A STN-Based Self-supervised Network for Dense Fingerprint Registration

Yang Yu, Haixia Wang, Yilong Zhang, Peng Chen

Abstract: Dense fingerprint registration in the preprocessing stage plays a vital role in the subsequent fingerprint fusion, mosaic, and recognition. However, the existing conventional methods are limited by handcraft features, while the methods based on deep learning lack a large amount of ground truth displacement fields. To overcome these limitations, we propose a self-supervised learning model to directly output densely registered fingerprints. With a spatial transformation network (STN) connected after fully convolutional network (FCN), image deformation interpolation can be achieved to obtain the registered image. Self-supervised training is achieved by maximizing the similarity of images, without the need for ground truth displacement fields. We evaluate the proposed model on publicly available datasets of internal-external fingerprint image pairs. The results demonstrate that the accuracy of the model is comparable to that of the conventional fingerprint registration while executing orders of magnitude faster.