Visual Feature-Guided Diamond Convolutional Network for Finger Vein Recognition

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Finger vein (FV) biometrics offer high security and non-contact identity authentication.
• Existing finger vein recognition (FVR) systems face challenges with limited training data and inconsistent image quality.

Purpose of the Study:

• To introduce a novel convolutional neural network, the visual feature-guided diamond convolutional network (VF-DCN), to improve FVR system performance.
• To address data scarcity and image quality issues in FVR through an innovative network architecture and unsupervised training.

Main Methods:

• Developed VF-DCN, a multi-scale, multi-orientation convolutional neural network utilizing Log-Gabor filters for kernel tuning.
• Implemented a diamond-shaped convolutional kernel architecture inspired by human vision, optimizing filter allocation across scales.
• Employed a three-layer configuration and fully unsupervised training for simplicity and performance.

Main Results:

• VF-DCN achieved exceptional Equal Error Rates (EERs) as low as 0.17% and Accuracy Rates (ACC) up to 100% across four diverse FV databases.
• Demonstrated superior recognition accuracy and robustness compared to existing FVR approaches.
• Exhibited a reduced number of parameters and lower model complexity.

Conclusions:

• The VF-DCN effectively overcomes limitations in FVR systems, offering high accuracy and robustness.
• The proposed network architecture and training strategy provide a promising solution for secure and efficient biometric authentication.
• VF-DCN presents a computationally efficient and highly accurate method for finger vein recognition.