Primerjava slikovnih deskriptorjev za samodejno razpoznavanje obrazov

GODEC, JAN (author), Štruc, Vitomir (mentor)

V diplomskem delu je predstavljena analiza računalniškega razpoznavanja obrazov z različnimi postopki. Funkcija, ki jo je opravljal biometrični sistem, je bila verifikacija oseb na podlagi razpoznavanja obrazov. Podano je teoretično ozadje, orodja in metode za izvedbo ter rezultati eksperimentov. Razpoznavanje je bilo izvedeno na sivinskih slikah obrazov iz dveh standardnih podatkovnih zbirk, in sicer XM2VTS in EYB. V prvi zbirki so bili zajeti obrazi brez večjih motenj v osvetlitvi, v drugi zbirki pa se je postopoma večala zatemnitev obrazov. Na zbirki XM2VTS so bili preizkušeni PCA, SURF, FREAK, BRISK, SURF+PCA, FREAK+PCA in BRISK+PCA postopek. Na zbirki EYB so bili preizkušeni PCA, SURF, FREAK in BRISK postopek. Pri vseh postopkih se je razdalja med značilnimi vektorji merila na tri načine, in sicer z Evklidovo, City Block ter kosinusno razdaljo. Rezultati razpoznavanja obrazov so prikazani v obliki tabel in grafov. Ovrednoteni so z napakami FAR, FRR, HTER in EER. Soodvisnost med FAR in FRR v različnih delovnih točkah ponazarjajo ROC krivulje. V tabelah so izpostavljeni rezultati verifikacije v treh delovnih točkah. Prikazana je tudi časovna zahtevnost izračuna posameznih značilnih vektorjev.

• Language: Slovenian
• Keywords: razpoznavanje obrazov, verifikacija oseb, PCA, SURF, FREAK, BRISK, deskriptor, XM2VTS, EYB, ROC krivulja, Evklidova razdalja, City Block, kosinusna razdalja
• Work type: Undergraduate thesis
• Organization: FE - Faculty of Electrical Engineering
• Year: 2016
• PID: 20.500.12556/RUL-84508
• Publication date in RUL: 26.08.2016

Secondary language

• Language: English
• Title: Comparative evaluation of image descriptors for automatic face recognition

Abstract

The diploma thesis presents an analysis of computer facial recognition with various procedures. The function performed by the biometric system was face verification on the basis of facial recognition. Theoretical background, tools, and methods for the experiments and results thereof are also presented. Recognition was performed on grey pictures of faces from two standard databases, specifically XM2VTS and EYB. The first database included faces without any significant lighting interference, whereas the pictures in the second database were increasingly darker. PCA, SURF, FREAK, BRISK, SURF+PCA, FREAK+PCA, and BRISK+PCA were tested on the XM2VTS database, whereas PCA, SURF, FREAK, and BRISK were tested on the EYB database. In all procedures, the distance between feature vectors was measured using three methods, specifically the Euclidean, City Block, and cosine distance. The results of facial recognition are shown in the form of tables and graphs. They are evaluated with FAR, FRR, HTER, and EER values. The correlation of FAR and FRR at various operating points are shown by ROC curves. The tables highlight the verification results in three operating points. The time requirement for the calculation of specific feature vectors is also shown.

• Keywords: facial recognition, face verification, PCA, SURF, FREAK, BRISK, descriptor, XM2VTS, EYB, ROC curve, Euclidean distance, City Block, cosine distance