A comparative analysis of kernel filtering techniques for convolutional neural networks in fingerprint recognition

Abstract

Fingerprint recognition remains one of the most widely adopted biometric methods for individual verification due to its uniqueness and reliability. This study investigates the role of kernel-based filtering in the feature extraction stage and evaluates the performance of four Convolutional Neural Network (CNN) architectures: Siamese Network, LeNet, ResNet, and VGG-19. Experiments were conducted on the SOCOFing dataset, with preprocessing steps involving image resizing, normalization, and augmentation. To enhance the discriminative quality of fingerprint features, Sobel, Gabor, Laplacian, and Gaussian kernels were applied prior to model training. Model performance was assessed using accuracy, loss, and the area under the ROC curve (AUC), with AUC emphasized as the primary metric to address class imbalance. The results demonstrate that Gaussian and Sobel kernels consistently yield superior and stable performance across all models. Furthermore, statistical validation through Friedman and Nemenyi tests confirmed significant differences among kernel–model combinations, underscoring the benefit of multi-kernel preprocessing in biometric classification. Overall, the findings highlight the critical role of kernel selection in optimizing fingerprint recognition systems, with Gaussian filtering showing particular promise for enhancing adaptability, robustness, and classification accuracy.