Klasificiranje slika Fisherovim vektorima izgrađenim nad slučajnim distribucijskim šumama značajki slike

master's thesis

• Title (english): Image classification with Fisher vectors over random density forests
• Author: Dino Pačandi
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2015-07-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

U okviru ovog rada proučena je i implementirana metoda korištenja Fisherovih vektora preko distribucijskih šuma za klasifikaciju slika. Motivacija korištenja distribucijskih šuma kao generativnog modela umjesto uvriježene Gaussove mješavine je ubrzanje samog postupka izračuna Fisherovih vektora.
Postignuto je ubrzanje izračuna od oko 8 puta naspram korištenja Gaussove mješavine, uz gubitak prosječne srednje preciznosti (mAP) klasifikacije od oko 3%. Također je i proučen utjecaj globalne i lokalne analize glavnih komponenata, broja stabala u šumi kao i broj listova u stablu na točnost klasifikacije. Dodatno, predstavljen je novi način treniranja distribucijskih stabala; EM distribucijska stabla.

Abstract (english)

In this work, usage of Fisher vectors over random density forests for image classification is studied and implemented. Usage of random density forests as generative models instead of Gaussian mixture model was motivated by the desire to reduce computation time of a single Fisher vector.
Computation time was reduced approximately 8 times versus Gaussian mixture usage, while loss of around 3% mean average precision (mAP) is reported. Effect of global and local PCA, number of trees per forest and number of leaves per tree on classification precision is also studied. Additionally, a novel approach to density tree training is introduced; EM density trees.

• Keywords: Fisherovi vektori; distribucijska stabla; distribucijske šume; generativni modeli; klasifikacija slika
• Keywords (english): Fisher vector; density trees; density forests; generative models; image classification
• Language: croatian
• URN:NBN: urn:nbn:hr:168:361801
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2018-02-06 11:02:34