Novel Local Coding Algorithm for Finger Multimodal Feature Description and Recognition

Shuyi Li, Haigang Zhang, Yihua Shi, Jinfeng Yang2019 · DOI

TL;DR

This paper introduces a novel local coding algorithm, Gabor-GSLGS, for finger multimodal biometrics (fingerprint, finger-vein, and finger-knuckle-print). It addresses challenges like pose variation and illumination inconsistencies by combining Gabor filters for enhancement with a generalized symmetric local graph structure for robust feature extraction, achieving improved accuracy and efficiency.

What Problem Does This Paper Address?

Biometric identification is increasingly important, and finger-based traits like fingerprints, finger-veins, and finger-knuckle-prints offer advantages in convenience and user-friendliness. However, relying on a single modality (unimodal biometrics) has security risks. Fusing these three traits (multimodal biometrics) promises higher accuracy. The main challenges in finger multimodal recognition are:

  • Pose Variation: Fingers can rotate during image capture, leading to image mismatch and reduced recognition accuracy.
  • Illumination Inconsistency: Variations in lighting can degrade image quality and make feature extraction unreliable.
  • Robust Feature Representation: Developing methods that can effectively capture and represent the unique features of each modality while being resilient to these variations is crucial.

Existing local coding methods, while useful, often struggle with these issues, either by only considering fixed neighborhoods or by having imbalanced feature representations due to asymmetric pixel weighting.

Methods

The proposed approach, Gabor-GSLGS, tackles these problems in several steps:

  1. Pose Correction: A novel algorithm is introduced to correct for finger rotation. It detects the finger's edge lines, calculates the midpoint coordinates, and then uses these to determine the rotation angle. The entire finger image is then rotated to a standard orientation.
  2. Region of Interest (ROI) Extraction: Specific methods are used to extract relevant regions for each modality: core point detection for fingerprints, convex direction coding for finger-knuckle-prints, and an interphalangeal joint prior method for finger-veins. These ROIs are then standardized in size.
  3. Image Enhancement (Gabor Filters): A bank of k-channel even symmetric Gabor filters is applied to the finger ROIs. Gabor filters are known for their ability to extract texture information in multiple directions and reduce the impact of illumination.
  4. Novel Local Coding Algorithm (Gabor-GSLGS): This is the core contribution.
  5. It uses a Generalized Symmetric Local Graph Structure (GSLGS) operator. Unlike previous methods, GSLGS considers relationships between a target pixel and its neighbors in multiple orientations.
  6. Crucially, it assigns equal weights to symmetric pixels on the left and right sides of the neighborhood. This ensures a balanced feature representation, mitigating issues from asymmetric weighting in prior methods.
  7. The Gabor-enhanced images are processed by this GSLGS operator. For each pixel, the algorithm computes coded values based on pixel comparisons in different orientations.
  8. The final coded value for a pixel is the maximum among these orientation-specific coded values, inspired by the optimal response of Gabor filters.
  9. Feature Fusion and Matching:
    • The coded trimodal feature images are divided into non-overlapping blocks.
  10. Histograms are generated for each block to represent local features.
  11. These local histograms are concatenated to form a global feature vector for each finger.
  12. Matching is performed using the intersection coefficient between feature vectors.

Datasets

The authors created a homemade finger trimodal database for their experiments. This database consists of:

  • 17,550 images from 585 individuals.
  • Each individual contributed images from index, middle, and ring fingers of both hands.
  • Each finger provided 30 images (10 per modality: FP, FV, FKP).

For their experimental database, they randomly selected 3000 images from 100 individuals, with 10 images per modality for each individual.

Key Insights and Ideas

  • Symmetry is Key for Balanced Representation: The paper highlights that assigning equal weights to symmetric pixels in local neighborhood comparisons is crucial for achieving a balanced and robust feature representation, overcoming limitations of non-symmetric approaches.
  • Gabor Filters Enhance, GSLGS Codes: The synergy between Gabor filters for noise reduction and illumination invariance, and the GSLGS operator for capturing complex local relationships, is a powerful combination for multimodal biometrics.
  • Pose Correction is Essential: Even small rotations can significantly impact recognition. The proposed pose correction method is shown to be effective in standardizing finger orientation.
  • Multimodal Fusion Outperforms Unimodal: As expected, combining FP, FV, and FKP consistently yields better accuracy than any single modality, demonstrating the complementary nature of these features.
  • FV is the Dominant Unimodal Trait: Among the single modalities, finger-vein (FV) shows the best performance, suggesting it carries the most discriminative information on its own.

Key Results

  • The proposed Gabor-GSLGS algorithm achieved excellent performance in improving matching accuracy and recognition efficiency.
  • Experimental results showed that the proposed approach achieved the lowest Equal Error Rate (EER) compared to other feature extraction methods like LBP, GLBP, LLBP, SLGS, and MOW-LGS.
  • The Gabor-GSLGS method also demonstrated the lowest computation cost for feature extraction among the compared methods.
  • The optimal parameters identified were a neighborhood size of 5x5 and 8x8 image division blocks.
  • The multimodal combination (FP + FV + FKP) achieved the best EER of 0.022%, significantly outperforming unimodal traits (e.g., FP EER of 4.26%, FV EER of 0.19%, FKP EER of 0.40%).

Limitations and Future Work

  • Computational Efficiency: While the proposed method is efficient, the authors aim to further reduce the dimensions of the feature vector and improve the overall efficiency of multimodal recognition.
  • Robustness to Other Databases: The current experiments were conducted on a homemade database. Future work will involve applying the algorithm to other public biometric databases to assess its generalizability.
  • Advanced Fusion: The authors plan to explore more robust and effective fusion methods that can better integrate multiple modal features.
  • Generalizability: The paper focuses on finger-based biometrics. Extending the local coding algorithm to other biometric modalities or applications could be a future direction.

Summary updated

BibTeX
@misc{li2019novel,
  author = {Shuyi Li and Haigang Zhang and Yihua Shi and Jinfeng Yang},
  title = {{Novel Local Coding Algorithm for Finger Multimodal Feature Description and Recognition}},
  year = {2019},
  doi = {10.3390/s19092213},
}
Figure