This diploma thesis addresses the challenge of person recognition based on sclera images from the open dataset SBVPI. The work focuses on the use of deep learning to improve the recognition accuracy between two segmented vasculature sclera images. We developed and evaluated two different approaches: Siamese neural networks and binary classification. The results show that we have successfully improved the recognition accuracy. Furthermore, we found that the approach based on separating learning according to the direction of gaze is promising for further improvements in system performance.
|
