Self-developed neural engines
End-to-end edge pipeline spanning detection, 98-pt alignment, compliance & ISO quality gating, dual-spectrum PAD, and adaptive matching.
SUNTEK BIOMETRIC ENGINE · 7-MODULE EDGE ARCHITECTURE
In real-world access control, end-to-end reliability depends on every stage working in concert. SUNTEK integrates 7 purpose-built neural modules into a single memory-streamed RK3576 NPU engine — trained on tens of millions of authorized portraits and verified against physically isolated blind-test sets.
End-to-end edge pipeline spanning detection, 98-pt alignment, compliance & ISO quality gating, dual-spectrum PAD, and adaptive matching.
Real-time 30fps tracking decoupled from on-demand deep verification so live preview never stalls under heavy gate load.
Trained on tens of millions of authorized portraits and strictly validated against tens of thousands of physically isolated blind-test samples.
On-device NPU embeddings maintain high cosine consistency with cloud server enrollment across bare faces, masks, and lighting shifts.
01RK3576 DUAL-CORE LOAD BALANCING & 7-MODULE MATRIX
SUNTEK Biometric Engine · Streamed Dual-Core NPU Execution
All 7 self-developed neural modules operate within a unified memory-streamed container with zero disk unpacking, while hardware image acceleration prepares camera frames directly for NPU inference without CPU overhead.
Lightweight detection, landmarks & gating
Executes multi-scale face tracking, 98-pt dense landmarks, wearing & ISO quality gating, 850nm NIR material checks, and blink verification at full 30fps without dropping frames.
Triggered only after quality & pose pass
Runs spatial-frequency visible-light anti-spoofing and 512-D adaptive feature extraction on the companion NPU core—never blocking front-end 30fps tracking.
Throughput + sub-pixel angular accuracy
Hardware-fused input normalization paired with INT8 backbone efficiency and FP16 precision on coordinate regression and projection heads preserves edge-cloud consistency.
| Module & Name | Functional Positioning | Input Spectrum | Execution Domain | Quantization Strategy | Core Business Value |
|---|---|---|---|---|---|
| Module 01 Multi-Scale Face Detection & Tracking | Multi-scale face localization, stable ID tracking, and coarse anchor generation across indoor/outdoor backgrounds | Visible RGB + 850nm NIR | Real-Time Core (30fps) | INT8 Backbone + FP16 Box Heads | Millisecond capture with minimal false triggers across 0.3–2.5 m |
| Module 02 98-Pt Dense Landmarks & Adaptive Alignment | High-density anatomical facial landmark localization and per-point visibility sensing for adaptive upper-face alignment | Visible RGB | Real-Time Core (30fps) | FP16 High-Precision Regression | Stable facial alignment under large poses, masks, and eyewear |
| Module 03 Multi-Dim Wearing Compliance & ISO Quality Gate | Unified 5-zone visibility, 5-state wearing compliance, sunglasses & dual-occlusion veto, 3D pose, and ISO/IEC 29794-5 UQS | Visible RGB | Real-Time Core (< 2 ms) | INT8 + FP16 Sensitive Heads | Accurate mask routing, half-mask prompts, and blur/glare rejection |
| Module 04 Visible-Spectrum Spatial-Texture PAD | 1:1 native spatial-frequency texture sensing in the 112×112 facial core with a 56px concentric zero-ring (and 2.0× context on mobile) against screen & print spoofs | Visible RGB (Nested Dual-Stream) | On-Demand Core (Post-Gate) | INT8 + High-Freq Texture Fidelity | 0.00% lab APCER under fixed threshold; wide-CCT and glasses-friendly |
| Module 05 850nm NIR Physical Material PAD | Exploits dermal subsurface scattering vs. screen/toner/silicone absorption with optical eyeglass hotspot suppression | 850nm Near-Infrared | Real-Time Core (Dual-Cam Sync) | High-Precision Lightweight | 0.00% screen & toner APCER in blind tests; reliable night & backlight defense |
| Module 06 Tri-Zone Adaptive Face Recognition | Generates a unified 512-D hypersphere feature organized into three orthogonal anatomical zones with bounded isometry | Aligned Visible RGB | On-Demand Core (Edge-Cloud Parity) | INT8 Backbone + FP16 Projection | Single template for bare, masked & bangs modes; 50K 1:N search in < 1 ms |
| Module 07 Dual-Spectrum Dynamic Eye-Blink Verifier | Sub-millisecond eye open/closed transition tracker operating simultaneously in visible light and 850nm near-infrared | RGB + 850nm NIR Dual-Domain | Real-Time Core (< 1 ms Trigger) | Lightweight High-Precision | 98.37% F1 optional dynamic liveness lock for high-security checkpoints |
02TRI-ZONE 512-D ADAPTIVE FEATURE SPACE
SUNTEK Tri-Zone Orthogonal Subspace · Single Root Template Across All Wearing Modes
Rather than maintaining separate enrollment templates for bare faces, masks, and heavy bangs, SUNTEK structures one 512-D unit vector across three orthogonal anatomical bands with bounded isometric regularization—enabling dynamic routing from a single root template.
