Striver: an image descriptor for fingerprint liveness detection

Discriminant image feature plays a key role in fingerprint liveness detection. In this paper, we propose a Fisher vector learning based image representation method for fingerprint liveness detection. Different from conventional methods, we consider the Fisher vector based image feature learning both in the spatial domain and the frequency domain. The contributions of our method are summarized as follows: (1) Image are transformed to the local frequency domain and global frequency domain by using the local Fourier transform and global Fourier transform, respectively. The frequency domain can not only preserve the image contextual information in the original spatial domain but also robust for image representation. (2) In the global frequency domain, image high frequency information is introduced in the feature extraction process rather than discard as usually did. (3) To take full advantage of the complementary in spatial and frequency domains, image local spatial feature, local frequency feature and global frequency feature are fused in the Fisher vector learning process. Extensive experimental results conducted on three benchmark databases demonstrate the superior performance of our method compared with other peer methods. Specifically, on LivDet 2011, LivDet 2013, and LivDet 2015 database, the average classification error are reduced to 5.16%, 1.40%, and 7.51%, respectively.