Detekcija uhljev s konvolucijskimi nevronskimi mrežami

GABRIEL, LUKA LAN (author), Peer, Peter (mentor)

Zaznavanje objektov na slikah je še zmeraj zahteven problem na področju računalniškega vida. Zaznavanje uhljev je v zadnjih letih postala popularna aplikacija zaznavanja objektov, z vedno večjim zanimanjem za identifikacijo ljudi glede na biometrijo uhlja. Kolikor vemo, se je problem zaznavanja uhljev do zdaj reševal s kombinacijami zaznavanja kože, zaznavanja robov, histogramov in algoritmi ujemanja predloge. V tem delu predstavimo metodo za detekcijo uhljev brez ujemanja predloge, z uporabo konvolucijske nevronske mreže, ki opravlja segmentacijo. S to metodo, ki je invariantna na kot, pod katerim je slika zajeta, obliko uhlja, barvo kože, osvetljitev, delno prekrivanje in dodatke na uhljih, smo uspeli natančno zaznati območje slike, kjer se uhelj nahaja. Nadalje, čas, potreben za zaznavo, se je zelo izboljšal v primerjavi z ostalimi metodami za reševanje enakega problema. Predvidevamo, da bo naša metoda uporabljena v orodju Annotated Web Ears Toolbox.

• Language: English
• Keywords: računalniški vid, segmentacija, konvolucijske nevronske mreže, detekcija uhljev
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2016
• PID: 20.500.12556/RUL-84351
• Publication date in RUL: 16.08.2016

Secondary language

• Language: Slovenian
• Title: Ear detection with convolutional neural networks

Abstract

Object detection is still considered a difficult task in the field of computer vision. Specifically, earlobe detection has become a popular application as the interest in human identification using earlobe biometry has increased. So far earlobe detection problem has been solved using a combination of skin detection, edge detection, segmentation by fusion of histogram-based k-means, and template matching algorithms. In this work we present a method of earlobe detection without template matching by using a convolutional neural network, performing image segmentation. With this method, which is invariant to angle at which the photo was taken, earlobe shape, skin color, illumination, occlusions, and earlobe accessories, we were able to accurately detect the area of the image, where an earlobe is present. Moreover, detection time was significantly improved when compared to other methods for solving the same task. We expect our method to be used in Annotated Web Ears Toolbox.

• Keywords: computer vision, segmentation, convolutional neural networks, earlobe detection