FV-EffResNet: an efficient lightweight convolutional neural network for finger vein recognition.

Area of Science:

• Biometrics
• Computer Vision
• Machine Learning

Background:

• Deep neural networks have advanced finger vein recognition but often require high computational resources.
• Current state-of-the-art models are complex and computationally expensive, limiting real-time and embedded applications.

Purpose of the Study:

• To develop a lightweight convolutional neural network (CNN) for finger vein recognition.
• To achieve a balance between network size, speed, and accuracy for practical implementation.

Main Methods:

• Introduced FV-EffResNet, a novel lightweight CNN architecture.
• Proposed an Efficient Residual (EffRes) block using decomposed convolutions (pointwise and depthwise) with rectangular dimensions.
• Utilized squeeze units, depthwise convolution, pooling strategy, Swish activation, and cyclical learning rates for efficiency and performance.

Main Results:

• The EffRes block significantly improved finger vein recognition performance.
• FV-EffResNet demonstrated state-of-the-art results in identification and verification tasks.
• The network achieved high accuracy with reduced computational costs and complexity.

Conclusions:

• The proposed FV-EffResNet provides an effective and efficient solution for finger vein recognition.
• The lightweight design and novel EffRes block enable practical real-time applications on embedded systems.