A single scale retinex based method for palm vein extraction

Chongyang Wang, Ming Peng, Lingfeng Xu, Tong ChenarXiv 2016 · arXiv

Abstract

Biometric palm vein recognition often fails when lighting is uneven, creating shadows that mask underlying vascular patterns. This paper introduces a preprocessing pipeline centered on the Single Scale Retinex (SSR) algorithm to strip away these illumination artifacts. The proposed method results in significantly clearer vein extraction and an 18.4% boost in image contrast compared to traditional enhancement techniques.

Introduction

Palm vein recognition is a highly secure biometric because vascular patterns are internal and nearly impossible to forge. However, capturing a high-quality "map" of these veins is difficult in real-world scenarios. Because the human palm is not a flat surface and lighting is rarely perfectly symmetrical, raw images often suffer from "asymmetric illumination"—heavy shadows and bright spots that confuse standard image processing algorithms.

Traditional enhancement techniques like Contrast Limited Adaptive Histogram Equalization (CLAHE) or Gaussian filters often amplify shadows rather than removing them. This research seeks to isolate the "true" reflection of the veins from the "noise" of the light source using Retinex theory.

Methodology

The researchers developed a multi-stage pipeline to move from a raw, shadowed image to a clean binary skeleton of palm veins:

  1. Hyperspectral Acquisition: Data is captured at the 850nm wavelength. This specific band is the "sweet spot" where deoxygenated hemoglobin in the blood absorbs light most effectively, making veins appear dark against surrounding tissue.
  2. Normalization and Segmentation: The raw image is normalized to a standard gray-scale range, and the central part of the palm (the area of interest) is cropped.
  3. Single Scale Retinex (SSR): This is the core innovation. Based on Edwin Land’s Retinex theory, an image is viewed as the product of two components: Illumination (the light hitting the object) and Reflection (the intrinsic properties of the object). By using a Gaussian function to estimate the illumination component and subtracting it (in the log domain) from the original image, the researchers effectively "cancel out" the shadows.
  4. Post-Processing: After SSR, the image's dynamic range is stretched using Histogram Equalization (HE) and smoothed with a Median Filter to remove salt-and-pepper noise.
Figure 5: Comparison of (a) the original shadowed image, (b) the image after SSR processing, and (c) the final result after SSR and Histogram Equalization.
  1. Binarization and Morphology: Finally, the image is converted to black and white (binarization). Morphological operations delete "false" vein fragments, and a thinning algorithm produces the final vein lines.

Key Results and Insights

The proposed SSR-based method was compared against three common techniques: CLAHE, DoG-HE (Difference of Gaussian-Histogram Equalization), and Gaussian Low Pass Filtering. The SSR method outperformed them across three key metrics:

  • Contrast Ratio: Increased by 18.4% (reaching a value of 76.71 compared to the original 32.82).
  • Definition (Sharpness): Increased by 18.8% (reaching 14.27, whereas the next best method, DoG-HE, reached 12.01).
  • Entropy (Information Content): Increased by 1.07%, indicating that the method preserves more useful biometric data while discarding noise.
Figure 4: The final stages of the pipeline: (a) the binarized vein image and (b) the thinned skeleton used for biometric matching.

Qualitatively, the SSR method was the only one that successfully removed the large central shadow of the palm without blurring the fine vein structures. A critical insight is that SSR alone compresses the dynamic range; following it with Histogram Equalization is essential to making the veins "pop" after the shadows have been neutralized.

Conclusion and Future Work

While the SSR-based pipeline demonstrates a clear improvement in image quality for palm vein extraction, several areas remain open for exploration. The computational complexity of Gaussian convolutions and log-domain transformations may pose challenges for real-time biometric scanners. Future work could explore Multi-Scale Retinex (MSR) to balance color rendition and dynamic range compression further, as well as testing the impact of this improved extraction on final recognition accuracy (False Acceptance Rates).

Summary updated

BibTeX
@misc{wang2016single,
  author = {Chongyang Wang and Ming Peng and Lingfeng Xu and Tong Chen},
  title = {{A single scale retinex based method for palm vein extraction}},
  year = {2016},
  eprint = {1605.08154},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1605.08154},
}
Figure