Two Stream Convolutional Neural Network for Full Field Optical Coherence Tomography Fingerprint Recognition

Full-Field Optical Coherence Tomography (FF-OCT) for fingerprint imaging has been recently explored to counter presentation attacks (previously referred as spoofing attacks). The ability to acquire discriminant information under the surface of the external fingerprint can not only detect such attacks, but also provide supplementary information to make fingerprint recognition superior. In this work, we present a new approach for robust fingerprint recognition by learning deep representation by employing multiple subsurface fingerprints. Specifically, we design a new Two Stream - Convolutional Neural Network (TS-CNN) to employ internal fingerprint images captured at 6 different depths. Further, to accelerate the learning of the features, we transfer the weights of the AlexNet for each stream. With a semi-public in-house FF-OCT database of 200 unique fingerprints, we demonstrate the applicability of the proposed approach by achieving 0.17% Equal Error Rate (EER). While the proposed TS-CNN is trained and fine-tuned on a development subset of fingerprints from FF-OCT fingerprint image database, the final results are reported on the disjoint testing set of fingerprints from the same dataset.