IrisVision — Human Iris Localization & Validation

Real-time, in-browser localization of the human iris with an independent, classical-vision validation stage. Everything runs on the client (WebAssembly + WebGL) — no frame ever leaves the device.

Live demo: https://subzero-005.github.io/iris-vision/

Features

• Live webcam, image upload or bundled sample photo (drag & drop supported; ?sample auto-runs it)
• Iris localization from MediaPipe Face Landmarker's 478 landmarks (5 per iris)
• Sub-pixel circle fit — centred algebraic (Kåsa) least-squares fit with RMS residual
• Independent validation of every fit:
  • Daugman's integro-differential operator on the lateral limbus arcs (pure JS)
  • OpenCV.js HoughCircles on an upsampled, median-filtered window around the iris
  • Radial intensity contrast between iris and sclera, sampled inside the eyelid polygon
  • Eye openness (EAR) so closed eyes are reported as OCCLUDED, not guessed
• Eye-state analytics — blink detection with hysteresis, blink rate, gaze direction with recentring
• Distance estimation from the ~11.7 mm human iris diameter and a pinhole camera model
• Live charts for EAR and validation score, plus a gaze heat map
• Export an annotated PNG or a JSON measurement report

Pipeline

frame ─▶ Face Landmarker ─▶ iris limbus points ─▶ least-squares circle
                                                       │
            ┌──────────────────────────────────────────┤
            ▼                    ▼                     ▼
  Daugman + Hough check   radial contrast      eye aspect ratio
            └──────────┬─────────┴─────────────────────┘
                       ▼
         score = 0.45·geometry + 0.35·contrast + 0.20·openness
                       ▼
            PASS (≥ 0.60) · FAIL · OCCLUDED (EAR < 0.17)

Signal	Definition
Geometry	0.65·g(Daugman) + 0.35·g(Hough), with g = (1 − Δc / 0.6r) · (1 − |1 − r′ / r| / 0.5)
Contrast	(I_ring − I_iris) / I_ring, iris ≤ 0.7r, ring 1.25r–1.9r
Openness	clamp((EAR − 0.15) / 0.12)
Distance	Z = f · 11.7 mm / d_px, f = (W/2) / tan(HFOV/2), HFOV ≈ 60°

If OpenCV.js fails to load, geometry falls back to the Daugman operator alone.

Run locally

Camera access requires a secure context, so serve the folder over localhost:

python3 -m http.server 8000
# open http://localhost:8000

Deploy

The site is static. On GitHub: Settings → Pages → Deploy from branch → main / root.

Tech

MediaPipe Tasks Vision · OpenCV.js 4.9 · Canvas 2D · vanilla ES modules — no build step.

Limitations

IrisVision is a research and portfolio demo, not a certified biometric system. Distance assumes a ~60° webcam field of view, and gaze is a geometric estimate without per-user calibration beyond recentring.

Credits

Sample photo: astronaut Kalpana Chawla, official NASA portrait (public domain, via Wikimedia Commons).

License

MIT