20.500.12556/DKUM-70287 Primerjava segmentacijskih algoritmov na posnetkih zdravih in obolelih očeh Comparison of segmentation algorithms on images of healthy and sick human eyes V tej magistrski nalogi smo se ukvarjali s segmentiranjem človeškega očesa ter prepoz-navanjem bolezni na očesu. Preštudirali in primerjali smo obstoječe algoritme za segmen-tacijo očesa. V praktičnem delu naloge smo razvili lastno hibridno metodo, s pomočjo kate-re detektiramo človeško oko ter prepoznamo bolezni. Naš algoritem in sorodne metode smo preizkusili na treh testnih zbirkah slik s človeškimi očmi, in sicer na lastni zbirki ter na javnih zbirkah IITD in UTIRIS. Algoritme smo primerjali kvalitativno in s pomočjo indeksa Jaccard. Rezultati so pokazali, da s pomočjo naše hibridne metode dosežemo najboljše rezultate. Pri zaznavi zenice smo na lastni podatkovni zbirki dosegli 80 % natančnost, na testnih zbirkah IITD in UTIRIS pa 96 % oziroma 97 % natančnost. Pri zaznavi šarenice smo bili najbolj natančni na zbirki UTIRIS (89 % natančnost), sledita pa lastna zbirka (75 %) ter zbirka IITD (59 %). Eksperimentalno smo potrdili domnevo, da s kombinacijo splo-šnih segmentacijskih postopkov in postopkov na osnovi geometrijskih modelov izboljšamo natančnost segmentacije na posnetkih zdravih in obolelih oči. Bolezni oči smo prepozna-vali le na lastni zbirki. Metrika priklic je variirala od 38 % za bolezen Arcus senilis do 93 % za bolezen miozo. Podobno smo opazili pri metriki točnost, ki je variirala od 60 % (vnetje šarenice), pa vse do 92 % (siva mrena). Preliminarni rezultati potrjujejo uspešnost našega pristopa. In this master's thesis we were focused on image segmentation techniques and diseases recognition of the human eye. We have researched and compared existing algorithms for eye segmentation. In the practical part of the master thesis we developed our own hybrid method with the help of which we can detect the human eye and recognize diseases. Our algorithm and related methods were tested on three human eyes image collections: our collection, IITD and UTIRIS. The algorithms were compared qualitatively and with Jaccard index. The results showed that we achieve best results with our hybrid method. At pupil detection we achieved 80% accuracy on our own collection, 96% accuracy on IITD and 97% accuracy on UTIRIS image collection. In detecting the iris, we got the best results on the UTIRIS (89% accuracy), followed by our own image collection (75%) and the IITD (59%). Experimentally, we have confirmed hypothesis that with the combination of general segmentation procedures and procedures based on geometric models we improve the accuracy of image segmentation on healthy and diseased eyes. We tested eye disease recognition procedure only on our own collection. The recall rate varied from 38% for Ar-cus senilis to 93% for the Miosis. Precision varied from 60% (Iritis) up to 92% (Cataract). Preliminary results confirm the success of our approach. napredna obdelava slik segmentacija očesa segmentacijski postopki zaznava bolezni človeškega očesa. image processing eye segmentation image segmentation algorithms detection of eye diseases. true true false S. Krivec Slovenski jezik Angleški jezik Magistrsko delo/naloga 2018-04-25 23:44:25 2018-06-14 08:00:43 2022-08-01 19:02:06 2019-07-11 18:02:03 2018 0 [Maribor 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 004.932:617.7(043.2) 21498134 URN:SI:UM:DK:AXTQSPKO MAG_Krivec_Sandi_2018.pdf MAG_Krivec_Sandi_2018.pdf 1 DCEF1E640769C7F9ACD4514490147D53 e01415466e8273691782edf2777f417a961755eeb601c6f550d1b4b16b26afd5 04d75609-7c10-11eb-bb7a-00155d0001ca 20.500.12556/dkum/1d3a7a36-070f-4fb0-ba0c-85a96609d483 https://dk.um.si/Dokument.php?lang=slv&id=125527 Fakulteta za elektrotehniko, računalništvo in informatiko 0 0 0