💻 computer science

Topology-Aware Global-Local Mamba Networks for Palm Vein Biometrics

The paper proposes a topology-aware global-local Mamba network that integrates multi-scale local features, structure-guided edge priors, and four-directional state-space scanning to achieve state-of-the-art palm vein recognition accuracy and low error rates with minimal parameters on public datasets.

Original authors: Zhengxi Wu, Felix Marattukalam, Waleed H. Abdulla

Published 2026-08-11

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhengxi Wu, Felix Marattukalam, Waleed H. Abdulla

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to unlock a secret door, but instead of a key, you need to prove who you are. This is the world of biometrics, the science of using your body's unique traits—like your fingerprint or face—to identify you. While we often think of fingerprints, there is a hidden layer of identity just beneath your skin: your veins. These tiny, winding rivers of blood form a unique map inside your palm that is invisible to the naked eye unless you shine a special near-infrared light on it. Because these maps are deep inside your body, they are incredibly hard to fake or steal, making them a super-secure way to prove you are "you." However, reading these maps is tricky. The veins are thin, faint, and twisty, like delicate threads in a foggy forest. To recognize them, computers need to be smart enough to see both the tiny texture of a single thread and the big picture of how the whole forest connects.

This is where a new study comes in, tackling the challenge of teaching computers to read palm veins with superhuman precision. The researchers, working with data from two public palm-vein datasets, noticed that while some computer models were good at seeing the "big picture" and others were good at seeing "tiny details," they often missed the specific shape of the veins themselves. To fix this, they built a new type of digital brain called a Topology-Aware Global-Local Mamba Network. Think of this network as a team of three specialized detectives working together. The first detective uses a magnifying glass to study the rough texture of the skin and veins. The second detective carries a special "edge-detecting flashlight" (based on a fixed mathematical rule called a Sobel operator) that highlights the exact lines and directions of the veins, ignoring the blurry background. The third detective is a master navigator who uses a "four-direction scan" to trace the entire path of the veins from every angle, understanding how they connect over long distances.

The magic happens when these three detectives share their notes. The team designed a system where the texture and direction clues are combined first, creating a "structurally aware" map, before being merged with the long-distance navigation map. This staged teamwork ensures the computer doesn't get lost in the details or miss the big connections. When they tested this new system, the results were impressive. On one dataset, it achieved a 99.13% accuracy rate in identifying people and an incredibly low error rate of 0.08%. On another dataset, it reached 92.42% accuracy and a 0.61% error rate. While another model called GLVM was slightly faster and better at pure identification, this new method was the most accurate at verification (proving a specific match) and used the fewest computer resources (only 7.2 million parameters) to get there. The researchers suggest that by explicitly teaching the computer to respect the line-like, directional nature of veins, they created a system that is exceptionally good at spotting the unique "signature" of a palm, even when the images are noisy or imperfect.

Technical Summary: Topology-Aware Global-Local Mamba Networks for Palm Vein Biometrics

Problem Statement
Palm-vein recognition is a fine-grained biometric task characterized by thin, low-contrast vascular structures that are observable only under near-infrared (NIR) illumination. While these features offer robustness against spoofing and surface contamination, effective recognition is challenging due to limited public datasets, intra-class variations (pose, noise, illumination), and the need to model both local vessel textures and the global connectivity of the vascular tree. Existing deep learning approaches face specific limitations: Convolutional Neural Networks (CNNs) require aggressive downsampling to capture long-range topology, which suppresses thin vessel structures; Transformers and self-attention mechanisms are computationally expensive for high-resolution imagery and may dilute sparse vascular patterns; and existing State-Space Models (SSMs) like Vision Mamba (Vim) and Global-Local Vision Mamba (GLVM) lack explicit encoding of the line-like, direction-sensitive geometry inherent to vascular structures.

Methodology
The authors propose a Topology-Aware Global-Local Mamba Network, a backbone architecture designed to jointly model local texture and global vascular topology. The network processes single-channel NIR images ( $128 \times 128$ ) through a stem, a body of six identical Topology-Aware Blocks (TABs), and a classification head.

