Raspoznavanje lica u domeni kompresije

doctoral thesis

• Title (english): Face Recognition in Compressed Domain
• Author: Krešimir Delač
• Mentor: Mislav Grgić (mentor)
• Committee member: Sven Lončarić (predsjednik povjerenstva)
• Committee member: Mislav Grgić (član povjerenstva)
• Committee member: Dragan Gamberger (član povjerenstva)
• Committee member: Borivoj Modlic (član povjerenstva)
• Committee member: Branka Zovko-Cihlar (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Communication and Space Technologies); Zagreb
• Defense date and country: 2007-03-05, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Electrical Engineering; Radio Communications
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing; Data Processing
• Universal decimal classification (UDC): 621.3 - Electrical engineering ; 004 - Computer science and technology. Computing. Data processing

Abstract

U ovoj disertaciji istraživanje je usmjereno na uporabu JPEG i JPEG2000 normi za kompresiju slika u sustavima za raspoznavanje lica. Kompresija je nužna da bi se smanjio prostor za pohranu slika. Kako bi se kompresija slika mogla koristiti u sustavima za raspoznavanje lica, potrebno je ispitati utjecaj koji izobličenja uzrokovana kompresijom imaju na točnost raspoznavanja. Razvijena je metodologija za objektivnu usporedbu točnosti metoda za raspoznavanje lica korištenjem komprimiranih i nekomprimiranih slika te je pokusom utvrđeno da se točnost raspoznavanja ne mijenja značajno. Zaključeno je da je, budući da kompresija ne narušava točnost raspoznavanja, opravdano cijeli sustav za raspoznavanje lica pokušati implementirati u domeni kompresije. Rad u domeni kompresije podrazumijeva uporabu transformacijskih koeficijenata na ulazu u metode raspoznavanja, umjesto uporabe elemenata izvorne slike. Proveden je pokus u istim radnim uvjetima uporabom transformacijskih koeficijenata dobivenih iz postupka kompresije te je utvrđeno da točnost raspoznavanja nije značajno smanjena u odnosu na točnost dobivenu uporabom nekomprimiranih slika. Uporabom transformacijskih koeficijenata, tj. izbjegavanjem inverzne transformacije prilikom dekompresije, postiže se znatna ušteda na složenosti proračuna. Dodatna ušteda na složenosti proračuna postiže se predloženom metodom odabira važnih transformacijskih koeficijenata. Predložena metoda temeljena je na mjerenju varijance pojedinog transformacijskog koeficijenta na pažljivo odabranom skupu slika za uvježbavanje sustava te zadržavanje samo koeficijenata sa najvišom varijancom. Pokusom je utvrđeno da predložena metoda ne smanjuje točnost raspoznavanja te da ju u mnogim slučajevima čak i povećava.

Abstract (english)

The main focus of this doctoral dissertation is the use of JPEG and JPEG2000 image compression scheme in face recognition systems. Image compression in necessary is order to reduce the image storage requirements. The basic precondition of using image compression in face recognition systems is that it should not significantly deteriorate system's recognition rate. Methodology for objective comparison of face recognition systems using both compressed and uncompressed images as input was developed and three face recognition methods were tested. It was concluded that compression does not deteriorate recognition rate significantly and efforts to implement face recognition into the compressed domain is thus justified. Working in the compressed domain means that transform coefficients are used instead of pixels. Using the described approach, an experiment was conducted in the same working conditions and using the same proposed methodology for objective comparison of face recognition systems. It was concluded that recognition rate is not significantly deteriorated when using transform coefficients as input to face recognition methods. By using transform coefficients, and thus avoiding inverse transformation, a significant computational complexity reduction is achieved. Additional computational complexity can be achieved when using the method for selecting significant transform coefficients proposed in this doctoral dissertation. The method is based on selecting only part of the transform coefficients by analyzing their variance across a carefully selected image set. It was experimentally confirmed that the proposed approach does not deteriorate recognition rate. Instead, it even improves it in many cases.

• Keywords: raspoznavanje lica; kompresija slika; domena kompresije; JPEG; JPEG2000; PCA; LDA; ICA
• Keywords (english): face recognition; image compression; compressed domain; JPEG; JPEG2000; PCA; LDA; ICA
• Language: croatian
• URN:NBN: urn:nbn:hr:168:575041
• Study programme: Title: Electrical Engineering and Computing; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti elektrotehnike i računarstva (Doktor znanosti elektrotehnike i računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2020-03-18 19:01:34