Deidentifikacija osoba u slikama

undergraduate thesis

• Title (english): Person De-Identification in Images
• Author: Luka Dragutin
• Mentor: Zoran Kalafatić (mentor)
• Committee member: Zoran Kalafatić (predsjednik povjerenstva)
• Committee member: Ivona Brajdić (član povjerenstva)
• Committee member: Zoran Medić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Čovjekova privatnost se može narušiti sa snimkama nadzornih kamera koje neprestano snimaju javni prostor, te samim time i osobe u njemu. Jer je identitet osoba rijetko važan u službi nadzornih kamera, poželjno je identitete osoba ukloniti sa snimki.
Najjednostavnija metoda za to je de-identifikacija potpomognuta umjetnom inteligencijom i računalnim vidom. Postoji mnogo metoda, neke veoma jednostavne, druge poprilično zahtjevne za implementaciju, te je potrebno odabrati najbolju s obzirom na okolnosti.
Snimke sa nadzornih kamera nisu poznate po svojoj visokoj kvaliteti, pa je testiranje provedeno na skupu slika koje su lošije kvalitete, te time bitno utječu na rezultate testiranja.
Zbog njihove inače poprilično jednostavne nulifikacije mnogi otpišu jednostavne metode de-identifikacije (tzv. naivne metode) poput zamućivanja lica (ugl Gaussovim filterom) ili pikselizacije, ali u slučaju kao što je ovaj, kad su u pitanju niskokvalitetne slike, takav pristup možda se isplati zbog njegove jednostavne implementacije i komputacijske ne zahtjevnosti.
U nekoliko slučajeva ispitana je učinkovitost naivne de-identifikacije nasuprot na FaceNet-u temeljene InceptionResnet V1 mreže za prepoznavanje lica. Pokazalo se da ni uz upotrebu inače učinkovitih metoda reverziranja naivnih metoda de-identifikacije, model za prepoznavanje lica nije imao mnogo uspjeha, baš zbog niske kvalitete slika koje su korištene.

Abstract (english)

Human privacy can be easily violated with surveillance camera footage that constantly captures public space, and therefore the people in it. Because the identity of persons is rarely important in the surveillance camera service, it is desirable to remove the identities of persons from the recordings.
The simplest method for this is de-identification aided by artificial intelligence and computer vision. There are many methods, some very simple, others quite demanding to implement, and it is necessary to choose the best one given the circumstances.
Surveillance camera images are not known for their high fidelity, so the testing was conducted on a set of images that are of poorer quality, and thus affect the test results.
Because of their otherwise fairly simple nullification, many write off simple de-identification methods (so-called naive methods) such as facial blur (Gaussian filter) or pixelation, but in a case like this, when it comes to low-quality images, such an approach may pay off due to its simple implementation and non-demanding computational complexity.
In several cases, the effectiveness of naive de-identification versus FaceNet-based InceptionResnet V1 face recognition network was examined. It turned out that even with the use of otherwise effective methods of reversing naive de-identification methods, the face recognition model did not have much success, precisely because of the low quality of the images used.

• Keywords: de-identifikacija u slikama; detekcija lica; prepoznavanje lica; računalni vid; MTCNN; ResNet
• Keywords (english): de-identification in images; face detection; face recognition; computer vision; MTCNN; ResNet
• Language: croatian
• URN:NBN: urn:nbn:hr:168:603150
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-14 08:41:59