Deep Learning in Biometry

                                                    Peter Peer
                                          Computer Vision Laboratory
                                  Faculty of Computer and Information Science
                                             University of Ljubljana
                                    Večna pot 113, 1000 Ljubljana, Slovenia
                                              peter.peer@fri.uni-lj.si




                                                           Abstract
                        In the recent years we are witnessing a dominance of deep neural net-
                        work learning approach in the machine learning field. Eventhough the
                        neural network concept is more than 50 years old, only the recent
                        developments enabled its wide use. Namely, availability of process-
                        ing power, especially graphical processing units, availability of large
                        databases, and the refined knowledge about the approach itself. Con-
                        volutional neural network as a special case of deep learning approach
                        is widely used in computer vision domain, also in biometry. A lot of
                        image material is obtained through surveillance scenarios, where we
                        can use modalities like face, gait, and ears for recognition, normally
                        within a multibiometrics system. Each modality has to be first de-
                        tected and then recognized. The following examples are discussed in
                        this context: 1) network for ear detection in the wild, where, different
                        from competing techniques from the literature, our approach does not
                        simply return a bounding box around the detected ear, but provides
                        accurate and detailed, pixel-wise information about the location of the
                        ears in the image; 2) training network with limited training data for ear
                        recognition in the wild, where we explore different strategies towards
                        model training with limited amounts of training data and show that by
                        selecting an appropriate model architecture, using aggressive data aug-
                        mentation, and selective learning on existing (pre-trained) models, we
                        are able to learn an effective model; 3) face deidentification (anonymiza-
                        tion) with generative network that provides privacy guaranties and at
                        the same time retains certain important characteristics of the data even
                        after deidentification.




Copyright cc 2017
Copyright      by thebypaper’s
                         the paper’s authors.
                               authors.       Copying
                                         Copying        permitted
                                                  permitted        for private
                                                             for private       and academic
                                                                         and academic        purposes.
                                                                                        purposes.
In: S.A.Hölldobler,
         Editor, B.A.Coeditor
                        Malikov, (eds.):
                                  C. Wernhard   (eds.): of
                                          Proceedings   YSIP2   – Proceedings
                                                           the XYZ    Workshop,of the Second Country,
                                                                                  Location,  Young Scientist’s International
                                                                                                       DD-MMM-YYYY,          Workshop
                                                                                                                          published at
on Trends in Information Processing, Dombai, Russian Federation, May 16–20, 2017, published at http://ceur-ws.org.
http://ceur-ws.org




                                                                  1
                                                                 15