Physically isolated blind-test samples
Over 500,000 overlapping identities purged via strict hash deduplication across tens of millions of compliant training images.
PHYSICAL TEST ISOLATION · 11 BLIND-TEST GATES · FIXED COMMERCIAL THRESHOLDS
Before training SUNTEK, tens of thousands of blind-test sample hashes across 13 evaluation protocols were locked read-only and physically isolated from the training corpus, purging over 500,000 overlapping identities. Every metric below is measured under fixed commercial production thresholds and Out-of-Distribution (OOD) protocols.
Over 500,000 overlapping identities purged via strict hash deduplication across tens of millions of compliant training images.
99.91% bare-face TPIR @ FPIR=10⁻³ and 99.97% masked TPIR on 50,000 enrolled + 20,000 unregistered impostors.
Measured at a single fixed commercial threshold (0.50) across visible-light nested PAD and 850nm NIR subsurface sensing.
Balanced 24,000-pair multi-demographic blind audit (99.95% / 99.78% / 99.75% / 99.70% across 4 groups).
01COMPLETE G-01 ~ G-11 BLIND-TEST SCORECARD
Frozen Hold-Out Benchmarks + Horizontal Flip + Unseen OOD Audits
Instead of tuning separate thresholds per dataset, SUNTEK is evaluated under fixed commercial contracts across 50K open-set galleries, medical/N95 masks, heavy makeup, heavy bangs, disguise blocking, and dual-spectrum spoofing.
| Gate ID | Evaluation Scenario | Conventional Baseline | Release Gate Target | SUNTEK Measured Result (with OOD & Flip) |
|---|---|---|---|---|
| G-01 | Cross-4-Demographic Fairness Blind Test (24,000 balanced verification pairs) | Mean ~93% · Spread > 3.0% | Mean ≥ 98.80% · Spread ≤ 0.60% | Mean 99.80% · Max Spread 0.25% (99.95% / 99.78% / 99.75% / 99.70%) |
| G-02 | 50,000-Person Open-Set 1:N + 20,000 Unregistered Impostor Stress Test | Open-set recall drops sharply under masks | Bare ≥ 93.8% · Mask ≥ 83.0% | Bare TPIR@10⁻³ = 99.91% · Masked Upper-Zone Subspace TPIR@10⁻² = 99.97% |
| G-03 | Real-World Street Cross-Mask 1:1 Verification (Unmasked vs. Masked) | Severe cross-mask drift at low FAR | Street Cross-Mask TAR@10⁻⁴ ≥ 94.5% · Masked ≥ 97.0% | Street Cross-Mask TAR@10⁻⁴ ≥ 99.9% (Acc 99.98%) · Masked Acc = 99.88% (TAR@10⁻⁴ ≥ 99.9%) |
| G-04 | Heavy Makeup ↔ Bare Face Cross-Modal Verification & Makeup Liveness | Elevated liveness false rejects under heavy makeup | TAR@10⁻⁴ ≥ 98.0% · BPCER ≤ 0.40% | TAR@10⁻⁴ ≥ 99.9% (Acc 99.95%) · Fixed-Threshold (0.50) Heavy Makeup Liveness BPCER = 0.00% |
| G-05 | Heavy Bangs Covering Eyebrows & Hairstyle Change Blind Test | Notable cosine drift under forehead occlusion | Δcos < 0.0150 · BPCER ≤ 0.40% | Mid-Lower Subspace Δcos = 0.0017 (Full = 0.0032, Flip = 0.0078) · BPCER = 0.00% |
| G-06 | Access Wearing Compliance & Disguise Gate (Compliant Mask vs. Sunglasses, Half-Mask & Full-Veil/Half-Face) | Disguise leaks & half-mask false alarms | Mask recall ≥ 99.2% · Near-zero disguise pass | Compliant Mask Recall 99.33% · Sunglasses Block 98.93% · Dual/Full-Veil/Half-Mask/Half-Face Pass = 0.00% (@0.50) |
| G-07 | Full Beard & Multi-Skin-Tone Bare Face False-Mask Immunity | Beards & dark skin tones trigger false masks | Beard ≤ 0.15% · Bare ≤ 0.15% | Full Beard False Mask = 0.00% · Dark-Skin Bare Face = 0.09% (Unseen OOD Set = 0.00%) |
| G-08 | Extreme Frontal-Profile Pose & 30-Year Cross-Age Blind Benchmarks | Accuracy degrades on steep yaw & multi-decade aging | Extreme Profile ≥ 99.30% · 30-Year Age ≥ 98.30% | Extreme Profile = 99.97% (Single 99.99%) · 30-Year Age = 99.98% · Standard 1:1 = 99.82% · Cross-Age = 95.78% |
