ConvNet Regression for Fingerprint Orientations

Estimation of orientation fields is a crucial task in fingerprint recognition. Many processing steps depend on their precise estimation and the direction of fingerprint minutiae is a valuable information. But especially for regions of low quality the task is not trivial and engineered approaches on local features may fail. Methods that combine local and global features learned from the data are state of the art and benchmarked with the framework FVC-ongoing. We propose to use Convolutional Neural Networks trained in a regression to estimate the orientation field (ConvNetOF). Regression is more accurate than classification in this case. Our approach achieves an RMSE of 8.53◦ on the Bad Quality Dataset of the FVC-ongoing benchmark. This is the best result reported so far. 1 Motivation and Introduction Fingerprint recognition is one of the most wide spread biometric modalities, when it comes to identification and verification of individuals. Recognition algorithms make use of the distinctive features in the fingerprints. Fingerprint minutiae are features, which are typically used for recognition. Minutiae are characteristic points of the papillary ridges, e.g. an ending and a bifurcation [13]. The spatial distribution and relations of positions and directions of minutiae are unique for every finger which allows to distinguish fingerprints. The direction of a fingerprint minutia is one of its most valuable informations for recognition besides its type and position. It directly depends on the local orientation at its location. The orientation field (OF) of the papillary ridges (see figure 3a) is itself another important feature in fingerprint recognition [13]. Besides this, the OF is relevant information for image enhancement and many processing steps along the workflow of a biometric feature extraction [13]. Deviations between the estimation and the real OF have to be as small as possible for the whole fingerprint area [3]. Otherwise biometric features may not be extracted correctly or spurious features may be generated.