Palm Recognition¶
One-line summary: Biometric identification using the principal lines, wrinkles, ridge patterns, and vein networks of the human palm — increasingly deployed in contactless retail payment systems.
Modality: Palm (palmprint + palm vein)
Related concepts: Fingerprint Recognition, Deep Learning Architectures for Biometrics, Anti Spoofing Techniques, Real World Biometric Deployments, Multimodal Biometrics
Last updated: 2026-04-04
Overview¶
Palm recognition encompasses two related but distinct sub-modalities:
- Palmprint — Surface-level features captured in visible or NIR light: principal lines (heart, head, life), wrinkles, and minutiae-like ridge details at high resolution.
- Palm vein — Subcutaneous vein patterns captured by NIR illumination (850–940 nm); veins absorb NIR and appear dark. Harder to spoof because features are internal.
Pipeline¶
- Acquisition — Contactless (hovering hand over camera) or contact (pegged/pegless scanners). NIR for vein; visible/NIR for palmprint.
- ROI extraction — Locate the palm region using finger-valley landmarks; crop and align to a canonical square.
- Feature extraction — Gabor filters, competitive coding, ordinal measures (classical); CNN/ViT embeddings (modern).
- Matching — Hamming distance on binary codes or cosine similarity on deep embeddings.
Technical Details¶
Classical Methods¶
- CompCode (Competitive Coding, Kong & Zhang, 2004) — 2D Gabor filters at 6 orientations; winner-take-all encoding into a 3-bit code per pixel. Matching via bitwise AND + Hamming distance.
- Ordinal Code (Sun et al., 2005) — Encodes relative magnitude of Gabor responses. Robust to illumination changes.
- RLOC (Robust Line Orientation Code) — Improved line detection with noise robustness.
- PalmCode (Zhang et al., 2003) — Single Gabor filter per pixel, encoding phase. Simpler but less discriminative than CompCode.
Deep Learning Approaches¶
| Model | Year | Approach |
|---|---|---|
| PalmNet (Genovese et al.) | 2019 | CNN + triplet loss for contactless palmprint |
| ArcPalm | 2022 | ArcFace-style margin loss on palm embeddings |
| CMOS-Palm (Amazon One) | 2022 | Proprietary system; contactless, NIR-based, enrolled via hover |
| PalmViT | 2024 | Vision Transformer for palmprint; outperforms CNN baselines on cross-session matching |
| MediaPipe Hands | 2020+ | 21-landmark hand skeleton; useful for ROI extraction and palm alignment |
Palm Vein Recognition¶
- NIR imaging reveals deoxyhemoglobin-absorbing vein networks unique to each individual.
- Vein patterns are stable across years and difficult to forge (internal structure).
- Fujitsu PalmSecure is the major commercial system; bank ATMs in Japan and Brazil.
- Deep vein matching using U-Net segmentation + CNN embeddings achieving EER < 0.01% in constrained settings.
Amazon One (Retail Deployment)¶
- Launched 2020; deployed at 500+ Whole Foods and Amazon Go stores by 2025.
- Contactless NIR palm scanner; creates a "palm signature" from surface + subsurface features.
- Enrollment: hover palm twice; verification in ~1 second.
- Privacy: palm data stored encrypted in AWS cloud; users can request deletion.
- See Real World Biometric Deployments.
Datasets¶
| Dataset | Size | Type | Notes |
|---|---|---|---|
| PolyU Palmprint (v2) | 7.7K images / 386 hands | Contact, grayscale | Standard academic benchmark |
| IITD Palmprint | 2.6K images / 460 hands | Contactless | Indian institute benchmark |
| CASIA Palmprint | 5.5K images / 312 subjects | Contact, NIR + visible | Multi-spectral |
| Tongji Contactless Palmprint | 12K images / 600 hands | Contactless | Large-scale, 2 sessions |
| VERA Palm Vein | 2.2K images / 110 subjects | NIR vein | Vein recognition benchmark |
| PolyU Multi-Spectral | 6K images / 250 hands | Multi-spectral (4 bands) | Blue, green, red, NIR |
Challenges¶
- Contactless ROI consistency — Hand pose variation makes ROI extraction less repeatable than contact scanners.
- Cross-session variation — Skin condition, hand dryness, minor injuries change surface appearance.
- Spoof resistance — Palmprint can be attacked with printed images; palm vein is more robust but not immune (warm water balloon + printed vein pattern). See Anti Spoofing Techniques.
- Resolution trade-off — Consumer cameras capture principal lines easily but not minutiae-level ridge detail.
- Privacy concerns — Amazon One and similar systems raise questions about biometric consent in retail. See Privacy Preserving Biometrics.
State of the Art (SOTA)¶
As of early 2026: - Contact palmprint (PolyU): EER < 0.01% (CompCode, deep methods). - Contactless palmprint (Tongji): EER 0.1–0.5% (PalmViT, ArcPalm). - Palm vein (VERA): EER < 0.05% in constrained settings. - Amazon One (operational): <1 second latency; claimed FAR < 1 in 1M. - Cross-session contactless: EER 0.5–2.0% depending on time gap and conditions.
Open Questions¶
- Can contactless palmprint achieve parity with contact methods using only phone cameras?
- How to standardize palm biometric templates across vendors (ISO/IEC 19795-9)?
- Will palm replace fingerprint for retail payments due to hygiene and contactless preference post-COVID?
- Can palm vein + palmprint fusion be performed with a single NIR sensor in consumer hardware?
References¶
- Kong, A. & Zhang, D. (2004). Competitive Coding Scheme for Palmprint Verification. ICPR.
- Zhang, D. et al. (2003). Online Palmprint Identification. IEEE TPAMI.
- Amazon One. https://one.amazon.com/
Backlinks: Fingerprint Recognition, Deep Learning Architectures for Biometrics, Anti Spoofing Techniques, Real World Biometric Deployments, Multimodal Biometrics, Privacy Preserving Biometrics