| G-09 | Multi-Scale Face Detection, 98-Pt Dense Landmarks & Dynamic Eye Blink | Background ghost boxes & coarse landmark error | AP ≥ 93%, 0 ghost boxes, NME -30% | Multi-Scale Medium AP = 94.06% · Background Ghost Boxes = 0 · 98-Pt NME = 3.23% (-65.5%) · Blink F1 = 98.37% (AUC = 99.66%) |
| G-10 | Visible-Light Cross-CCT (2200K/6500K) Anti-Spoofing & Screen-Replayed Deepfake Test | Color-temp drift & high-res Deepfake screen leaks | APCER ≤ 0.40% · CCT drift < 0.30% | Fixed Threshold 0.50: 6500K/2200K BPCER = 0.00% · Physical & Deepfake APCER = 0.00% (Blue-Light Glasses BPCER 0.30%) |
| G-11 | Bimodal 850nm NIR Material PAD (OLED/LCD Screens, Laser Toner Print & 3D Silicone Masks) | Screens, toner prints & 3D heads bypass RGB-only PAD | Screen 0.0%, Toner ≤ 0.2%, Silicone ≤ 0.6% | 850nm Real BPCER = 0.00% · OLED/LCD APCER = 0.00% · Laser Toner APCER = 0.00% · 3D Silicone APCER ≤ 0.55% |
Note: All metrics above are measured on SUNTEK standard isolated blind-test sets under fixed commercial thresholds and do not constitute a guarantee for every environment or deployment. Real-world performance depends on ambient lighting, mounting geometry, and user cooperation—validate with field tuning on target hardware before production.
02ARCHITECTURAL PARADIGMS & OOD ROBUSTNESS
Design Fit Across Compute Budgets · SUNTEK Isolated Blind-Test Evidence
Different face recognition architectures are engineered for different compute budgets and operating assumptions: heavy cloud networks maximize capacity on unconstrained server GPUs, lightweight global CNNs optimize general verification efficiency, and SUNTEK specializes in offline RK3576 NPU access control under partial occlusion.
| Dimension / Scenario | Lightweight CNN | Cloud Heavy (ResNet / ViT) | SUNTEK (Edge 3-Zone) | Design Trade-off & Fit |
|---|---|---|---|---|
| Edge NPU Deployability | Engineered for low-power edge inference; lightweight quantized path suited to standard single-stream checks | Designed for floating-point server GPUs; large parameter footprint and global attention ops constrain embedded INT8 NPU deployment | Edge-quantized 101-layer backbone (24.18 GFLOPs; edge-cloud cosine fidelity ≥ 0.996) | Server-class networks excel in centralized datacenters; SUNTEK natively fits RK3576 dual-core NPU offline scheduling |
| Standard Unoccluded 1:1 Verification | Mature and stable commercial verification accuracy on standard frontal unoccluded faces | Delivers top-tier representation capacity across standard unoccluded portrait benchmarks | 99.82% (Single-frame) / 99.78% (Multi-frame fused) on isolated blind tests | Matches server-grade baseline precision on clean frontal captures within a compact edge compute budget |
| Large-Angle Profile & Cross-Age Span | Constrained feature capacity leads to gradual drift under steep yaw angles and multi-decade aging | High upper bound in unconstrained wild poses, complex age spans, and high-capacity server matching | 99.97% Large-angle profile · 99.98% 30-yr age span · 61.40% Complex-pose TAR@10⁻⁴ | Bounded isometric geometry maintains turnstile capture stability and minimizes periodic gallery re-enrollment |
| 50,000 Open-Set Masked Search | Coupled full-face vector representation without dedicated partial-occlusion subspace isolation | Trained for holistic unoccluded semantics; strict-threshold masked search experiences lower-face dilution unless adapted | 99.97% Masked Upper-Zone (FPIR=10⁻²) / 99.91% Bare-Face (FPIR=10⁻³) | Difference stems from task assumptions: cleanroom turnstiles require wearing-aware routing and orthogonal upper-zone isolation |
