Biometric Systems

In the era of advanced biometric technologies, the demand for secure and efficient identification methods has significantly increased. Our project focuses on the finger-photo identification system, an innovative alternative to traditional fingerprint-based identification techniques. By leveraging image processing and machine learning, this system identifies individuals based on unique patterns and details captured in finger photos, particularly using mobile-resolution images.

The motivation behind this project stems from the advantages finger-photo identification offers over other biometric approaches:

• Standard equipment: Finger photos can be captured using common devices, such as modern mobile cameras.

• Non-intrusive capture: The process of taking finger photos is simple, user-friendly, and does not require specialized sensors or tools.

• Integration capability: The system seamlessly integrates with existing identification frameworks, enabling comparisons with traditional fingerprint systems.

The system operates as a finger-photo identification framework that matches a probe image against a gallery of stored finger images. To achieve accurate results, the system assumes a closed-set identification approach, where at least one match exists in the gallery. However, the framework can be extended to support open-set identification scenarios with minor modifications.

At a high level, the system follows the following pipeline:

• Image Preprocessing - To ensure quality and consistency of finger images, the following preprocessing steps are applied:

  • Background removal,

  • Orientation alignment,

  • Cropping fingertips,

  • CLAHE enhancement (Contrast Limited Adaptive Histogram Equalization),

  • Thresholding,

  • Ridge thinning.

• Contrastive Learning: a deep learning approach based on contrastive learning is employed to extract and compare features effectively between the probe image and the gallery images.

• Gallery Evaluation: the system evaluates the probe image against the gallery and identifies the image with the highest matching score.

The pipeline extensively employs image processing techniques to simplify feature extraction, followed by the utilization of a pretrained backbone model for image embedding, with contrastive fine-tuning inspired by SimCLR. The results highlight the effectiveness of this approach and suggest further exploration.

The full report for this project, which covers the implementation details, experimental settings, and evaluation results, can be downloaded from the following link:

Download Report.

Additionally, an interactive demo is available on Google Colab, which you can access here.