Unsupervised Learning of Fingerprint Rotations

Abstract

The alignment of fingerprint samples is a preprocessing step in fingerprint recognition. It allows an improved biometric feature extraction and a more accurate biometric comparison. We propose to use Convolutional Neural Networks for estimation of the rotational part. The main contribution is an unsupervised training strategy similar to Siamese Networks for estimation of rotations. The approach does not need any labelled data for training. It is trained to estimate orientation differences for pairs of samples. Our approach achieves an alignment accuracy with a mean absolute deviation 2:1 on data similar to the training data, which supports the alignment task. For other datasets accuracies down to 6:2 mean absolute deviation are achieved.