noviscanby Secufor

Technical transparency

Which models
actually run under the hood.

Noviscan trains no proprietary model and does not ask you to trust a black box. It orchestrates published research models that you can name, download and verify yourself. This page says which ones, and where they come from.

See supported cameras Model repository

The recognition engine

Detect, then recognise

Two chained models, published by research and distributed through OpenCV Zoo. They run locally, on your own machines.

Detection — YuNetLocates faces in the image. File face_detection_yunet_2023mar.onnx (~230 KB).
Recognition — SFaceTurns each face into a 128-dimension embedding, compared against the faces you enrolled. File face_recognition_sface_2021dec.onnx (~37 MB).
Apache-2.0 licencePermissive: commercially usable, royalty-free. You can inspect the models, their code and their papers without depending on us.

A face embedding is not a photo: it is a list of numbers, and the original face cannot be reconstructed from it. It only serves to measure similarity against the faces you enrolled yourself.

Supply chain

Verifiable models, not a black box

The weights are not shipped with Noviscan: they come from their original repository, and their integrity is checked.

Official sources, verified checksumsDownloaded from the opencv/opencv_zoo repository — not from a Secufor mirror — then checked against a pinned SHA-256: a truncated or tampered transfer is rejected.
Offline installation supportedAutomatic download can be turned off: you place the files by hand, on a machine with no Internet access.
100% local processingInference runs on your machine. No image and no face embedding is sent to any third-party service.

Verifiable, not just claimed

You don't have to take our word for it

Most face recognition products rest on a proprietary algorithm: you have to trust the vendor. Noviscan takes the opposite route — the models are public, named, and checkable by you.

Public models, not a black boxWe name the exact models in use. Their papers, code and results are published by their authors: you can review them without going through us.
A chain you can verify end to endModel origin, licence, checksum of the installed file: every link can be checked. A proprietary algorithm, however well rated, gives you no such verification.
Your data serves nothing elseNo training on your footage, no upload to our servers: the models arrive already trained and run on your machine, as they are.

One point of honesty that applies across the field: published face recognition scores are measured on academic datasets, often frontal and good quality — they do not transfer as-is to surveillance footage (angle, distance, lighting). That is exactly why Noviscan applies its own quality thresholds: a face too far in profile or too blurry raises no alert, rather than raising a false one.