Palmprint Recognition Based on Local Fisher Discriminant Analysis

Jinyu Guo, Haibin Chen, Yuan LiJ. Softw 2014 · DOI

TL;DR

This paper introduces a novel palmprint recognition method called Local Fisher Discriminant Analysis (LFDA). By considering local data structures, LFDA significantly outperforms existing methods like PCA, FDA, ICA, and KPCA, achieving a remarkable 98.95% recognition rate with a fast 0.031s recognition time, making it suitable for real-time applications.

What Problem Does This Paper Address?

Biometric recognition systems, particularly palmprint recognition, are crucial for security and identification. While various feature extraction methods exist, many struggle with the "small-size-sample" problem, where the number of samples is limited compared to the dimensionality of the data. This can lead to singularity issues in eigenvalue equations. Furthermore, existing methods often focus on global features, potentially missing important local patterns within the palmprint, which can be affected by variations in illumination and image quality. The paper aims to address these limitations by developing a more robust and accurate palmprint recognition technique.

Methods

The core of the proposed method is the Local Fisher Discriminant Analysis (LFDA). LFDA builds upon the principles of Fisher Discriminant Analysis (FDA) but incorporates local information.

Here's a breakdown of the key steps:

  1. Dimensionality Reduction: To combat the small-size-sample problem, palmprint images are first down-sampled to reduce their dimensionality.
  2. Local Fisher Discriminant Analysis (LFDA):
    • LFDA introduces weighting matrices ($W_b$ and $W_w$) into the between-class and within-class scatter matrices. These matrices are designed to capture the local relationships between data points.
  3. An affinity matrix ($A_{ij}$) is used to quantify the "closeness" of data pairs, with a standard exponential decay function based on Euclidean distance.
  4. The method then performs a generalized eigenvalue decomposition on these locally weighted scatter matrices to find projection vectors that maximize class separation while minimizing within-class scatter, considering local structures.
  5. The resulting generalized eigenvectors are weighted by the square root of their corresponding eigenvalues to form the transformation matrix ($T_{LFDA}$).
  6. Feature Extraction: Palmprint images are projected onto the derived LFDA vectors to obtain local palmprint feature vectors.
  7. Matching: Cosine distance is used to compare the feature vectors of two palmprints for matching.

Datasets

The PolyU palmprint database was used for training and evaluation. This database contains 600 grayscale images from 100 different palms, with 6 samples per palm collected in two sessions. The authors used 4 images per palm for training and the remaining 2 for testing.

Key Insights and Ideas

  • Local Structure Matters: The paper strongly emphasizes that considering local data structures is critical for robust palmprint recognition, especially when dealing with variations in illumination and image quality.
  • LFDA's Advantage: By incorporating local relationships through weighting matrices and affinity, LFDA effectively handles multimodal data (classes with multiple clusters) and achieves better class separation than global methods.
  • Addressing Singularity: Down-sampling is a practical approach to mitigate the singularity issues often encountered in eigenvalue decomposition with small sample sizes.
  • Trade-off in Speed: While LFDA offers superior accuracy, its feature extraction process is slightly more computationally intensive than simpler methods like FDA, but the overall recognition time remains very fast.

Key Results

  • Highest Recognition Rate: LFDA achieved the highest recognition rate among all tested algorithms, reaching 98.95%.
  • Superiority over Baselines: LFDA significantly outperformed PCA (91.51% at $\theta=0.95$), FDA (91.95% at $\theta=0.95$), ICA (92.66% at $\theta=0.95$), and KPCA (94.33% at $\theta=0.95$).
  • Fast Recognition Time: The proposed LFDA algorithm boasts a recognition time of just 0.031s, meeting real-time system specifications.
  • Feature Number Impact: The recognition rate varied with the number of selected features (controlled by $\theta$), with LFDA showing consistently high performance across different feature sets.

Limitations and Future Work

  • Feature Extraction Time: The paper acknowledges that LFDA's feature extraction time is longer than FDA due to the introduction of weighting matrices and the calculation of affinity matrices.
  • Future Work: Not explicitly detailed in the provided text, but the success of LFDA suggests potential for its application in other biometric modalities or pattern recognition tasks where local structure is important. The authors' focus on real-time performance indicates a continued interest in efficient and accurate biometric systems.

Summary updated

BibTeX
@article{guo2014palmprint,
  author = {Jinyu Guo and Haibin Chen and Yuan Li},
  title = {{Palmprint Recognition Based on Local Fisher Discriminant Analysis}},
  journal = {J. Softw.},
  year = {2014},
  doi = {10.4304/jsw.9.2.287-292},
}
Figure