STEP / Computer vision for access
What if finding a patient’s information didn’t depend on a reliable internet connection? STEP explores face-assisted record lookup on a local computer.
01 / Under the hood
Follow one capture through the current pipeline.
Interactive illustration · Fictional data. This page explains the process; it does not identify faces or access your camera.
Locate → align → crop
128 values · A learned face template
Illustrative colors, not an actual biometric template.Highest score → threshold check
Linked by unique patient ID
A possible match is not a verified identity.
Step 01 of 05
The desktop app accepts a webcam frame or an existing photo. Processing happens on the same device. Enrollment requires exactly one detected face.
The actual app processes captures in memory rather than saving them as photographs.
YuNet detects the face and landmarks. SFace aligns the crop and produces 128 values. SQLite links that template to patient details by ID. A new capture is compared to stored templates using cosine similarity; the highest score yields a candidate if it meets the threshold. Otherwise the result is unknown. An operator reviews the candidate before opening its record. Direct ID lookup is also available.
Recognition implementation ↗Storage implementation ↗Try the decision rule
Move the cutoff or change the sample. The same score can produce a different decision.
Candidate A → operator review
These are made-up similarity scores, not confidence percentages. The repository’s default 0.363 threshold has not been calibrated for this application.
02 / Why it matters
In a clinic with limited connectivity or a small technology budget, a cloud-dependent lookup system may be difficult to sustain. My goal is to explore a simpler path to finding information, including in lower-income countries.
Choose a constraint to see the design response.
The design response
After installation and model downloads, enrollment, recognition, and record lookup work offline. Losing connectivity does not require sending a new image to a cloud service.
What I demonstrated
The original face_recognition/dlib version achieved approximately 85% identification accuracy in my testing with 25 sample faces, and won first place statewide across all divisions at the New York STEP Science Fair.
What comes next
The current OpenCV version has not had a new accuracy evaluation. Faster lookup and more time for care are intended benefits, not measured clinical outcomes. Real-world use would require representative testing, consent, secure storage, access controls, and safeguards against false matches.
Educational prototype · not for clinical use or identity verification.
03 / The project in pictures
Select a screenshot to take a closer look.
Enrollment, live recognition, and local record lookup.
First place · New York STEP Science Fair.
A space for the poster, demonstration, or science-fair story.
Technical walkthrough and application screenshot: STEP repository, reviewed September 15, 2026. Historical award and testing results are from my project history. STEP branding belongs to its respective owner.