Research map / Image generation / Editing and style transfer
PCE-Palm: Palm Crease Energy Based Two-Stage Realistic Pseudo-Palmprint Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PCE-Palm, a two-stage generator that maps Bézier curves to a palm crease energy (PCE) domain and then to realistic palmprints, plus a plug-and-play line feature enhancement block, to produce identity-controllable pseudo palmprints for recognition under few-ID settings.
Problem
Palmprint recognition is limited by scarce large-scale data; existing Bézier-curve-based pseudo palmprint generators (BézierPalm, RPG-Palm) show a large domain gap to real palmprints, degrading performance when few real IDs are available.
Approach
Decompose the Bézier-to-real gap into crease and texture differences by introducing a differentiable palm crease energy domain (via Gaussian-MFRAT line kernels) as an intermediate bridge; stage one translates Bézier curves to realistic creases (unpaired), stage two translates PCE images to textured palmprints (paired-like, with cycle PCE consistency), and an LFEB enhances line features.
Key results
- Under 40 real training IDs (1:1 open-set), PCE-Palm reaches TAR@FAR=1e-6 of 0.7694 vs 0.4796 for RPG-Palm (about 29% improvement) and exceeds ArcFace trained on all 1600 IDs by more than 6 points.
- Generated-image FID of 40.3 versus 188.9 for RPG-Palm.
- With ResNet50 at 1:3 open-set, TAR@FAR=1e-6 reaches 0.9271 compared with 0.8698 for RPG-Palm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PCE-Palm instead of RPG-Palm | MobileFaceNet palmprint recognition · Open-set, train:test = 1:3 | TAR@FAR=1e-6: 0.8485 -> 0.9079 | Helped |
| Use PCE-Palm instead of RPG-Palm | MobileFaceNet palmprint recognition · Open-set, 40 real training IDs | TAR@FAR=1e-6: 0.4796 -> 0.7694 | Helped |
| Add PCEM to the baseline generation model (ablation) | PCE-Palm · 40 real IDs, open-set train:test = 1:1 | TAR@FAR=1e-6: 0.4144 -> 0.5525 | Helped |
| Add data augmentation (ablation) | PCE-Palm with PCEM · 40 real IDs, open-set train:test = 1:1 | TAR@FAR=1e-6: 0.5525 -> 0.6077 | Helped |
| Use improved Bézier curve synthesis (ablation) | PCE-Palm with PCEM and data augmentation · 40 real IDs, open-set train:test = 1:1 | TAR@FAR=1e-6: 0.6077 -> 0.6385 | Helped |
| Add LFEB to the generation model (ablation) | PCE-Palm · 40 real IDs, open-set train:test = 1:1 | TAR@FAR=1e-6: 0.6385 -> 0.7060 | Helped |
| Add LFEB to the recognition model (ablation) | PCE-Palm · 40 real IDs, open-set train:test = 1:1 | TAR@FAR=1e-6: 0.7060 -> 0.7694 | Helped |
Surprising findings
- Pretraining on synthetic data with only 40 real IDs outperforms ArcFace trained on all 1600 real IDs at TAR@FAR=1e-6.
Limitations
- Relies on hand-designed Bézier templates derived from dermatoglyphics priors.
- Requires real palmprints to extract PCE images for stage-two training, so some real data is still needed.
- Evaluated only on palmprint datasets.
Experimental setup
Datasets: 13 public palmprint datasets (PolyU, etc.) BézierPalm synthetic Bézier curves
Benchmarks: 13 public palmprint datasets under open-set protocol (1:1 and 1:3 train:test)
Metrics: TAR@FAR FID
Baselines: BézierPalm RPG-Palm ArcFace C-LMCL pix2pixHD CycleGAN BicycleGAN CompCode LLDP BOCV RLOC DOC
Builds on: BézierPalm RPG-Palm MFRAT CycleGAN ArcFace
Closest papers by meaning
- Diff-Palm: Realistic Palmprint Generation with Polynomial Creases and Intra-Class Variation Controllable Diffusion Models (CVPR 2025 · similarity 0.91)
- PVTree: Realistic and Controllable Palm Vein Generation for Recognition Tasks (AAAI 2025 · similarity 0.84)
- Towards Explainable and Unprecedented Accuracy in Matching Challenging Finger Crease Patterns (CVPR 2025 · similarity 0.80)
- FoundHand: Large-Scale Domain-Specific Learning for Controllable Hand Image Generation (CVPR 2025 · similarity 0.78)
- RHanDS: Refining Malformed Hands for Generated Images with Decoupled Structure and Style Guidance (AAAI 2025 · similarity 0.77)
- HanDiffuser: Text-to-Image Generation With Realistic Hand Appearances (CVPR 2024 · similarity 0.77)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.