Učinkovito generiranje hipotetičnih slikovnih regij za detekcijo polipov

Zavrtanik, Vitjan (author), Kristan, Matej (mentor), Maver, Jasna (co-mentor)

V tej nalogi naslavljamo problem detekcije polipov v slikah. Moderne metode detekcije so pogosto sestavljene iz dveh sklopov. Najprej se na obravnavani sliki hitro generirajo hipotetične regije na mestih, kjer naj bi se nahajali iskani objekti, nato pa se na predlaganih lokacij izvede še verifikacija hipotez z močnim klasifikatorjem. V tem delu se osredotočamo na prvi korak detekcij in poskušamo optimizirati novo metodo za reševanje problema predlaganja regij za detekcijo polipov na slikah ostrig. Rezultate predlagane metode smo primerjali z drugimi metodami za predlaganje regij kot sta ACF in Selective Search. Predlagali smo modificirano mero za ocenjevanje uspešnosti, ki bazira na meri AR. Problem detekcije polipov na slikah je težko rešljiv zaradi veliko različnih faktorjev, ki vplivajo na izgled polipa, ki ga je potrebno prepoznati. Polipi so različnih velikosti, so različno orientirani, se lahko zaradi svoje prosojnosti zlijejo z ozadjem in se velikokrat tudi prekrivajo kar zanesljivo detekcijo posameznih polipov lahko onemogoči.

• Language: Slovenian
• Keywords: predlaganje regij, polipi, detekcija objektov
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2016
• PID: 20.500.12556/RUL-85566
• Publication date in RUL: 16.09.2016

Secondary language

• Language: English
• Title: Efficient region proposals for polip detection

Abstract

In this thesis we deal with the issue of polyp detection in images of oysters. Modern methods of object detection are often composed of two parts. Firstly we use a fast region proposal method to generate hypothetical regions in places where objects are located with a higher probability. We then use a strong classifier to verify the proposed hypothetical regions. We address the issue of the first part of the object detection pipeline and we propose a new region proposal method for the purpose of detecting polyps in images of oysters. We compared the results of our method with other region proposal methods such as ACF and Selective Search. We also propose a metric for region proposal method performance based on the AR metric. Polyp detection is a hard problem to solve due to the visual properties of individual polyps as they often vary in size, orientation and opacity which causes them to blend with their surroundings and can often make reliable detection difficult.

• Keywords: region proposal, polyps, object detection