Phase-aggregated Dual-branch Network for Efficient Fingerprint Dense Registration

TL;DR

Experimental results demonstrate that the proposed Phase-aggregated Dual-branch Registration Network reaches the state-of-the-art registration performance in terms of accuracy and robustness, while maintaining considerable competitiveness in efficiency.

Abstract

Fingerprint dense registration aims to finely align fingerprint pairs at the pixel level, thereby reducing intra-class differences caused by distortion. Unfortunately, traditional methods exhibited subpar performance when dealing with low-quality fingerprints while suffering from slow inference speed. Although deep learning based approaches shows significant improvement in these aspects, their registration accuracy is still unsatisfactory. In this paper, we propose a Phase-aggregated Dual-branch Registration Network (PDRNet) to aggregate the advantages of both types of methods. A dual-branch structure with multi-stage interactions is introduced between correlation information at high resolution and texture feature at low resolution, to perceive local fine differences while ensuring global stability. Extensive experiments are conducted on more comprehensive databases compared to previous works. Experimental results demonstrate that our method reaches the state-of-the-art registration performance in terms of accuracy and robustness, while maintaining considerable competitiveness in efficiency.

Figures (12)

• Figure 1: Flowchart of proposed two-step fingerprint dense registration. Green areas indicate overlapping ridges, gray and red indicate non-overlapping ridges of the two fingerprints respectively.
• Figure 2: An overview of our dense deformation estimation network. 'Fp.' and 'Reg.' are the abbreviations of 'fingerprint' and 'registration' respectively. The network includes two encoder branches for extracting features of texture and correlation information, a multi-stage interaction module for fusing multi-scale and multi-semantic features, and an registration estimation module to predict the pixel-wise deformation field. The specific details of block architectures and image preprocessing flow are shown in Fig. \ref{['fig:block_architecture']} and Fig. \ref{['fig:preprocess_flowchart']} respectively.
• Figure 3: The specific architecture of blocks utilized in the proposed fingerprint dense registration network. '*' indicates that this layer connects convolution, batch normalization and ReLU in series. Numbers on the right side of each module and arrows represent the stride and current channel number respectively.
• Figure 4: A visual example of image preprocessing. Red rectangles represent the intermediate results of fingerprint enhancement. Blue rectangle represents the final result of image preprocessing, which corresponds to the output of the same module in Fig. \ref{['fig:network']}. The structure of 'EnhNet' is shown in Fig. \ref{['fig:network_preprocess']}.
• Figure 5: A simplified schematic of fingerprint enhancement network. The main structure is separated from FingerNet tang2017fingernet, and only the output is adjusted according to Equation \ref{['eq:enh_phase']}.