Open-Set Palmprint Verification

• Open-set palmprint verification is a biometric approach that distinguishes enrolled identities from unknown individuals using strict thresholding and robust feature embeddings.
• Synthetic pretraining via geometric, texture, and identity-aware generative models enhances generalization on unseen data while addressing data scarcity.
• Advanced matching protocols leveraging deep metric learning, angular margin losses, and feature-space quantization ensure ultra-low false accept rates in real-world applications.

Open-set palmprint verification is a biometric authentication task in which the system must determine whether a given probe palmprint corresponds to any identity in an enrolled gallery or belongs to an unseen (unknown) individual. Unlike closed-set scenarios, all test identities are disjoint from the training set, and the operational requirements emphasize extremely low false accept rates, high discrimination on unseen classes, and robust generalization in the presence of limited, privacy-constrained real data.

1. Problem Definition and Fundamental Challenges

In open-set palmprint verification, the core requirement is to correctly accept genuine matches among a gallery of enrolled subjects while reliably rejecting impostors—probe images from identities not present in the gallery or at training. This setting is operationalized by requiring, for each probe $x_q$, a binary decision: accept as some known enrolled person (assign an ID) or reject as "unknown" (Shen et al., 2023, Yang et al., 5 Mar 2025, Gao et al., 2024).

Key technical challenges:

• Unseen identity generalization: Models must avoid overfitting to seen classes and construct discriminative representations for genuinely novel palmprints.
• Ultra-low FAR: Real deployments (e.g., border, fintech authentication) demand FAR $\leq 10^{-6}$, where performance gains are scarce and data is often extremely imbalanced.
• Data scarcity: Real palmprint datasets are expensive and privacy-sensitive, severely limiting exposure to true intra- and inter-class diversity (Shen et al., 2023).

2. Synthesis-driven Large-Scale Pretraining

The dominant methodological advance in recent years has been the large-scale generation of synthetic palmprints for pretraining discriminative networks, effectively bridging the gap left by limited real datasets. Multiple synthesis paradigms are employed:

Geometric Synthesis with Bézier Curves

Each identity is uniquely defined by a set of principal lines and wrinkles, parameterized as 2D Bézier curves. Identity variability is achieved by:

• Randomly sampling control points for principal and wrinkle lines
• Imposing constraints on stroke width, color, and spatial arrangement
• Rendering palm images over natural-image backgrounds with optional blur (Zhao et al., 2022)

Empirical results show that pretraining with 4,000–20,000 synthetic identities and finetuning on real data produces up to 15 percentage point improvements in TAR@1e-6 over strong ArcFace baselines, with saturation proportional to the number of unique synthetic IDs rather than samples per ID (Zhao et al., 2022).

Texture- and Block-level Synthesis

Canny2Palm advances synthesis realism by:

• Affine-normalizing real palms to a canonical pose
• Extracting Canny edge maps for palm texture
• Generating synthetic images via Pix2Pix conditional GANs
• Assembling new identities via block-level (3×3 grid) recombination from multiple real ID textures, enforcing distinctness between virtual IDs

This approach yields 5–7.2 percentage point superior open-set identification rates over competing methods (e.g., BézierPalm, NS-PVDB) and avoids the performance saturation inherent to less-controllable generative models (Lan et al., 8 May 2025).

Generative Pseudo-data with Identity-aware Loss

RPG-Palm utilizes a U-Net generator, where physical crease maps and noise vectors encode the identity and intra-class variability. An explicit identity-aware loss ensures that intra-ID synthetic images have consistent embeddings, while enhanced pseudo-identity generation is enforced by similarity-minimizing constraints between different synthetic IDs (Shen et al., 2023).

On benchmarks with 1:1 and 1:3 train:test splits, synthetic pretraining with RPG-Palm yields +5% to +14% gain in TAR@FAR=1e-6 compared to BézierPalm, and even exceeds models trained on 100% real data with only 10% real-data fine-tuning (Shen et al., 2023).

Method	TAR@FAR=1e-6 (1:1)	TAR@FAR=1e-6 (1:3)
BézierPalm	84.37%	70.12%
RPG-Palm	89.46%	84.85%

3. Feature Representation and Matching Protocols

Verification frameworks increasingly utilize learned deep embeddings, often trained under proxy classification losses such as ArcFace:

• Training: Discriminative representation learning via large-scale (often synthetic) pretraining, then fine-tuning on limited real palmprints. Losses emphasize angular margin separation (e.g., ArcFace loss with $m=0.5$, $s=48$) (Zhao et al., 2022, Lan et al., 8 May 2025, Shen et al., 2023).
• Scoring: Cosine similarity in embedding space between probe and gallery templates.
• Decision: If $\max_i s_i \geq \tau$ (where $s_i$ is similarity to the $i$th enrolled ID, and $\tau$ is a threshold set for desired FAR), accept; otherwise, reject as "unknown" (Yang et al., 5 Mar 2025, Khan et al., 2014).

