Duboko učenje za automatizirano spajanje i ispravljanje slika iz dronova temeljeno na dubokom učenju

master's thesis

• Title (english): Deep Learning for Automatized Correction and Joining of Images Taken from Drones
• Author: Marko Puk
• Mentor: Vladimir Čeperić (mentor)
• Committee member: Vladimir Čeperić (predsjednik povjerenstva)
• Committee member: Siniša Šegvić (član povjerenstva)
• Committee member: Mirko Poljak (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-14, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Tema ovog rada je ispravljanje nedostataka na slikama snimljenih dronom i njihovo povezivanje u mapu prostora. Rješenje se sastoji od detaljnog opisa dobivanja skupa podataka potrebnog za učenje neuronske mreže, konfiguracije konvolucijske neuronske mreže u arhitekturi autoenkodera i naposljetku izrade dekodera kojim se dobiva mapa prostora. Za kreiranje skupa podataka za učenje korištena je metoda augmentacije koja uz pomoć transformacija vrši promjene na slikama. Za potrebe krajnjeg izgleda arhitekture neuronske mreže najprije je napravljen običan autoenkoder prilagođen ovom problemu. Na temeljima te arhitekture napravljena su poboljšanja u vidu dodavanja potpuno povezanih slojeva kako bi autoenkoder bio sposoban ispravljati nedostatke na slikama. Završni dekoder dizajniran je na način da koristi latentne prostore pripadnih slika i spaja ih i rekonstruira u mapu prostora.

Abstract (english)

The topic of this paper is the correction of imperfections in images taken by a drone and their connection to a map. The solution consists of a detailed description on how to obtain the data set needed to learn the neural network, the configuration of the convolutional neural network architecture and finally the creation of a decoder which results in map. To create a set of learning data, the augmentation method was used, which makes changes to the images with the help of transformations. For the needs of the final solution of the neural network architecture, a simple autoencoder adapted to this problem was created firstly. Based on this architecture, improvements have been made in the form of adding fully connected layers to make the autoencoder able to correct image imperfections. The final decoder is designed to use the latent spaces of the associated images and merge them into a map.

• Keywords: Konvolucijska neuronska mreža; Autoenkoder; Augmentacija; Slika drona
• Keywords (english): Convolutional neural network; Autoencoder; Augmentation; Drone image
• Language: croatian
• URN:NBN: urn:nbn:hr:168:008329
• 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: 2023-01-13 15:58:24