TransFCN: A Novel One-Stage High-Resolution Fingerprint Representation Method

Since pores are widely used to represent high-resolution fingerprint images, the detection and representation of pores are essential for high-resolution fingerprint recognition. The latest method uses only one fully convolutional network to represent high-resolution fingerprint images for subsequent recognition by combining pore detection and pore representation into one stage, showing good generalization and pore detection ability. Nevertheless, it still has limitations in feature learning and pore detection due to its network architecture and the loss used. To tackle the limitations, in this paper, we propose a novel network architecture, namely TransFCN, for one-stage high-resolution fingerprint representation. We introduce the transformer and attention module into our network architecture and combine them with the fully convolutional network to effectively learn both global and local information. In addition, we employ the adaptive wing loss and weighted loss map to further improve the pore detection capability. Experimental results on the PolyU HRF dataset demonstrate the effectiveness of our proposed method in pore detection and feature learning. Furthermore, the experimental results on an in-house dataset demonstrate the excellent generalization capability of our proposed method when compared to the state-of-the-art two-stage method.