CNN-Based Biometric Recognition Using Palmprint and Fingerprint Images
Biometric recognition systems are widely used in modern security because they provide reliable identity verification based on unique human traits such as fingerprints and palmprints. However, unimodal systems often suffer from limitations such as noise sensitivity, variability in data, and reduced performance under poor image quality.
Approach
This work presents a Convolutional Neural Network (CNN) approach using two biometric modalities: palmprint and fingerprint images. The objective is to develop and evaluate independent CNN models for each modality as a foundation for future multimodal biometric fusion.
Datasets and Preprocessing
The BMPD dataset was used for palmprint recognition, while the FVC2004 DB1-B dataset was used for fingerprint recognition. All images were converted to grayscale, resized to 64×64 pixels, and normalized to improve model performance.
Model Overview
Separate CNN architectures were designed for each modality. The models were trained independently to extract discriminative features from palmprint and fingerprint images.
Results
The palmprint model achieved moderate performance with approximately 75% accuracy, indicating effective feature learning. However, the fingerprint model showed weak performance due to class imbalance and misclassification issues, often predicting a dominant class.
Conclusion
This study establishes a baseline for biometric recognition using CNNs on palmprint and fingerprint images. While the palmprint model performed reasonably well, the fingerprint model requires further improvement. The work highlights the potential of deep learning in biometrics and sets the stage for multimodal fusion in future research.
Future Work
Future improvements will focus on enhancing dataset size, optimizing CNN architectures, and applying advanced techniques such as metaheuristic-based feature selection using a Golf Optimization-inspired algorithm. Most importantly, multimodal fusion of palmprint and fingerprint features will be explored to improve overall system accuracy and robustness.
Comments
Post a Comment