A Study of Hand Back Skin Texture Patterns for Personal Identification and Gender Classification

Jin Xie, Lei Zhang, Jane You, David Zhang, Xiaofeng Qu2012 · DOI

TL;DR

This paper introduces a novel approach to personal identification and gender classification using hand back skin texture (HBST). By developing a specialized imaging system and employing a sparse representation-based texton learning method, the authors demonstrate that HBST patterns can provide valuable biometric information, either independently or as a complementary trait to existing methods like palmprints.

What Problem Does This Paper Address?

The paper addresses the challenge of finding new and effective biometric traits for personal identification and gender classification. While established biometrics like fingerprints and iris scans are widely used, there's a continuous need for complementary or alternative methods, especially for multi-modal systems that enhance robustness and accuracy. The authors specifically investigate hand back skin texture (HBST) as a potential biometric, noting its unique patterns that vary between individuals and genders, while remaining relatively stable over time for a given person. Existing texture analysis methods, while successful on standard datasets, struggle with the specific characteristics of HBST, such as the lack of distinct interest points and the high inter-class similarity (different people's skin textures can look similar).

Methods

The core of the paper's methodology involves:

  • Specialized HBST Imaging System: A custom device was designed to capture high-resolution (450 dpi) HBST images. This system uses a ring of LED lights for uniform illumination and a CCD camera with a fixed position to ensure consistent data acquisition. The design aims to capture the skin surface as a 2D plane, simplifying modeling.
  • Texton Learning with Sparse Representation (TL_SR): Instead of traditional methods like k-means clustering for texton learning, the authors adopt a sparse representation (SR) technique.
  • Filter Bank Responses: HBST images are first processed through an MR8 filter bank to extract rotation-invariant features.
  • Dictionary Learning: A dictionary of "textons" (representative micro-structures) is learned from these filter responses using SR, specifically minimizing the L1-norm. This approach is presented as more robust than k-means for HBST.
  • Feature Extraction: For each pixel, a sparse representation coefficient vector is computed using the learned dictionary. A histogram of these coefficients is then formed to represent the HBST image. This histogram serves as the final feature descriptor.
  • Classification: A nearest neighbor classifier is used, with the chi-squared distance metric to compare feature histograms for classification.

Datasets

  • Newly Introduced HBST Database: The authors established a novel HBST image database for their experiments.
  • Size: Consists of 1,920 images from 160 hands (80 individuals, 61 males, 19 females).
  • Collection: Images were collected in two sessions with a time interval of approximately 30 days.
  • Purpose: This dataset is designed to be challenging due to high inter-class similarity (different individuals' skin textures can look alike) and serves as a benchmark for HBST analysis.

Key Insights and Ideas

  • HBST as a Viable Biometric: The research strongly suggests that hand back skin texture, despite its subtle nature, contains sufficient discriminative information for personal identification and gender classification.
  • Limitations of Standard Texture Methods: Traditional texture analysis methods (like multi-fractal spectrum and basic LBP) perform poorly on HBST, highlighting the need for specialized approaches due to the unique characteristics of skin texture.
  • Power of Sparse Representation for Textons: Employing sparse representation for learning textons and extracting features proves more effective for HBST than standard clustering methods like k-means. This is attributed to SR's ability to capture more nuanced representations and achieve lower reconstruction errors.
  • Complementary Nature of Left and Right Hands: Fusing features from both left and right hands significantly boosts identification accuracy, indicating that these patterns offer complementary information.
  • HBST for Multi-Modal Biometrics: The study demonstrates that HBST can effectively augment other biometrics (like palmprints), leading to substantially improved identification rates.

Key Results

  • Personal Identification:
    • In experiments where left and right hands were treated as different classes (160 classes), the proposed TL_SR method achieved 86.81% accuracy, outperforming DLBP (75.56%), CLBP (84.51%), and TL_KM (84.40%).
  • When left and right hands of the same person were treated as the same class (80 classes), TL_SR achieved 90.17% accuracy.
  • Fusing left and right hand HBST data further improved accuracy, reaching 92.51% with TL_SR.
  • Fusion with palmprint data achieved a remarkable 99.58% accuracy.
  • Gender Classification:
    • The TL_SR method achieved 98.65% accuracy for gender classification, significantly outperforming other methods.
  • Error rates were low: 1.23% for males and 1.75% for females, even with an imbalanced dataset.

Limitations and Future Work

  • Dataset Size and Balance: The authors acknowledge that their HBST database, while reasonably sized for illustration, may not be large or comprehensive enough for very strong conclusions. They plan to collect more samples, especially to balance the male-female ratio and cover a wider age range.
  • Factors Affecting Performance: External factors like hairs on the skin and skin humidity can influence performance. The current database has limited hair due to the demographic of the collected samples.
  • Modeling Skin Texture Over Time: The paper notes that skin texture changes with age. Future work will focus on modeling these long-term variations to improve biometric performance over extended periods.
  • Algorithm Development: The established HBST dataset is presented as a valuable platform for developing and evaluating new texture feature extraction and classification algorithms.

Summary updated

BibTeX
@misc{xie2012study,
  author = {Jin Xie and Lei Zhang and Jane You and David Zhang and Xiaofeng Qu},
  title = {{A Study of Hand Back Skin Texture Patterns for Personal Identification and Gender Classification}},
  year = {2012},
  doi = {10.3390/s120708691},
}
Figure