Thresholds are tuned to control FAR/EER, as open-set error separation is largely dictated by tail behaviors in the impostor score distribution.

4. Algorithmic Enhancements and Domain Generalization

Feature-Space Alignment via Vector Quantization

PalmBridge introduces post-hoc domain shift mitigation by learning a codebook $C=\{c_k\}$ in embedding space and mapping each feature vector to its nearest codeword:

• Blending: Final feature $\widetilde{f} = \alpha c_{k^*} + (1-\alpha)f$, $\alpha\in[0,1]$
• Objectives: Jointly optimize discrimination (task loss), feature/quantized-feature consistency, and codeword orthogonality
• Evaluation: Reduces open-set EER by 0.3–1.0% (absolute) across datasets and enhances cross-domain generalization (Zhang et al., 28 Jan 2026)

Assignment consistency ($p_{\text{same}}$) and collision rate ($p_{\text{collide}}$) quantify intra- and inter-class blending accuracy for unseen identities.

Federated and Privacy-preserving Learning

FedPalm employs distributed training with local personalization for heterogeneity and a global open-set expert aggregated by federated averaging:

• Interaction Module: Fuses locally-extracted textural features with those routed from the most similar peer experts
• Results: Achieves consistent open-set EER improvements (e.g., 4.27% on Tongji, 6.18% on IITD) and significant privacy guarantees, as raw biometrics never leave the client device (Yang et al., 5 Mar 2025)

5. Beyond Identity: Cross-Chirality and Multispectral Verification

The CCPV framework generalizes open-set verification to cross-chirality scenarios, enabling left–right palmprint matching with a single template per individual:

• Matching: Four-match rule (involving flipped and unflipped counterparts) and averaging of cosine distances over all permutations
• Loss: Aggregated cross-chirality contrastive term ensures that embeddings of all flips and hands per-ID cluster together
• Empirical gains: CCPV obtains [email protected]%FAR of up to 95.2% (1:1 protocol) and EER ≈ 0.15% under trans-domain splits (Gao et al., 2024)

6. Early Foundations: Contour Coding and Classical Pipelines

Prior to deep learning, robust orientation-based binary hash encoding (Contour Code) yielded highly compact palmprint representations. Matching was performed by score-thresholding, enabling both closed- and open-set rejection without additional score normalization. On benchmarks:

• EER=0.0030% (PolyU), GAR≅100% @ 0.1% FAR (ATM matching)
• Open-set extension is formulated as: reject probe if all max gallery scores $< \tau_{\text{open}}$ (from impostor distribution tail) (Khan et al., 2014)

7. Future Directions and Open Questions

While synthetic data generation has greatly expanded open-set palmprint generalization, several outstanding issues remain:

• Domain adaptation: Existing pipelines rely on simple fine-tuning; advanced unsupervised or domain-invariant approaches could potentially further close the synthetic–real domain gap (Zhao et al., 2022).
• Scaling of synthesis diversity: Performance gains saturate with sample count per ID; progress depends critically on increasing the effective number of high-diversity synthetic identities (Lan et al., 8 May 2025).
• Modality fusion: Initial advances in multispectral and vein-inclusive features, as in Contour Code, remain largely unexplored in the most recent deep synthetic and federated paradigms (Khan et al., 2014).

A plausible implication is that further open-set accuracy improvements will require synthesizing multimodal palm representations, dynamic threshold calibration for unknown-class generalization, and more universal codebook/product-quantization–based architectures for embedding alignment.

---

References:

• "Geometric Synthesis: A Free lunch for Large-scale Palmprint Recognition Model Pretraining" (Zhao et al., 2022)
• "Canny2Palm: Realistic and Controllable Palmprint Generation for Large-scale Pre-training" (Lan et al., 8 May 2025)
• "RPG-Palm: Realistic Pseudo-data Generation for Palmprint Recognition" (Shen et al., 2023)
• "PalmBridge: A Plug-and-Play Feature Alignment Framework for Open-Set Palmprint Verification" (Zhang et al., 28 Jan 2026)
• "Cross-Chirality Palmprint Verification: Left is Right for the Right Palmprint" (Gao et al., 2024)
• "FedPalm: A General Federated Learning Framework for Closed- and Open-Set Palmprint Verification" (Yang et al., 5 Mar 2025)
• "Multispectral Palmprint Encoding and Recognition" (Khan et al., 2014)