Finger Knuckle Print classification: Leveraging vision Mamba for low complexity and high accuracy
Abstract
Although FKP has been recognized as a viable alternative to other biometric modalities, its adoption remains
in the early stages due to its accuracy being lower than that of more competitive options. This research aims
to bridge this performance gap by presenting an advancement in the classification of Finger Knuckle Prints
(FKPs) using the Vision Mamba (ViM) model. The experimental study, conducted on a dataset from Hong Kong
Polytechnic University containing 7,920 images from 165 individuals, evaluated the ViM model’s performance
against several pretrained classification models. ViM achieved an impressive accuracy of 99.1%, outperforming
other models such as AlexNet (96.2%), SCNN (98.3%), and EfficientNet (98.0%), highlighting its superior
capability in FKP classification. With around 7 million parameters, ViM balances complexity and performance,
engineered specifically to capture fine-grained FKP features, such as texture and line patterns. Its use of weight
decay mitigates overfitting, and it demonstrates resilience in occlusion scenarios by maintaining performance
despite missing FKP components. Spatial attention mechanisms further enhance classification accuracy by
prioritizing the most informative regions. Seventeen pretrained deep neural networks were evaluated for their
effectiveness in classifying FKPs, with experimental results consistently demonstrating the superior performance
of the ViM model. ViM exemplifies an advanced deep learning approach for biometric applications, combining
high precision with efficient resource usage. Its versatile design – incorporating bidirectional SSMs for global
context modeling and position embeddings for spatial awareness – extends its applicability to visual tasks
beyond FKP identification. However, users should take its complexity and resource requirements into account
for practical implementation.