale.
Does it work on bottles and curved packaging?
Not reliably without cylindrical unwarping, because the geometric comparison assumes a flat face. A label on a round bottle mis-rectifies and can produce false geometry failures. Flat printed cartons and panels are the strong case.
Can I use photos customers send on WhatsApp?
As a triage signal only. Messaging apps re-encode images, which destroys the fine print structure the forensic layers depend on and adds noise to the geometric anchors. Verification should use images captured in the app and stored untouched.
How long does a verification take?
A genuine multi-layer verification takes tens of seconds — rectification and measurement, then several model calls for extraction, comparison and adjudication. Anything advertised as instant is running one shallow check rather than a layered one.