Abstract

Biometrics play an important role in security measures, such as border control and online transactions, relying on traits like uniqueness and permanence. Among the different biometrics, the fingerprint stands out for their enduring nature and individual uniqueness. Fingerprint recognition systems traditionally rely on ridge patterns (Level 1) and minutiae (Level 2). However, these systems suffer from recognition accuracy with partial fingerprints. Level 3 features, such as pores, offer distinctive attributes crucial for individual identification, particularly with high-resolution acquisition devices. Moreover, the use of convolutional neural networks (CNNs) has significantly improved the accuracy in automatic feature extraction for biometric recognition.

A CNN-based pore fingerprint recognition system consists of two main modules, pore detection and pore feature extraction and matching modules. The first module generates pixel intensity maps to determine the pore centroids, while the second module extracts relevant features of pores to generate pore representations for matching between query and template fingerprints. However, existing CNN architectures lack in generating deep-level discriminative feature and computational efficiency. Moreover, available knowledge on the pores has not been taken into consideration optimally for pore centroids and metrics other than Euclidean distance have not been explored for pore matching.

The objective of this research is to develop a CNN-based pore fingerprint recognition scheme that is capable of providing a low-complexity and high-accuracy performance. The design of the CNN architecture of the two modules aimed at generating features at different hierarchical levels in residual frameworks and fusing them to produce comprehensive sets of discriminative features. Depthwise and depthwise separable convolution operations are judiciously used to keep the complexity of networks low. In the proposed pore centroid part, the knowledge of the variation of the pore characteristics is used. In the proposed pore matching scheme, a composite metric, encompassing the Euclidean distance, angle, and magnitudes difference between the vectors of pore representations, is proposed to measure the similarity between the pores in the query and template images.

Extensive experiments are performed on fingerprint images from the benchmark PolyU High-Resolution-Fingerprint dataset to demonstrate the effectiveness of the various strategies developed and used in the proposed scheme for fingerprint recognition.