A hybrid finger identification pattern using Polarized depth-Weighted Binary Direction Coding

Wenming Yang, Wenyang Ji, Jing-Hao Xue, Yong Ren, Qingmin LiaoNeurocomputing 2019 · DOI · OpenReview

TL;DR

This paper introduces a novel hybrid finger identification system that captures both finger vein and dorsal texture patterns in a single image using a single camera. They propose a new feature extraction method, Polarized depth-Weighted Binary Direction Coding (PWBDC), which effectively combines information from both modalities, achieving superior identification accuracy compared to existing methods on both a new hybrid dataset and a public finger vein dataset.

What Problem Does This Paper Address?

Traditional biometric systems often rely on a single modality, such as finger vein patterns or surface textures. While finger veins are difficult to spoof, they can be challenging to capture clearly, especially for individuals with faint vein patterns. Conversely, surface textures are easier to capture but are more susceptible to environmental changes and manipulation. Existing methods that attempt to fuse information from multiple modalities often require separate imaging devices, increasing cost and complexity. This paper addresses the need for a more robust and cost-effective finger identification system by proposing a method to capture and process a hybrid pattern of dorsal finger vein and texture simultaneously from a single image.

Methods

The core innovation lies in the Polarized depth-Weighted Binary Direction Coding (PWBDC) method. This approach tackles the challenge of extracting features from a single image containing both vein and texture information. PWBDC has three key components:

  1. Polarized Direction Extraction: Instead of focusing on a single dominant direction (as in traditional methods like LDC), PWBDC identifies the two deepest directions at each pixel. This allows it to capture information from both the vein (which absorbs light) and the dorsal texture (which reflects light differently due to its sunken nature).
  2. Extended Normalized Angular Binary Coding: To represent these two directions, a novel binary coding scheme is introduced. This scheme is an extension of existing angular binary coding methods, designed to handle eight directions and ensure that adjacent orientations have smaller binary differences, leading to more accurate matching.
  3. Self-adaptive Depth-based Weighting: The contributions of the deepest and second deepest directions to the final matching score are not treated equally. A weighting mechanism is employed, where the depth of the vein and texture at a pixel is used to adaptively determine the importance of each direction's feature. Deeper features are given more weight, reflecting their greater discriminative power.

The system also utilizes a custom-designed image acquisition device that uses near-infrared light to penetrate the finger and capture both vein and texture patterns simultaneously.

Datasets

  • Newly Established Hybrid Finger Image Database: The authors created a new database containing 210 independent finger samples. This database captures the hybrid pattern of dorsal finger vein and texture. Samples were collected from both male and female volunteers over two sessions (training and testing).
  • Traditional Finger Vein Database: For comparative purposes, they also used a traditional database of finger vein images from the same 210 finger samples.
  • HKPU-FV Database: The proposed method was also evaluated on a popular public database of traditional finger vein images to demonstrate its generalizability.

Key Insights and Ideas

  • Synergy of Vein and Texture: The paper strongly suggests that combining finger vein and dorsal texture information is crucial for robust identification, especially when one modality might be weak or difficult to capture.
  • Two Deepest Directions are Better Than One: By considering the two most prominent directions at each pixel, PWBDC captures richer information than methods that only focus on the single deepest direction, effectively leveraging both vein absorption and texture reflection.
  • Adaptive Weighting for Feature Fusion: The self-adaptive depth-based weighting is a critical innovation. It intelligently fuses information from the two directions based on their actual depth, acknowledging that not all features are equally informative.
  • Single Camera, Multi-modal Advantage: The development of a single-camera system for capturing both vein and texture patterns significantly reduces hardware costs and system complexity compared to traditional multi-modal approaches.

Key Results

  • Superior Performance on Hybrid Data: On their newly established hybrid database, PWBDC achieved an accuracy of 98.89% with an EER of 1.3%. This significantly outperformed baseline methods and even other advanced techniques like PDC and PWDC.
  • Strong Performance on Traditional Data: PWBDC also demonstrated excellent performance on the traditional finger vein database, achieving 96.38% accuracy and an EER of 2.5%, outperforming most other methods.
  • Validation on Public Dataset: The method also showed competitive results on the HKPU-FV public dataset, outperforming several established methods and even deep learning-based approaches like DeepVein and CNN+FCN in terms of EER.
  • Component Contribution: Experiments showed that each component of PWBDC (polarized direction extraction, extended angular binary coding, and adaptive weighting) contributed to the overall performance improvement.

Limitations and Future Work

  • Extreme Conditions: The authors acknowledge that failed identification cases can still occur under extreme conditions, such as significant light variations or large finger rotations.
  • Future Work:
    • Expanding the established hybrid database.
  • Developing enhanced preprocessing methods to handle extreme conditions.
  • Exploring the fusion of additional modalities (e.g., fingerprint, knuckle texture) for even more reliable identification.
  • Investigating extensions of other advanced feature extraction methods (like multi-orientation weighted symmetric local graph structure) to their hybrid finger pattern.

Summary updated

Abstract

Highlights•Develop a new finger identification system to synchronously capture a two-modal image of dorsal finger vein and texture by a single camera.•Establish a corresponding two-modal hybrid images database of 210 finger samples.•Present a feature extraction method named by “Polarized depth-Weighted Binary Direction Coding”, which explores features of two modalities simultaneously.
BibTeX
@article{yang2019hybrid,
  author = {Wenming Yang and Wenyang Ji and Jing-Hao Xue and Yong Ren and Qingmin Liao},
  title = {{A hybrid finger identification pattern using Polarized depth-Weighted Binary Direction Coding}},
  journal = {Neurocomputing},
  year = {2019},
  doi = {10.1016/j.neucom.2018.10.042},
}
Figure