Palmprint Recognition across Different Devices

Wei Jia, Rong-Xiang Hu, Jie Gui, Yang Zhao, Xiao-Ming Ren

TL;DR

This paper tackles the challenge of palmprint recognition across different devices (PRADD), a crucial problem for real-world biometric systems where images are captured by various sensors. The authors introduce a new PRADD dataset and a robust scale normalization method, demonstrating that orientation coding-based methods, particularly CompCode and Ordinal Code, offer promising performance for this cross-device recognition task.

What Problem Does This Paper Address?

Palmprint recognition, a biometric technique that uses the unique patterns on a person's palm for identification, has shown great promise due to its rich features and ease of use. However, most existing research assumes that all palmprint images for training and testing are captured by the same device. In reality, individuals might have their palmprints captured by a variety of devices, such as digital cameras, smartphones, or even older devices. This variation in sensors introduces significant challenges, including differences in image quality, resolution, and acquisition conditions, leading to a drop in recognition accuracy. This problem, known as "Palmprint Recognition Across Different Devices" (PRADD) or sensor interoperability, has not been extensively studied. The paper aims to address this gap by investigating PRADD and proposing solutions to improve cross-device palmprint recognition.

Methods

The authors propose a novel approach to tackle PRADD, which involves several key components:

  • Dataset Creation: A new non-contact PRADD image database, named the Chinese Academy of Science—HeFei Institutes of Physical Science (CASHF) image database, was created. This database contains 12,000 grayscale images from 100 subjects, captured using three different devices: a Canon IXUS 950 IS digital camera and two smartphones (Motorola ME525 and Nokia 5800 XpressMusic). This dataset is crucial for evaluating PRADD performance.
  • Preprocessing:
    • Rotation Normalization: A classical algorithm is employed to correct for rotational variations in palmprint images. This involves converting the image to binary, identifying key points related to finger gaps, and rotating the image to make a specific line segment horizontal.
  • Scale Normalization: A new, robust algorithm for scale normalization is proposed. This method focuses on identifying a stable reference point along the heart line of the palm and then resizing the image to a consistent height. This is critical because many palmprint recognition methods are sensitive to scale variations.
  • ROI Extraction: After normalization, a Region of Interest (ROI) is cropped from the palmprint image. The authors refine the standard ROI extraction by slightly shifting the center point to ensure better cropping across different hand poses.
  • Recognition Methods: The paper evaluates three categories of palmprint recognition methods for PRADD:
  • Subspace Learning Methods: This includes Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and their variants (KPCA, KLDA, 2DPCA, 2DLDA, CSA, MDA). These methods aim to find a low-dimensional subspace for feature representation.
  • Correlation-Based Methods: Specifically, the Band-Limited Phase-Only Correlation (BLPOC) method is investigated. This method uses the phase information of the Fourier transform for matching.
  • Orientation Coding-Based Methods: Three methods are evaluated: Ordinal Code, Competitive Code (CompCode), and Robust Line Orientation Code (RLOC). These methods extract orientation features from the palmprint.

Datasets

  • CASHF Image Database: A newly created dataset for PRADD research.
  • Content: 12,000 grayscale palmprint images.
  • Subjects: 100 individuals (200 hands).
  • Devices: Canon IXUS 950 IS digital camera, Motorola ME525 smartphone, Nokia 5800 XpressMusic smartphone.
  • Acquisition: Non-contact, under indoor and daylight conditions.
  • Structure: Divided into three sub-databases (N5800, M525, C950) based on the capturing device.

Key Insights and Ideas

  • PRADD is a real and significant challenge: The study clearly demonstrates that palmprint recognition performance drops considerably when images are captured by different devices, highlighting the need for robust cross-device solutions.
  • A new dataset is essential: The creation of the CASHF database is a significant contribution, providing a much-needed benchmark for PRADD research.
  • Scale normalization is critical: The proposed scale normalization method is effective in reducing scale variations, which is a major hurdle for many existing palmprint recognition algorithms in a multi-device scenario.
  • Orientation coding methods excel in PRADD: Among the evaluated methods, orientation coding-based approaches, particularly CompCode and Ordinal Code, show superior performance for PRADD compared to subspace learning and correlation-based methods.
  • Palmprints are robust across devices: Despite the challenges, the study suggests that palmprints remain a viable biometric trait for cross-device identification, with orientation coding methods achieving high accuracy.

Key Results

  • Intra-sensor recognition: Orientation coding methods (CompCode, Ordinal Code, RLOC) achieved very high Best Identification Rates (BIRs), often near 100%, with low Equal Error Rates (EERs). BLPOC performed less well. Subspace learning methods showed satisfactory but not outstanding BIRs (around 85-90%).
  • Inter-sensor (PRADD) recognition:
    • Subspace learning methods performed poorly in PRADD scenarios, with BIRs significantly lower (74%-83%) compared to intra-sensor recognition.
  • Orientation coding methods (CompCode and Ordinal Code) maintained strong performance in PRADD, achieving average BIRs of around 99.7% and low average EERs of approximately 0.37%. RLOC also performed well but was slightly less stable.
  • BLPOC's performance also decreased in PRADD scenarios.
  • Best performing methods for PRADD: CompCode and Ordinal Code were identified as the most suitable methods for PRADD due to their stable and high recognition performance across different devices.

Limitations and Future Work

  • Limitations:
    • The quality of images captured by the M525 and N5800 smartphones was observed to be worse than that of the C950 camera, with less clear principal lines and wrinkles. This difference in image quality can impact recognition.
  • The study notes that some samples in the C950 database might have been poorly cropped or had affine transformations, potentially affecting matching with test samples from smartphones.
  • The performance of RLOC was noted to be easily influenced by image quality.
  • Future Work:
    • The authors suggest exploring other strategies to further improve PRADD recognition performance.
  • Developing multi-feature based methods is proposed as a promising direction to achieve even better performance for PRADD.

Summary updated

Abstract

In this paper, the problem of Palmprint Recognition Across Different Devices (PRADD) is investigated, which has not been well studied so far. Since there is no publicly available PRADD image database, we created a non-contact PRADD image database containing 12,000 grayscale captured from 100 subjects using three devices, i.e., one digital camera and two smart-phones. Due to the non-contact image acquisition used, rotation and scale changes between different images captured from a same palm are inevitable. We propose a robust method to calculate the palm width, which can be effectively used for scale normalization of palmprints. On this PRADD image database, we evaluate the recognition performance of three different methods, i.e., subspace learning method, correlation method, and orientation coding based method, respectively. Experiments results show that orientation coding based methods achieved promising recognition performance for PRADD.
BibTeX
@article{jia2012palmprint,
  author = {Wei Jia and Rong-Xiang Hu and Jie Gui and Yang Zhao and Xiao-Ming Ren},
  title = {{Palmprint Recognition across Different Devices}},
  journal = {Sensors},
  year = {2012},
  doi = {10.3390/s120607938},
}
Figure