magistrsko delo
Abstract
Sodobna družba se vse bolj zaveda pomembnosti varovanja osebnih podatkov in zasebnosti. Prikrivanje identitet posameznikov na fotografijah ali video posnetkih je zato pomembno opravilo. Sodoben pristop k reševanju tega problema je generiranje nadomestnih obrazov, s katerimi zakrijemo izvirne. Algoritem imenovan $k$-Same-Net uporabi generativno nevronsko mrežo, ki sintetizira nadomestne obraze brez podobnosti resničnim osebam. Pristop je uspešen, vendar so generirani obrazi neraznoliki, slike pa niso povsem ostre. Naš cilj je izboljšava stabilnosti generativnega postopka in kvalitete rezultatov z uporabo najnovejših metod s področja generativnih modelov, natančneje z generativnimi nasprotniškimi mrežami. Preizkusili smo številne različne arhitekture in postopke učenja. Zaradi težavnega učenja generativnih nasprotniških mrež, ki je razvidno tudi iz našega dela, nismo prišli do zadovoljivih rezultatov. Nekatere implementacije uspešno posnemajo človeški obraz, izraz na obrazu in celo identiteto. Ni pa nam uspelo odkriti kombinacije parametrov, ki bi se naučila na podlagi več vhodnih identitet generirati prepričljiv nov obraz.
Keywords
deidentifikacija;generativne nasprotniške mreže;generiranje obrazov;računalništvo;računalništvo in informatika;magisteriji;
Data
Language: |
Slovenian |
Year of publishing: |
2019 |
Typology: |
2.09 - Master's Thesis |
Organization: |
UL FRI - Faculty of Computer and Information Science |
Publisher: |
[N. Sušin] |
UDC: |
004(043.2) |
COBISS: |
1538495939
|
Views: |
626 |
Downloads: |
212 |
Average score: |
0 (0 votes) |
Metadata: |
|
Other data
Secondary language: |
English |
Secondary title: |
Improving a deidentification model using generative adversarial networks |
Secondary abstract: |
In a privacy-concerned society it is a common challenge to protect the identities of people appearing in a photo or video. A modern approach is to generate surrogate faces to replace the originals. An algorithm called $k$-Same-Net uses generative neural networks to synthesize faces without any visual resemblance to real people. While highly successful, it suffers from low variety and blurriness of the generated faces. Our goal was to improve the quality and stability of this process by applying the latest methods in the field of generative neural networks, namely generative adversarial networks. We compare the quality of faces generated by several different implementations. Due to the difficulties of training generative networks, which are evident from our work, we were unable to achieve satisfactory results. Some of the methods we present are successful in imitating human faces, emotions and even identities. However, we were unsuccessful in finding a set of parameters that would result in convincing new identities based on multiple existing faces. |
Secondary keywords: |
deidentification;generative adversarial networks;generating faces;computer science;computer and information science;master's degree; |
Type (COBISS): |
Master's thesis/paper |
Study programme: |
1000471 |
Embargo end date (OpenAIRE): |
1970-01-01 |
Thesis comment: |
Univ. v Ljubljani, Fak. za računalništvo in informatiko |
Pages: |
64 str. |
ID: |
11381741 |