Biometric Recognition Using Feature Selection and Combination

Ajay Kumar, David ZhangAVBPA 2005 · DOI

TL;DR

This paper tackles the often-overlooked problem of feature selection in biometric recognition, demonstrating that a small subset of features can achieve high accuracy while significantly reducing computational complexity. They propose and evaluate a correlation-based feature selection method, showing its effectiveness in improving performance and efficiency for hand-based biometrics.

What Problem Does This Paper Address?

Biometric recognition systems, which identify individuals based on unique physiological or behavioral characteristics, typically involve two main stages: feature extraction and classification. While much research focuses on developing sophisticated feature extraction techniques and accurate classifiers, the crucial step of evaluating the usefulness of these extracted features has been largely ignored. This leads to systems that might use an excessive number of features, increasing computational cost and potentially introducing irrelevant or redundant information that can actually harm classifier performance. The paper addresses this by investigating feature subset selection to identify the most discriminative features, aiming to improve accuracy and efficiency, especially for online recognition systems. They also explore feature-level fusion, where information from different biometric modalities is combined at the feature stage, an area they note has received less attention.

Methods

The core of the paper's methodology revolves around feature subset selection and feature-level fusion.

  • Feature Subset Selection: They employ the Correlation-based Feature Selection (CFS) algorithm. CFS is a "filter-based" method, meaning it evaluates features independently of any specific classifier. Its heuristic is that a good feature subset should have high correlation with the class label (i.e., be discriminative) but low correlation among themselves (i.e., be non-redundant). The CFS algorithm uses a correlation-based objective function to guide a search for the best feature subset.
  • Feature-Level Fusion: The paper investigates combining features extracted from different biometric modalities (palmprint and hand-shape) at the feature level. This is contrasted with score-level or decision-level fusion, which are considered less effective by the authors.
  • Classification Schemes: A variety of popular machine learning algorithms are used to evaluate the performance of the selected feature subsets and fused features. These include:
  • Naive Bayes: (Normal, Estimated, Multinomial)
    • k-Nearest Neighbor (k-NN)
    • Support Vector Machine (SVM) (with polynomial kernel)
    • Feed-Forward Neural Network (FFN)
    • Decision Trees: (C4.5 and Logistic Model Tree - LMT)

Datasets

The experiments were conducted on a biometric image database collected from 100 subjects.

  • The dataset consisted of 1000 hand images in total, with 10 images per subject.
  • These images were obtained from a digital camera in an indoor environment.
  • The data was collected over two sessions with an average interval of three months to assess modality stability (though the primary focus is on feature selection).
  • The dataset was split into five image samples per user for training and the remaining for testing. The same splits were used for fair comparison.

Key Insights and Ideas

  • Feature Selection is Crucial: The paper strongly suggests that while many extracted biometric features might be useful, only a small subset is truly necessary for accurate identification. This has significant implications for efficiency and practicality.
  • Feature-Level Fusion Benefits from Selection: Combining features from different modalities at the feature level can improve performance, but this benefit is amplified when feature selection is applied to the combined set. This indicates that not all combinations of individual reduced feature sets are optimal; a new, combined subset might be more discriminative.
  • CFS Algorithm Effectiveness: The Correlation-based Feature Selection (CFS) algorithm proved effective in identifying relevant features and reducing dimensionality without sacrificing accuracy, and in many cases, improving it.
  • Classifier Sensitivity to Redundancy: Some classifiers, like Naive Bayes (especially the multinomial model), are more sensitive to irrelevant and redundant features, showing performance degradation when such features are present.
  • LMT's Robustness: The Logistic Model Tree (LMT) classifier demonstrated robustness to redundant features, maintaining good performance even with a larger feature set, likely due to its stagewise fitting process that selects relevant features.

Key Results

  • Palmprint Recognition (Table 2): Feature selection reduced the number of palmprint features from 144 to 69, improving accuracy for most classifiers. SVM achieved the best performance (95.8% with 69 features).
  • Hand-Shape Recognition (Table 3): Feature selection reduced hand-shape features from 23 to 15, generally improving accuracy. SVM and k-NN performed well (around 88.6% and 89.6% respectively with 15 features).
  • Combined Features (Table 4): CFS selected a 75-feature subset from the combined 167 features. SVM achieved the highest accuracy (98% with 75 features), closely followed by k-NN. The combination of features was beneficial for most classifiers.
  • Efficiency Gains: The reduction in feature dimensionality significantly reduces computational complexity, making online recognition more feasible. For example, the paper notes that feature selection found 20 best features (15 palmprint, 5 hand-shape) yielding ~96% accuracy with SVM.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the performance of feature selection must be measured indirectly by comparing classifier performance with and without selection.
  • The multinomial event model for Naive Bayes showed degraded performance on hand-shape features, attributed to potentially inappropriate estimation of term probabilities due to small feature vector sizes.
  • Future Work:
    • The authors are currently working to examine the issue of feature subset selection for other biometric modalities.
  • Further investigation into the usefulness of new hand-shape features (f3, f7, f8, f10-f14, f17-f18, f20-f23) is implied, as some showed promise while others did not establish significance.

Summary updated

BibTeX
@inproceedings{kumar2005biometric,
  author = {Ajay Kumar and David Zhang},
  title = {{Biometric Recognition Using Feature Selection and Combination}},
  booktitle = {AVBPA},
  year = {2005},
  doi = {10.1007/11527923\_85},
}
Figure