SmartScanPro runs two engines on one API surface: a remote photoplethysmography (rPPG) signal-processing stack for the human face, and a purpose-built OCR + layout model for medical device displays and clinical documents.
Every stage is auditable and benchmarked independently. Nothing is a black box.
Lightweight CNN locates the face at 30+ fps on a mid-range phone. ROI tracked across frames.
Forehead & cheek regions segmented — these maximise the green-channel pulsatile signal.
POS / CHROM / ICA combined with a learned skin-melanin-aware weighting for cross-Fitzpatrick fairness.
Per-frame RGB means → detrended, bandpass-filtered (0.7–4 Hz), windowed into 10 s overlapping frames.
SNR, motion, lighting-variance checks. Reject low-quality windows and tell the user why.
FFT + autocorrelation ensemble. Peak picking with prior from previous windows.
Kalman-smoothed PPG estimate — feeds HRV, SpO₂, respiratory rate, and pulse-wave features.
HRV (RMSSD, SDNN, pNN50), respiratory sinus arrhythmia, stress index, autonomic balance.
Age / gender / BMI conditioning heads trained against large-cohort reference data for wellness BP.
A skin region's green channel varies by ~1 % with each heartbeat. Here's what that looks like after our bandpass filter.
Purpose-trained OCR + layout model on 400k+ medical-device photos. Reads 7-seg LCD, e-ink, backlit LCD, thermal receipts.
Each extracted value is bound to a pixel bounding box — auditors can click a JSON field and see where it came from.
Every value carries a calibrated probability. Use it to triage manual review or set auto-approval thresholds.
The live demo uses the same engine we ship in production.