A loan, an insurance card, a food ration or a vote is meant for each person once. De-duplication checks every new enrollment against everyone already registered, so a second registration under another name is caught before anything is issued.
If you keep a list of people who receive something, a few names on it will belong to the same person: two offices, two spellings, sometimes a deliberate second try. Each of them costs you a second payment, card or ration every time you hand things out. An iris scan at sign-up finds that person while they are still at the desk, before the second record exists.
An example with round numbers. Suppose you register 10,000 people for a monthly benefit worth $50, and 1 in 100 signs up twice. That is 100 extra payments, $5,000 every month. The cloud costs $99 a month for a register of this size, and the $990 starter kit adds two scanners, the first three months and setup. These numbers are only an example.
To start, you need a scanner for each desk and a plan that fits your number of people. See the prices or tell us about your program.
The search runs in memory and in parallel. On two CPU cores our matcher searches 200,000 enrolled eyes in about 0.17 seconds, and the time grows linearly with the gallery, so even large registers answer while the person is still at the desk.
De-duplication is harder than verification. A 1:1 check compares a person with one claimed record; a 1:N search compares them with everyone, so the chance of a false match grows with the size of the register. The modality has to produce very few false matches per comparison. Iris codes do: the scores of different people cluster tightly around chance, which is why iris has been the modality of choice for large registers since Daugman's work in the 1990s.
It also suits the people these programs serve. Capture is contactless, it works for farmers and manual workers whose fingerprints are worn, and the iris pattern does not change with age the way a face does. Face recognition is easier to capture but needs much more care at 1:N, because lookalikes and relatives produce close scores.
With N people enrolled and a false match rate p per comparison, a new applicant falsely matches someone with a probability of roughly N × p. For a register of 100,000 people that product has to stay well below one, which means a per-comparison false match rate far under one in 100,000. Two measures get there:
The Cloud Iris API de-duplicates at enrollment by default. Each organization has its own gallery; a duplicate is refused with the id of the person already enrolled:
POST /persons
{ "name": "...", "leftEye": "<base64>", "rightEye": "<base64>" }
409 Conflict
{ "message": "This iris is already enrolled", "duplicateOf": "8f3c...", "eye": "left", "score": 0.81 }If images must stay on your servers, the same search runs on premises, or in your own code with the open-source iris recognition SDK (identify() returns the best match or nothing). Templates and images can be exported in ISO/IEC 19794-6, so the register is not tied to us.
The enrollment is stopped and the operator sees the existing record. Your procedure decides the rest: a returning client is simply recognized, a second application under another name is refused or sent for review.
Yes, if it holds iris images. We run every record against all others and return groups of records that belong to the same person, with scores, for your team to review.
Any scanner that produces standard near-infrared iris images, from a USB scanner on an agent's laptop or Android tablet to dual-iris cameras. See pricing for the kits we supply.
Tell us who you enroll, how many people and where. We will propose a pilot on your own population with success criteria agreed in advance.