An end-to-end Algorithm Based on Spatial Transformer for Fingerprint Matching

Siqi Tang, Congying Han, Mingqiang Li, Tiande Guo

Abstract: Fingerprint recognition is not only applied in traditional criminal investigation, but also widely applied in civil field. Therefore, higher requirement is put forward for accuracy and efficiency of fingerprint recognition. In this paper, we propose an end-to-end fingerprint matching algorithm based on derivable spatial transformer. The algorithm consists of regression and classification network, which are connected into an end-to-end structure through affine transformation. Firstly, the core points of fingerprint are obtained by regression network. Then we use affine transformation to align and crop fingerprint image. Finally, classification network is used to extract fingerprint feature vector for fingerprint matching. The combination of alignment and feature extraction in an end-to-end structure is conducive to avoid information loss and promote each other to obtain higher matching accuracy. The experiment is tested on the public fingerprint dataset, which shows the improvement of efficiency and effectiveness.