Palmprint Decomposer: Interpretable and Editable Factorization via Synthetic Learning and Instance-Wise Refinement
Authors: Licheng Yan, Yuqi Wang, Jingtao Wang, Bob Zhang, Andrew Beng Jin Teoh, Lu Leng, Shuyi Li, Ziyuan Yang
Organizations: PAMI Research Group, Department of Computer and Information Science, University of Macau, Macau, China · School of Electrical and Electronic Engineering, College of Engineering, Yonsei University, Seoul 120749, Republic of Korea · School of Software, Nanchang Hangkong University, Nanchang, Jiangxi 330063, P. R. China · School of Cyber Science and Engineering, Sichuan University, Chengdu, Sichuan 610065, P. R. China
Palmprint recognition and synthesis methods focus on extracting discriminative textures or generating images with varied textures and appearances. However, they do not generally provide an explicit image-level representation that separates palmprint texture from its underlying base appearance for independent analysis or controlled editing. We introduce Palmprint Decomposer, an interpretable framework that represents a palmprint through independently editable factors and reconstructs the image through a multiplicative composition model. A color map and a light map characterize the base appearance, while a structure map and a global texture strength parameter describe texture modulation. However, accurate reconstruction alone does not guarantee separation of palmprint texture from base appearance, as fine palmprint textures may leak into the estimated base appearance. To mitigate this ambiguity, we train the decomposer exclusively on synthetic data with direct factor supervision, using asymmetric decoding paths and decomposition constraints to suppress texture leakage. Instance-Wise Refinement further improves real-image reconstruction by directly optimizing the predicted factors for each input without relying on the trained network. Experiments across seven datasets demonstrate high reconstruction fidelity, with the extracted structure maps generally achieving verification performance comparable to the original images. Compared with existing methods, our approach achieves better factor separation while preserving more discriminative palmprint texture information.
Figures & tables
Figure 1: Palmprint decomposition and recomposition. The color map C and light map L define base appearance B , while the structure map S and global texture strength a define texture modulation T . A multiplicative composition model combines B and T to reconstruct the input palmprint.
Figure 2: Overview of the proposed Palmprint Decomposer framework. During synthetic-only training, FD estimates four factors from synthetic images using asymmetric decoding and decomposition losses. For real palmprints, IWR directly optimizes the initial factor estimates to improve reconstruction fidelity.
Figure 3: Synthetic factor generation and composition. (a) Bézier curves are rasterized and converted into PCE responses, then augmented to generate structure maps. (b) Examples of synthetic factors and their composed images.
Reconstruction
Original similarity
Separation
Original
Structure Map S
Reconstruction
Dataset/subset
PSNR ↑
SSIM ↑
LPIPS ↓
Sim. ↑
Intra ↑ /Inter ↓
C/L/B–O ↓
EER ↓
d′↑
Δ EER ↓
Δd′↑
Δ EER ↓
Δd′↑
IITD
42.62
0.980
0.030
0.99
0.79 / 0.55
0.54 / 0.56 / 0.56
1.21
4.60
−0.05
−0.03
+0.02
−0.01
PolyU
48.38
0.999
0.003
0.99
0.78 / 0.54
0.56 / 0.55 / 0.57
0.10
5.67
+0.01
+0.01
0.00
+0.02
PolyU_CF
43.19
0.974
0.020
0.97
0.77 / 0.56
0.56 / 0.53 / 0.58
2.40
3.94
+0.03
−0.09
+0.07
−0.01
Tongji
42.84
0.981
0.032
0.98
0.77 / 0.55
0.55 / 0.55 / 0.58
1.11
4.07
−0.22
+0.06
−0.01
+0.01
Zhou
46.40
0.989
0.022
0.98
0.80 / 0.54
0.55 / 0.55 / 0.57
2.49
3.95
−0.33
+0.02
0.00
+0.02
Table 1: FD + IWR results across all datasets and subsets. Sim. denotes reconstruction-to-input texture similarity; Intra/Inter denotes mean same-/different-identity similarities of original images; C/L/B–O gives the texture similarity of each of C , L , and B to the original input. Original reports EER (%) and d′ , while Structure Map S and Reconstruction report changes relative to Original. All similarities are computed using MTCC. Green / red denotes improvement/degradation; changes rounding to zero are black.
Figure 4: Decomposition on real palmprints. Original images, final reconstructions, refined factors, and the derived base appearance B^=C^(1+L^) and texture modulation T^=a^S^ across datasets and subsets.
Figure 5: Independent factor editing. Each group shows an edited factor and the resulting recomposed image, with the other three factors fixed. Primes denote the edited factors and recomposed images.
