Deep Expectation for Estimation of Fingerprint Orientation Fields

Estimation of the orientation field is one of the key challenges during biometric feature extraction from a fingerprint sample. Many important processing steps rely on an accurate and reliable estimation. This is especially challenging for samples of low quality, for which in turn accurate preprocessing is essential. Regressional Convolutional Neural Networks have shown their superiority for bad quality samples in the independent benchmark framework FVC-ongoing. This work proposes to incorporate Deep Expectation. Options for further improvements are evaluated in this challenging environment of low quality images and small amount of training data. The findings from the results improve the new algorithm called DEX-OF. Incorporating Deep Expectation, improved regularization, and slight model changes DEX-OF achieves an RMSE of 7.52° on the bad quality dataset and 4.89° at the good quality dataset at FVC-ongoing. These are the best reported error rates so far.