Home / Detection
AI Detection for Redaction
Redaction is only as good as detection. Obscuring every face in a frame does not de-identify anyone if a distinctive tattoo is still visible, the street sign names the road, or the mobile data terminal shows the record. Redactor detects across all of it, and lets you tune how sure it has to be.
How Detection Works
Detection runs at upload through content processing, or on demand against an asset already in the library. You select the object and PII classes to look for and the duration to analyze, so part of a video can be processed on its own.
Detections are stored as timed data, not as marks on frames. That is what lets the studio show an object as a track across the timeline, and what makes a detection searchable and correctable rather than baked in.
What Detection Covers
Objects in video and images
| Class group | Detected |
|---|---|
| People | Persons, faces, heads, body parts |
| Vehicles | Car, bus, truck, bike, boat, airplane, train |
| License plates | Detected and read, so a plate is searchable as text rather than only as an object |
| Weapons | Firearms and other weapons |
| PPE | Including negative forms, so a missing item is detectable |
| Devices and screens | Monitors, displays, laptops, mobile phones, notebooks, and the mobile data terminal in a patrol vehicle |
| Identity documents | Detected as a class of their own |
| Traffic signs and signals | Street signs and house numbers |
| Environmental hazards | Scene-level hazard classes |
| Signatures | Handwritten signatures in documents and images |
| Tattoos | Identifying body art |
Two of those deserve calling out, because they are the ones that quietly break a release.
Tattoos appear as a redaction class in the HIPAA and GDPR compliance profiles for a reason: obscuring a face while leaving a distinctive tattoo visible does not de-identify anyone.
Screens are named explicitly in the CJIS profile. A body-worn camera recording inside a patrol vehicle captures the terminal, and whatever record is displayed on it.
PII in text, transcripts and documents
Classes include person, age, date and time, identifiers, email, phone, URL, IP address, postcode, location, credit card, IBAN, cryptocurrency address, medical license, organization, profession and username, plus country-specific identifiers for the United States, United Kingdom, Spain, Italy, Poland, India, Australia and Singapore.
Custom entities
Organization-specific identifiers are defined three ways: a regular expression, a context-word rule, or a vocabulary list. A case number format that exists only in your agency is a definition, not a feature request.
Clinical text gets a clinical model
Medical narrative is de-identified with a RoBERTa model trained on the i2b2 clinical corpus, running as a recognizer alongside the general ones. The identifiers in a medical record take forms they do not take in a contract.
Tuning
| Control | Range | Default |
|---|---|---|
| Object detection threshold | 10 to 90 | 45 |
| Persistence before a detection counts | 3 to 30 consecutive frames | — |
| OCR thresholds | Set separately for video, image and document | — |
Precision and recall are a tuning decision, not a fixed property of the software. Redaction favors recall, because a missed detection is a disclosure. Investigation review often favors precision. Raising the persistence dial suppresses single-frame false positives at the cost of a few frames of latency on a genuine detection.
Correction
Detections are corrected before anything leaves: bounding boxes adjusted, duplicates merged, tracks split, objects renamed, false positives deleted, and missed objects added by hand. Automated detection proposes; a person disposes.
Where It Matters
- Body camera footage — screens, plates and bystanders in a moving scene
- CCTV and surveillance — crowds, occlusion and low light
- Face redaction — and everything that survives it
- Law enforcement — the CJIS class set
What Detection Does Not Do
- Detection quality follows source quality. A clear, well-lit source produces better detection than a poor one.
- PII detection covers 10 languages. Transcription covers 82. Do not read the second number as the first. Outside those 10, the transcript still supports keyword and pattern search.
- The clinical de-identification model is one recognizer among those the PII pipeline runs; its language coverage follows the pipeline.
- Automatic redaction applies what the detector found at the configured threshold. Where a release is consequential, detection review is the step that catches what the threshold admitted or missed.
How Detection Is Evaluated
- Object detection spans people, vehicles, weapons, PPE, devices and screens, identity documents, traffic signs and environmental hazards, with signatures and tattoos as separate classes.
- License plates are read as text, not only detected as objects.
- Confidence is configurable per class, range 10 to 90, default 45, with separate thresholds for detection, recognition and redaction.
- A frame-persistence dial of 3 to 30 consecutive frames trades false positives against latency.
- Custom entities are defined by regular expression, context-word rule or vocabulary list.
- A dedicated clinical de-identification model handles PHI alongside the general recognizers.
- Every detection is correctable before output: merge, split, rename, delete, adjust, or draw what was missed.
Test It on Your Own Footage
Send us a file with something difficult in it. Crowds, poor light, a screen in frame. We will show you what the detector returns and what the thresholds do to it.