Base appearance vs. input
Texture verification
Method
Hist. corr. ↑
B–O ↓
EER (%) ↓
d′↑
IITD
Gaussian
0.980
0.910
1.56
3.87
RTV [ 27 ]
0.826
0.788
1.33
4.35
Direct PCE [ 22 ]
N/A
N/A
6.32
3.02
VO [ 28 ]
N/A
N/A
2.59
3.26
Table 2: Comparison of factor separation and texture verification on IITD and XJTU-UP. B denotes the base appearance estimate (the filtered image for Gaussian and RTV), and O denotes the original input. Hist. corr. measures mean RGB-channel histogram correlation. N/A indicates that only the texture output is evaluated. Bold marks the best result per metric and dataset.
Figure 6: Visual comparison of base appearance estimates and texture outputs on IITD and XJTU-UP. For our method, B and T denote base appearance and texture modulation, respectively. Texture outputs are visualized with zero response mapped to mid-gray for consistent comparison. For Direct PCE and VO, only the texture outputs are shown.
Reconstruction
Separation
Structure Map S
Recon.
Variant
PSNR ↑
SSIM ↑
LPIPS ↓
Sim. ↑
C/L/B–O ↓
EER ↓ / d′↑
EER ↓ / d′↑
IITD
FD
38.45
0.971
0.081
0.96
0.49/ 0.53 / 0.52
1.20/ 4.61
1.26/4.59
w/o PCE
36.61
0.961
0.104
0.94
0.50/0.57/0.57
1.23/4.53
1.26/4.53
Full-res. Base
39.52
0.976
0.065
0.97
0.53/0.58/0.59
1.23/4.60
1.25/4.57
w/o Cons.
38.20
0.973
0.074
0.96
0.45 /0.60/0.58
1.23/4.59
1.26/4.55
Table 3: Ablations on the complete IITD and PolyU sets. Each training variant independently modifies FD. FD + IWR reuses the 100-step final results from Table 1 , which also defines the reconstruction metrics and C/L/B–O. Structure Map S and Recon. cells report EER (%)/ d′ . Bold marks the best displayed values within each dataset, including ties.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure S1: Factor Decomposer. The base appearance decoder operates at half resolution and omits the skip connection from the highest-resolution encoder stage. Its two heads predict the color map C and light map L , followed by bilinear upsampling. The texture modulation decoder retains skip connections at all encoder scales. Its spatial head predicts the structure map S , while global average pooling followed by a linear layer predicts the global texture strength a . RB denotes residual block, GN group normalization, and GAP global average pooling.
Dataset
Images
Palm identities
IITD
2,600
460
Tongji
12,000
600
PolyU
7,752
386
PolyU_CF
1,770
177
Zhou
1,295
166
XJTU-UP
7,880
200
Appendix
Table S1: Evaluated ROIs and palm identities. Each palm is counted once across acquisition conditions. Verification uses all unordered pairs of distinct images within each evaluation group.
Figure S2: Cross-dataset factor exchange. Left-column images supply base appearance and top-row images supply texture modulation. Each interior image combines the corresponding factors through the fixed composition model. Diagonal entries show self-reconstructions.
B / T
IITD
PolyU
PolyU_CF
Tongji
Zhou
XJTU- UP
CASIA
IITD
0.99/0.99
0.18/0.92
0.89/0.87
0.86/0.90
0.90/0.86
0.91/0.85
0.91/0.88
PolyU
0.25/0.82
0.99/0.99
0.31/0.82
0.34/0.85
0.25/0.79
0.16/0.77
0.24/0.82
PolyU_CF
0.90/0.86
0.49/0.91
0.98/0.97
0.90/0.87
0.90/0.83
0.87/0.82
0.89/0.85
Tongji
0.92/0.88
0.49/0.91
0.94/0.85
0.99/0.98
0.91/0.83
0.85/0.83
0.91/0.86
Zhou
0.91/0.88
0.26/0.92
0.89/0.86
0.86/0.89
1.00/0.98
0.90/0.84
0.90/0.87
XJTU-UP
0.87/0.90
0.01/0.93
0.76/0.87
0.68/0.90
0.83/0.87
0.99/0.98
0.83/0.89
Appendix
Table S2: Dataset-wide factor exchange. Rows supply base appearance and columns supply texture modulation. Cells report mean RGB histogram correlation to the base appearance donor / MTCC texture similarity to the texture donor (both ↑ ). Diagonal entries are self-reconstructions.
Figure S3: IWR on IITD and PolyU. Each input appears at left, followed by its base appearance B , texture modulation T=aS , and reconstruction residual. The residual is the per-pixel mean of ∣Iˉ−I∣ over RGB channels, shown on a common scale. (a) Four snapshots from a 200-step run. (b) The same inputs under the standard 100-step schedule with and without the overlap term. (c) Full-dataset metrics sampled every 20 updates along the runs in (a). Texture EER (%) is computed using MTCC on T=aS at each checkpoint.
Institute of Data Science, National University of Singapore, Singapore 117602, Singapore · Centre for Frontier AI Research (CFAR), A*STAR, Singapore, 138634 · School of Cyber Science and Engineering, Wuhan University, China, 430072 +2