izpis_h1_title_alt

Computer-vision based polyp detection in underwater images : bachelor's thesis
ID Shirgoski, Kristijan (Author), ID Kristan, Matej (Mentor) More about this mentor... This link opens in a new window

.pdfPDF - Presentation file, Download (10,74 MB)
MD5: 37035B9D6EDA13DEB86190C3E2567E64
PID: 20.500.12556/rul/e7ec158e-17a1-47f1-a066-b83edb1d088c

Abstract
Object detection is a popular topic in computer vision and machine learning. Numerous approaches have been proposed in literature to address the challenging task of general object detection. The approaches vary conceptually as well as in the level of computational intensity. The goal of this thesis was to develop a pipeline of state-of-the-art algorithms to detect polyps of the Aurelia aurita jellyfish, which are densely spread across corals. In object detection problems, a mandatory task is searching the image for regions of interest, preferably of several sizes. We propose a trained aggregated channel features (ACF) model to do that. In order to later classify these regions, first they need to have some features or characteristics extracted from them. In this thesis, this is performed by a convolutional neural network (CNN) trained on the MNIST dataset. Furthermore, a binary support vector machines (SVM) classifier with linear kernel and L2-regularized logistic regression is used to classify the features and determine the probability of correctly classifying them. It is very likely that several regions are proposed for each ground truth, so the regions must undergo a non-maximum suppression which uses the probability outputs from the logistic regression to group the local regions together, greedily prioritizing based on the probability distribution. The algorithms were trained and tested on 35 images consistent of nearly 40000 rectangle annotations from a newly annotated dataset. We have achieved very promising results and analyzed the strengths and weaknesses of our approach.

Language:English
Keywords:object detection, computer vision, machine learning, ACF, CNN, SVM, regression, non-maximum suppression
Work type:Bachelor thesis/paper
Typology:2.11 - Undergraduate Thesis
Organization:FRI - Faculty of Computer and Information Science
Publisher:[K. Shirgoski]
Year:2015
Number of pages:68 str.
PID:20.500.12556/RUL-73041 This link opens in a new window
COBISS.SI-ID:1536643523 This link opens in a new window
Publication date in RUL:07.10.2015
Views:1852
Downloads:314
Metadata:XML DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:Slovenian
Title:Zaznavanje polipov z računalniškim vidom v podvodnih slikah : diplomsko delo
Abstract:
Detekcija objektov je priljubljena tematika na področjih računalniškega vida in strojnega učenja. Reševanje problemov detekcije objektov predstavlja precejšenj izziv in v literaturi obstaja veliko različnih pristopov. Pristopi se konceptualno precej razlikujejo, različne pa so tudi njihove časovne zahtevnosti. Cilj diplomske naloge je ustvariti splošen pristop za zaznavanje objektov v sliki, v našem primeru polipe meduze Aurelia aurita. Za te polipe je značilno, da se na gosto širijo preko koral. Ena od značilnih tematik pri detekciji objektov je iskanje določenih regij v sliki, ki nas zanimajo. Ponavadi iščemo regije različnih velikosti. Zato smo predlagali uporabo modela agregirane značilnice po kanalih (ACF), ki se je učil na anotaciji iz naše zbirke. Iz vsake regije moramo izluščiti podatke, na podlagi katerih posamezni regiji določimo lastnosti ali karakteristike. V tej diplomski nalogi smo to izvedli s pomočjo konvolucijske nevronske mreže (CNN), ki je bila trenirana na podatkovni zbirki MNIST. Poleg tega sta za klasifikacijo in ocenjevanje njene pravilnosti uporabljena binarni klasifikator podpornih vektorjev (SVM) z linearnim jedrom ter L2 regularizirana logistična regresija. Zelo verjetno je, da vsaki anotaciji pripada več regij, ki jih ACF predlaga. Zato je potrebno veliko teh regij odstraniti s pomočjo metode tlačenja nemaksimumov, ki se naivno osredotoča na tiste regije, ki imajo višje ocene. Algoritme, ki smo jih uporabljali, so bili učeni in preizkušeni na novi zbirki podatkov, ki je sestavljena iz skoraj 40000 pravokotnih anotacij v 35 slik. Dosegli smo zelo obetavne rezultate ter analizirali prednosti in slabosti našega pristopa.

Keywords:detekcija objektov, računalniški vid, strojno učenje, ACF, konvolucijska nevronska mreža, metoda podpornih vektorjev, tlačenje nemaksimumov

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Back