20.500.12556/RUL-81161 Ovrednotenje metod za samodejno zaznavanje obrazov Comparative assessment of face detection techniques V pričujočem delu so predstavljeni problemi pri zaznavanju obrazov v slikah ter nekaj pristopov, kako te probleme rešiti. Opisan je sistem FDDB kot način enotnega sistema za ovrednotenje in primerjavo med različnimi pristopi zaznave. Izpostavljene so nekatere implementacije prosto dostopnih akademskih metod in brezplačne različice komercialnih rešitev. Med akademskimi metodami je povzeto ogrodje za zaznavanje objektov Viola-Jones, njegove izvedbe s Haarovo, LBP in SURF kaskado, rešitve z uporabo DPM, HOG piramide in SVM razvrščevalnikom ter primerjave svetilnosti pikslov. Predstavljene komercialne rešitve zajemajo Face++, ki koristi globoko učenje in nevronske mreže ter rešitve BetaFace, MS Project Oxford in Apple Photos, za katere metoda obdelave ni javno dostopna. Iz rezultatov je razvidno, da je možno s skrbno izbiro parametrov odprtokodnih rešitev dobiti vsaj primerljive rezultate s komercialnimi. Po primerjavi potrebnega časa za zaznavanje je mogoče zaključiti, da se najbolje obnese metoda pico. Najbolj natančen rezultat pa dajeta metodi HOG v izvedbi knjižnice dlib in DPM v izvedbi knjižnice voc-dpm. The paper introduces a few major issues with image face detection and presents some solutions to tackle them. It sets out with a description of the FDDB system as a means of evaluation and comparison of different approaches in this field. It furthermore examines some freely available implementations of academic methods and free versions of commercial solutions. Among the academic methods the paper concentrates on analysing the Viola-Jones object detection framework along with different implementations of the detection cascades based on Haar, LBP and SURF features. In addition, it describes solutions which use the DPM, HOG pyramids with a SVM classifier and pixel-intensity classifier. The selected commercial solutions presented in the paper include Face++ (which utilizes deep learning and neural network algorithms), BetaFace, MS Project Oxford and Apple Photos (which do not publicly reveal the method of detection). Finally, the paper concludes with establishing that a careful parameter selection of open-source solutions can bring about at least comparable results to those achieved with the commercial solutions. The comparison of detection time determines pico as the fastest method that still yields useful results while the HOG method implemented with dlib and the DPM method implemented with voc-dpm are found to give the most accurate results. zaznavanje obrazov kaskadni rzvrščevalnik primerjava zaznavalnikov Viola-Jones zaznavalnik objektov LBP SURF HOG DPM OpenCV libccv dlib pico FDDB face detection cacade classifier detector comparison Viola-Jones object detector LBP SURF HOG DPM OpenCV libccv dlib pico FDDB true false false Slovenski jezik Angleški jezik Diplomsko delo 2016-03-30 15:05:02 2016-03-30 15:05:05 2022-08-01 21:38:51 0000-00-00 00:00:00 2016 0 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 1970-01-01 34124 De_reggi_Andreas_luka_-_Ovrednotenje_metod_za_samodejno_zaznavanje_obrazov.pdf De_reggi_Andreas_luka_-_Ovrednotenje_metod_za_samodejno_zaznavanje_obrazov.pdf 1 D9030AD9D2FE170154BC3D9682EB8245 17d31801e052c033938e463fd4df08e1e11bffc6471694d374b72558ac9f8ce2 4bfe9df6-a1a9-11eb-a523-00155dcfd717 20.500.12556/rul/e1c40cdb-ed99-4453-8a3b-6719710384ef https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=81378 Fakulteta za elektrotehniko 0 0 0