Biometric Finger Vein Classification using Finger Inferred Images with ResNet50 and Attention Mechanisms

Finger vein recognition is gaining traction as a secure and reliable biometric authentication method due to its resistance to forgery and its internal nature, making it less vulnerable to manipulation compared to external biometrics like fingerprints. Finger vein-based authentication is a promising solution for secure and reliable identity verification, as finger vein patterns are highly unique and internal to the body, making them difficult to forge or replicate. This internal nature ensures resilience against environmental factors like dirt, moisture, or external wear, providing consistent and accurate results. This research focuses on classifying hand types (left, right) and specific finger categories (index, middle, ring) using finger vein inferred images, which present challenges such as varying vein visibility due to external factors like lighting and skin tone. In this study, we propose a model that integrates spatial attention module and a convolutional block attention module (CBAM) with pretrained ResNet50 and dense layers to enhance feature extraction from finger vein images. By incorporating these attention mechanisms with a pretrained ResNet50 architecture and dense layers, our model achieved a classification accuracy of 97.98%, precision of 98.92%, recall of 97.89%, and F1-score of 98.44% using the spatial attention mechanism with ResNet50. With CBAM, the model achieved an accuracy of 95.85%, precision of 95.85%, recall of 93.85%, and an F1-score of 95.54%. For ablation studies, we also evaluated several other models, including CNN, ANN, CNN with spatial attention and dense layers, and ANN with spatial attention and dense layers, achieving prediction accuracies of 93.32%, 93.85%, 93.06%, and 92.80%, respectively.