The accuracy of ocular biometric systems is critically dependent on the image acquisition conditions and segmentation methods. To minimize recognition error robust segmentation algorithms are required. Among all ocular traits, iris got the most attention due to high recognition accuracy. New modalities such as sclera blood vessels and periocular region were also proposed as autonomous (or iris-complementary) modalities.
In this work we tackle ocular segmentation and recognition problems using deep learning methods, which represent state-of-the-art in many computer vision related tasks. We individually evaluate three recognition pipelines based on different ocular modalities (sclera blood vessels, periocular region, iris). The pipelines are then fused into a single biometric system and its performance is evaluated. The main focus is sclera recognition in the scope of which we i) create a new dataset named SBVPI, ii) propose and evaluate segmentation approaches, which won the first place on SS(ER)BC competitions, and iii) develop and evaluate the rest of the sclera-based recognition pipeline. The next contribution of this work is multi-class eye segmentation technique, which gives promising results. We also propose and evaluate deep learning pipeline for periocular recognition. For iris recognition we use an existing pipeline and evaluate it on our dataset.
With deep learning we achieve promising recognition results for each individual modality. We further improve recognition accuracy with multi-modal fusion of all three modalities.
|