| Heavy Bangs Feature Preservation (Δcos) | Global feature vectors shift when eyebrows and forehead are heavily covered | Deep global receptive fields or holistic attention propagate forehead changes across the full embedding | 0.0017 (Mid-Lower Subspace) / 0.0032 (Full 512-D) | Tri-band orthogonal routing automatically relies on unoccluded mid-lower facial bands when heavy bangs appear—no re-enrollment needed |
| Access Gating & Dual-Spectrum PAD | Typically operates as a standalone feature extractor requiring separate external quality and liveness modules | Focuses on server-side static or single-stream comparison rather than terminal dual-ISP sync and voice compliance cues | 7 in-memory co-scheduled engines (ISO 29794-5 quality, half-mask coaching, RGB + 850nm NIR PAD) | Real-world turnstile reliability depends equally on full-pipeline optical co-tuning, wearing governance, and liveness defense |
Architecture Comparison Note & Objective Fairness Statement: This matrix compares engineering trade-offs across neural network paradigms from a system-selection perspective and does not target, map to, or evaluate any specific third-party vendor, commercial product, or open-source repository. Both lightweight global CNNs and heavy cloud networks perform strongly within their intended unoccluded or high-compute environments. All numerical results above reflect only SUNTEK self-measured benchmarks on internal physically isolated test sets under fixed commercial thresholds; field performance varies with lighting, mounting geometry, and hardware setup—validate on target devices prior to deployment.
Validated on unseen out-of-distribution test pairs degraded by heavy compression, optical defocus, and non-linear exposure shifts.
Pairwise monotonic ranking loss eliminates traditional MSE score inversion under severe blur, reaching +0.8805 SRCC against international quality yardsticks.
Slashed anti-blue-light lens false sunglasses rejections from 21.96% to 1.40% while maintaining a consistent 98.93% true sunglasses block rate.
03GLOBAL REGULATORY COMPLIANCE & PRIVACY BY DESIGN
EU AI Act · GDPR · ISO/IEC 29794-5 (OFIQ) · ISO/IEC 30107-3 · PIPL
In multinational tenders and critical infrastructure deployments, algorithmic fairness, statutory quality standards, and zero-leakage privacy architecture are foundational.
Designed from inception around data minimization and non-discrimination: the architecture inherently contains zero demographic or sensitive-attribute inference modules, extracting only anonymized identity geometry. Balanced 24,000-pair audits show 99.80% mean accuracy with only 0.25% spread across groups.
Natively embeds ISO/IEC 29794-5 OFIQ Unified Quality Score (UQS) alignment inside the 21-D assessment module, and validates dual-spectrum liveness under ISO/IEC 30107-3 APCER/BPCER protocols at fixed thresholds.
Transforms camera frames immediately into one-way irreversible 512-D hypersphere vectors in "template-only, zero raw image retention" mode, protected by on-device encrypted storage and hardware-isolated access controls.
All 7 neural networks execute 100% locally on the terminal NPU with zero external cloud calls or biometric data transmission—ready for air-gapped government, defense, finance, and industrial networks.
Enforces strict hash-based deduplication and physical isolation between training corpora and blind-test suites, paired with signed model pack verification and full version lifecycle governance.
Unified RK3576 NPU architecture combining face detection, 98-pt landmark alignment, wearing & ISO 29794-5 quality gating, RGB + 850nm NIR liveness, and adaptive subspace matching.
Explore algorithms & modelsCo-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 & solutionsUnified 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