||v||₂ = 1.0 · Dynamic routing driven automatically by wearing & visibility gating
Leverages the complete 512-D representation across all three zones with controlled truncated-cosine resilience against localized specular glare.
Links the Upper and Mid Core zones upon confirming a properly worn mask, enforcing upper-zone completeness, zero low-quality compensation, elevated decision margins, and multi-frame consensus.
Automatically shifts matching weight to the Mid Core and Lower zones when bangs obscure the forehead—slashing cosine drift to 0.0017 with no photo re-upload required.
Bounded isometric manifold constraints prevent cross-pose feature collapse and preserve biometric stability across wide yaw/pitch angles and multi-decade aging without re-enrollment.
03UNIFIED COMPLIANCE, ANTI-DISGUISE & ISO QUALITY ENGINE
Module 03 · Single-Pass Edge Gate Aligned with ISO/IEC 29794-5 (< 2 ms)
Replacing brittle single-purpose classifiers and heuristic blur rules, Module 03 evaluates facial zone visibility, 5-state wearing compliance, sunglasses & dual-occlusion redlines, 3D head pose, and ISO/IEC 29794-5 image quality in a single sub-2 ms pass.
04DUAL-SPECTRUM PHYSICAL & DIGITAL ANTI-SPOOFING
Modules 04, 05 & 07 · ISO/IEC 30107-3 Dual-Spectrum PAD Architecture
Single-spectrum liveness struggles against either high-refresh OLED replays or realistic 3D masks. SUNTEK fuses spatial-frequency visible-light sensing, 850nm NIR subsurface material verification, and sub-millisecond dynamic blink detection.
Preserves 1:1 native pixel sampling in the 112×112 facial core isolated by a 56px concentric zero-ring on edge/server (paired with a 2.0× outer context ring on mobile capture)—preventing frequency loss from naive scaling.
Exploits physical reflectance laws in the 850nm near-infrared band comparing live human dermal subsurface scattering against artificial materials, paired with optical eyeglass hotspot attenuation.
Tracks dense orbital landmarks synchronously in visible light and 850nm near-infrared to capture natural physiological micro-blinks as an optional high-security dynamic lock.
Co-engineered RK3576 binocular terminals with 9:16 touch displays, zero-lag dual-spectrum cameras, adaptive 850nm IR fill light, and solutions for campuses, cleanrooms, and banks.
Explore devices & solutions99.91% bare / 99.97% masked 50K open-set retrieval, 0.00% dual-spectrum PAD spoof pass rate, 0.25% demographic parity spread, and native EU AI Act, GDPR & ISO alignment.
View benchmarks & complianceUnified edge AAR, multi-arch Node.js server SDK, and 2.5 MB mobile capture across Android, iOS, HarmonyOS NEXT, Web WASM, Flutter, RN & UniApp with <1ms 50K search and QR/offline activation.
Read SDK & licensing guideSUNTEK FACE AI