From Noise to Feature: Exploiting Intensity Distribution as a Novel Soft Biometric Trait for Finger Vein Recognition

Wenxiong Kang, Yuting Lu, Dejian Li, Wei Jia

TL;DR

This paper introduces a novel approach to finger vein recognition by treating the intensity distribution in the background of vein images, previously considered noise, as a valuable soft biometric trait. By fusing this new trait with traditional vein features using a hybrid matching strategy, the authors achieve improved recognition accuracy and robustness across multiple public datasets.

What Problem Does This Paper Address?

Biometric authentication systems are crucial for security in our increasingly digital world. Finger vein recognition is a promising biometric technology due to its resistance to tampering and ability to identify living individuals. However, existing methods primarily focus on the vein texture itself, often discarding or treating the background intensity variations as noise. This paper argues that these background intensity patterns, influenced by individual finger structures (bone, tissue thickness, water content), contain discriminative information that can enhance recognition performance. The challenge is to effectively extract and utilize this "noisy" information as a complementary biometric trait.

Methods

The core of this work lies in treating the background intensity distribution as a soft biometric. The authors propose a two-stage process:

  1. Background Layer Extraction: To isolate the intensity distribution, they first separate the finger vein image into a foreground (vein texture) and a background layer. Two methods are explored:
  2. Image Layer Separation (ILS): A smoothness-based algorithm that decomposes the image into smooth and high-gradient layers.
  3. Gaussian Blur (GB): A simpler, more intuitive low-pass filter that smooths out the high-frequency vein texture, leaving the low-frequency background intensity. GB is found to be more efficient and equally effective, becoming the preferred method.
  4. Soft Biometric Trait Extraction: Three methods are proposed to represent the extracted background intensity distribution:
  5. Mean and Variance (M&V): Simple statistical measures of the background layer.
  6. Array of Mean and Variance (AM&V): The background image is divided into blocks, and the mean and variance of each block are concatenated to capture spatial variations.
  7. Histogram of Spatial Pyramid (HSP): A multiscale approach where the image is partitioned into grids, and histograms of image blocks are computed and concatenated.
  8. Hybrid Matching Strategy: Recognizing that primary vein features (high-dimensional) and soft biometric traits (low-dimensional) have different characteristics, a hybrid matching strategy is employed. The primary biometric trait is matched using Support Vector Machines (SVM), which are well-suited for high-dimensional data. The soft biometric trait is matched using Manhattan Distance, which is more appropriate for low-dimensional features. The scores from these two matchers are then normalized and fused using a weighted sum rule to produce the final matching score.

Datasets

The authors evaluated their method on three publicly available finger vein databases:

  • FV-USM Database: Captured from 123 subjects, with 4 finger veins per subject and 6 samples per finger per session.
  • MMCBNU_6000 Finger Vein Database: Captured from 100 subjects, with 6 finger veins per subject and 10 samples per finger.
  • SDUMLA-HMT Finger Vein Database: Captured from 106 subjects, with 6 finger veins per subject and 6 samples per finger.

Key Insights and Ideas

  • "Noise" as a Feature: The most significant insight is the re-framing of background intensity variations, previously dismissed as noise, into a valuable soft biometric trait. This highlights the potential for extracting useful information from overlooked aspects of biometric data.
  • Complementary Information: The intensity distribution provides complementary information to the vein texture, leading to improved discrimination power when fused.
  • Robustness of AM&V: Among the proposed soft biometric traits, AM&V demonstrated greater robustness to parameter variations (specifically, the sigma value in Gaussian blur) compared to M&V and HSP.
  • Hybrid Matching Necessity: Directly concatenating features from primary and soft biometrics can dilute the impact of the soft trait. A score-level fusion strategy, using different matching methods tailored to each trait's dimensionality, is crucial for effective integration.

Key Results

  • Improved Accuracy: Fusion of the proposed soft biometric traits with primary biometric traits consistently outperformed using primary traits alone across all three databases.
  • Gaussian Blur Efficiency: The Gaussian Blur method for background extraction was significantly faster (4.58 ms vs. 120.54 ms for ILS) while achieving comparable or better performance.
  • Hybrid Matching Superiority: The hybrid matching strategy (HM) achieved significantly lower Equal Error Rates (EER) compared to a single matching strategy (SM) that used Manhattan Distance for both primary and soft traits. For example, with LBP+AM&V, HM achieved an EER of 0.216 on FV-USM, compared to 0.257 for SM.
  • State-of-the-Art Performance: The proposed method achieved competitive or superior EERs compared to several state-of-the-art algorithms on the evaluated databases, particularly on the FV-USM database (0.216 EER).

Limitations and Future Work

  • Limited Primary Traits: The paper acknowledges that only classical texture extraction algorithms (LBP, WLD, HOG, SIFT) were used as primary biometric traits. The authors suggest that using more advanced primary biometric traits could lead to even better performance.
  • Preliminary Study: The authors consider this work a preliminary study of the intensity distribution as a soft biometric trait.
  • Future Directions: Future work will focus on further understanding finger vein patterns and developing more advanced features to enhance recognition performance. This could involve exploring more sophisticated methods for extracting and representing the intensity distribution or investigating other potential soft biometric traits within finger vein images.

Summary updated

Abstract

Most finger vein feature extraction algorithms achieve satisfactory performance due to their texture representation abilities, despite simultaneously ignoring the intensity distribution that is formed by the finger tissue, and in some cases, processing it as background noise. In this paper, we exploit this kind of noise as a novel soft biometric trait for achieving better finger vein recognition performance. First, a detailed analysis of the finger vein imaging principle and the characteristics of the image are presented to show that the intensity distribution that is formed by the finger tissue in the background can be extracted as a soft biometric trait for recognition. Then, two finger vein background layer extraction algorithms and three soft biometric trait extraction algorithms are proposed for intensity distribution feature extraction. Finally, a hybrid matching strategy is proposed to solve the issue of dimension difference between the primary and soft biometric traits on the score level. A series of rigorous contrast experiments on three open-access databases demonstrates that our proposed method is feasible and effective for finger vein recognition.
BibTeX
@article{kang2019from,
  author = {Wenxiong Kang and Yuting Lu and Dejian Li and Wei Jia},
  title = {{From Noise to Feature: Exploiting Intensity Distribution as a Novel Soft Biometric Trait for Finger Vein Recognition}},
  journal = {IEEE Trans. IFS},
  year = {2019},
  doi = {10.1109/tifs.2018.2866330},
}
Figure