Generative deep models for ear images

Za dobro delovanje potrebujejo globoke nevronske mreže veliko podatkov. V primeru biometrične modalnosti uhljev - največje anotirane baze slik uhljev v nekontroliranem okolju zajemajo nekaj tisoč slik, kar je premalo za globoko učenje in razpoznavo. Ta problem skušamo rešiti z uporabo generativnih nevronskih mrež za obogatitev baze. Implementiramo dva tipa generativnih nevronskih mrež: generativno mrežo in variacijski avtokodirnik. Obe mreži naučimo s pomočjo slik iz obstoječe baze in z vsako generiramo množico umetnih podatkov (slike uhljev). Z vsako od teh množic nato učimo mreže za razpoznavo in primerjamo rezultate. Kljub uporabi umetno generiranih slik, ne uspemo doseči visoke stopnje razpoznave na bazi AWE-v1, vseeno pa so opazne izboljšave v primerjavi z rezultati učenja razpoznave brez umetno generiranih slik.Deep neural networks require large amounts of data to perform well. In the case of the biometrical modality of the human ear, the largest annotated databases of images of ears in an uncontrolled environment consist of a few thousand images, which is insufficient for recognition using deep learning. We try to solve this problem using generative neural networks for data augmentation. We implement two types of generative neural networks: a generative network and a variational autoencoder. We train both networks on images from the existing database and then use them to generate a new set of artificial data (images of ears) with each. We then use each of these datasets to train neural networks for recognition and compare the results. Even using artificially generated images, we do not manage to achieve a high recognition rate on the AWE-v1 ear database. Despite that, there is a noticeable improvement compared to results of training for recognition without using generated data