Stem: Two strided $3 \times 3$ Conv-BatchNorm-SiLU layers reduce resolution from $128 \times 128$ to $32 \times 32$ and expand channels to $D=256$ .
Topology-Aware Blocks (TAB): Each block preserves spatial resolution and consists of three parallel branches and two stages of gated fusion:
- Multi-Scale Local Stream: Uses depthwise convolutions with $3 \times 3$ and $5 \times 5$ square kernels to extract isotropic vessel texture.
- Structure-Aware Directional Stream: Employs axis-aligned strip convolutions ( $1 \times 5$ and $5 \times 1$ ) to capture directional responses. Crucially, this stream is augmented by a fixed Sobel-magnitude edge prior. This prior is computed in feature space (on the first channel of the stem output) using fixed Sobel kernels ( $G_x, G_y$ ) to generate an edge map $E(u)$ . This map is projected via a learnable $3 \times 3$ convolution and added element-wise to the directional stream. This provides a structural inductive bias without learnable parameters in the prior itself.
- Global State-Space Stream: Implements a four-direction state-space scan (Left-Right, Right-Left, Top-Bottom, Bottom-Top) adapted from VMamba. It uses compact learnable per-channel scalars for discretization to suit smaller datasets, providing linear complexity while propagating information along scan directions aligned with vascular routing.
Staged Gated Fusion:
- Stage 1: The multi-scale and structure-augmented directional streams are fused via a learnable gating mechanism to produce a structurally aware local representation.
- Stage 2: This local representation is fused with the global state-space output. This ordering ensures the global scan operates on features already enriched with structural cues.
Residual Updates: The fused output passes through two residual connections with stochastic depth (DropPath), followed by a pointwise Feed-Forward Network (MLP ratio $r=2$ ).

Key Contributions
The paper identifies three primary contributions:

Explicit Structural Prior: The injection of a parameter-free, differentiable Sobel-magnitude edge map computed in feature space. Unlike previous models that leave structural geometry implicit, this prior explicitly marks vessel boundaries, guiding the directional and global branches.
Direction-Sensitive Local Encoding: A novel local block pairing multi-scale isotropic convolutions with axis-aligned strip convolutions, specifically designed to match the line-like, sparse geometry of palm veins.
Staged Fusion Strategy: A specific composition where local structural features are merged before the global scan. This ensures the four-direction state-space scan processes a representation that already carries structural cues, rather than unstructured local features.

Experimental Results
The method was evaluated on two public benchmarks: HKPU-NIR and VERA Palm Vein, under a unified training protocol. It was compared against ResNet50, Vim-S, ViT-S, and GLVM.

HKPU-NIR: Achieved 99.13% top-1 accuracy and 0.08% Equal Error Rate (EER) with 7.2M parameters.
VERA Palm Vein: Achieved 92.42% top-1 accuracy and 0.61% EER.
Comparative Performance: The proposed method achieved the lowest EER among all compared models on both datasets while maintaining the smallest parameter count (7.2M).
Trade-offs: While GLVM achieved slightly higher top-1 accuracy (99.40% on HKPU-NIR) and significantly lower computational cost (0.59 G FLOPs vs. 7.28 G FLOPs for the proposed method), the proposed architecture prioritized verification reliability (lowest EER). The higher FLOPs are attributed to the dense multi-branch blocks operating at full $32 \times 32$ resolution without downsampling.

Significance and Claims
The paper claims that the proposed architecture successfully addresses the challenge of modeling both local texture and global topology in palm-vein biometrics. By integrating a fixed structural prior with state-space modeling, the network achieves superior verification performance (lowest EER) with fewer parameters than existing CNN or Transformer-based baselines.

The authors acknowledge limitations: the structural prior is a fixed first-order edge operator (not learnable), the comparison relies on baselines trained under slightly different pipelines (though data partitions are identical), and the high computational cost restricts the design to scenarios where verification reliability outweighs inference budget. Future work is proposed to explore learnable connectivity-aware priors and leaner global operators.

Technical Summary: Topology-Aware Global-Local Mamba Networks for Palm Vein Biometrics

More like this