Finger Vein Recognition Based on Unsupervised Spiking Convolutional Neural Network with Adaptive Firing Threshold

Area of Science:

• Biometrics
• Computational Neuroscience
• Machine Learning

Background:

• Deep neural networks (DNNs) advance finger vein recognition (FVR) but require complex architectures and extensive training.
• Static, continuous-valued activations in DNNs limit efficiency and biological plausibility.

Purpose of the Study:

• To introduce an adaptive firing threshold-based spiking neural network (ATSNN) for improved FVR.
• To leverage discrete spike encodings for spatio-temporal feature extraction in finger vein images.
• To enhance network robustness and reduce model complexity compared to traditional DNNs.

Main Methods:

• Finger vein images are converted to spike latency encodings using Gabor and Difference of Gaussian filters.
• An ATSNN processes spike encodings, extracting features via biologically plausible local learning rules.
• Neuron firing thresholds are dynamically adjusted based on average potential tensors for adaptive response modulation.

Main Results:

• The ATSNN model achieves remarkable recognition accuracy across three benchmark datasets.
• The proposed method demonstrates reduced parameter count and model complexity compared to existing FVR techniques.
• ATSNN exhibits superior performance, surpassing several current FVR methods.

Conclusions:

• The ATSNN offers a more biologically plausible, efficient, and robust solution for FVR.
• Its sparse, event-driven nature presents an advantage over traditional DNNs in biometric applications.
• This approach paves the way for advanced, efficient biometric security systems.