Primerjava postopkov svetlobnega normiranja za samodejno razpoznavanje obrazov

PASTORČIČ, DINO (author), Štruc, Vitomir (mentor)

Razpoznavanje človeških obrazov se je v zadnjih letih pokazalo za eno od najpomembnejših področji analize slik. Sodobni sistemi razpoznavanja obrazov so sposobni identificirati in preveriti identiteto oseb na podlagi slik ali video posnetkov, pridobljenih iz digitalnih virov. Danes se sistemi za razpoznavanje obrazov uporabljajo v številnih aplikacijah na področju varnosti in zgodnje detekcije potencialnih osumljencev. Vse več pa se tovrstni sistemi uporabljajo na velikih javnih površinah, kot so letališča, nogometni stadioni, kjer lahko izkoriščamo ključno prednost sistemov, in sicer, da ne potrebujejo aktivnega sodelovanje opazovane osebe. Med glavne težave v uspešni implementaciji sistemov razpoznavanja obraza so spremembe osvetlitvenih pogojih in spremembe v položaju obraza opazovane osebe glede na kot kamere. Aldini et al. so v svojem delu dokazali, da so razlike med slikami istega obraza zaradi sprememb jakosti in kota osvetlitve pogosto večje kot razlike med slikami različnih obrazov [1]. V našem delu smo se odločili teoretično predstaviti, analizirati in primerjati metode svetlobnega normiranja in njihovo učinkovitost pri uporabi v sistemih za razpoznavanje obrazov. V ta namen smo uporabili pet različnih postopkov svetlobnega normiranja slik in jih implementirali na razširjeni Yale B zbirki obrazov. Uporabljene metode so retinex algoritem (angl. Single scale retinex algorithm), metoda samo-kvocientne slike (angl. Single scale self quotient image), anizotropna difuzijska metoda svetlobnega normiranja (angl. Anisotropic diffusion based normalization technique), svetlobno normiranje, zasnovano na Weberjevih obrazih (angl. Single scale Weberfaces normalization technique) in Tan in Triggs metoda svetlobnega normiranja (angl. Tan and Triggs normalization technique). Preizkusili smo tudi metodo ojačenja kontrasta, ki temelji na izravnavi histograma (angl. Histogram equalization), in njen vpliv na algoritme svetlobnega normiranja. Uspešnost metod smo ocenili z uporabo linearne diskriminantne analize (angl. Linear discriminant analysis, LDA). Uspešnost delovanja biometričnega sistema smo ocenili v identifikacijskem in verifikacijskem načinu delovanja. Rezultati so predstavljeni v obliki rang 1 rezultatov identifikacije in krivulj karakteristike delovanja sprejemnika (angl. Receiver operating characteristic curve) ter z deležem lažnih pozitivnih primerov (FAR 1 % in 0,1 %) v verifikacijskem načinu delovanja. Algoritmi so implementirani v Matlab programskemu okolju z uporabo INFace in PhD programskih orodji, razvitih v Laboratoriju za umetno zaznavanje, sisteme in kibernetiko Fakultete za elektrotehniko Univerze v Ljubljani.

• Language: Slovenian
• Keywords: razširjena Yale B zbirka obrazov, svetlobno normiranje obraznih področij, histogramska izravnava, retinex algoritem, Tan in Triggs normiranje, anizotropna difuzijska metoda, metoda Weberjevih obrazov, samo-kvocientna slika.
• Work type: Undergraduate thesis
• Organization: FE - Faculty of Electrical Engineering
• Year: 2016
• PID: 20.500.12556/RUL-83818
• Publication date in RUL: 01.07.2016

Secondary language

• Language: English
• Title: Comparative assessment of photometric normalization techniques for automatic face recognition

Abstract

Face recognition has recently become one of the most important areas of image processing. The current face recognition systems are capable of recognizing and identifying the subject of interest based on digitally acquired images or recordings. Face recognition systems are being applied in numerous ways in security and early detection of suspicious individuals. The use of these systems is also increasing in large public places such as airports, stadiums, etc. Here, the key advantage of face recognition systems compared to other biometrics can be exploited, as such systems do not require the cooperation of the tested subject. The main difficulties in the successful implementation of face recognition systems are changes in illumination conditions and subject pose variation. Aldini eta al. in [1] proved that the variation between the images of the same face due to illumination and viewing direction are almost always larger than variations due to changes of the subject’s identity. In this thesis, we analyzed and compared the photometric normalization techniques for face recognition applications. Five different preprocessing algorithms have been tested on the extended Yale Face Database B, namely, the single scale retinex algorithm, the single scale self-quotient image, the anisotropic diffusion based normalization technique, the single scale Weberfaces normalization technique and the Tan and Trigg’s normalization technique. Furthermore, we tested the influence of histogram equalization image enhancement technique on the above mentioned photometric normalization techniques. In the next step, we used the LDA (Linear Discriminant Analysis) algorithm to identify the best illumination invariant methods. The biometrical system was assessed in the identification and verification operating mode. The results are presented in the form of rank 1 results of the identification and the receiver operating characteristic curve, as well as a percentage of false positive cases (FAR 1 % and 0,1 %) in the verification operating mode. The algorithms are implemented in Matlab, with the use of the INface and PhD toolboxes which were developed at the Laboratory of Artificial Perception, Systems and Cybernetics at the Faculty of Electrical Engineering, University of Ljubljana.

• Keywords: extended Yale Face Database B, photometric normalization techniques, histogram equalization, single scale retinex algorithm, Tan and Triggs normalization technique, anisotropic diffusion based normalization technique, single scale Weberfaces normalization technique, single scale self-quotient image.