diplomsko delo
Bor Černe (Avtor), Simon Dobrišek (Mentor)

Povzetek

V tem dokumentu je opisan postopek izdelave zaključnega dela na prvi stopnji visokošolskega strokovnega študija na fakulteti za elektrotehniko v Ljubljani. Cilj zaključnega dela je razpoznava napak pri predmetih na rentgenskih slikah. Predstavljajmo si, da je pred nami predmet, ki na zunaj izgleda popolnoma nepoškodovan in po predpisih. Pri takem bi si sigurno mislili, da je zmožen iti na prodajno police oziroma v dnevno uporabo. Kaj pa če je predmet od znotraj poškodovan, odkrušen, ipd.? Predmeta ne moremo odpreti, ker je iz enega kosa in ga ne moremo kar odpreti in pogledati vanj, zato tukaj nastane težava. Obstajajo pa tehnološke rešitve, ki so zmožne rešiti ta problem. To je uporaba metod računalniškega in strojnega vida, ki se razvijajo na področju umetne inteligence. Ampak kako bi videli notranjost? To se lahko izvede z zajemom rentgenske slike predmeta. Sprva je bilo potrebno pridobiti slike za podatkovno bazo s katero lahko začnemo učiti stroj. Te slike se je pridobilo z rentgenskim zajemom. Sprva je potrebno osnovne slike programersko izboljšati z različnimi filtri, da je možno iz njih razbrati čim več. Po tem se na slikah izvede maskiranje, saj izvajamo segmentacijo in je potrebno označiti kateri del nas zanima, da se ga stroj nauči napovedovati.. Za programski jezik, ki je bil za ta projekt v uporabi pri programiranju filtrov ter modela nevronskega omrežja, je bil izbran Python, za knjižnico, strojnega učenja pa PyTorch. Model, kateri je bil uporabljen za to delo je U-Net, saj kaže ene izmed najboljših rezultatov za segmentacijo slik. To se je tudi videlo pri rezultatih učenja, ki in kasnejši napovedi z tem modelom nevronskega omrežja. Rezultati napovedi so bili na koncu kar uspešni, vendar zaradi težavnosti vhodnih slik in natančnosti mask je uspešnost malo manjša kot željeno. Izguba se je manjšala z vsako iteracijo pri učenju kar nam je dalo vedeti, da se model U-Net trenira na vhodnih podatkih. Kljub temu, da so vhodne slike težavne, je napovedovalni čas zelo majhen, okoli 0.01s, kar je za uporabo v realnem svetu zelo uporabno.

Ključne besede

predobdelava slik;modeli nevronskih mrež;napovedovanje;PyTorch;visokošolski strokovni študij;Aplikativna elektrotehnika;diplomske naloge;

Podatki

Jezik: Slovenski jezik
Leto izida:
Tipologija: 2.11 - Diplomsko delo
Organizacija: UL FE - Fakulteta za elektrotehniko
Založnik: [B. Černe]
UDK: 004.932:004.8(043.2)
COBISS: 121534723 Povezava se bo odprla v novem oknu
Št. ogledov: 13
Št. prenosov: 8
Ocena: 0 (0 glasov)
Metapodatki: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Ostali podatki

Sekundarni jezik: Angleški jezik
Sekundarni naslov: ǂThe ǂuse of a U-Net model on radiographs to recognize object damages
Sekundarni povzetek: This document describes the process of producing the final work at the first stage of higher vocational studies at the Faculty of Electrical Engineering in Ljubljana. The goal of the final part is to identify errors in objects which are photographed with an X-Ray device. Let's imagine that there is an object in front of us that looks completely unscathed on the outside and by the regulations. I'm sure you'd think it was capable of going on store shelves or day-to-day use. What if the object is damaged from the inside, chipped, etc.? We cannot open an object because it's made out of one piece so we can't just take a look at insides, so there's a problem here. However, there are technological solutions that are capable of solving this problem. This is the done with the help of computer and machine vision methods developed in the field of artificial intelligence. But how would you see the inside? This can be done by captureing an X-ray of the object. At first, it was necessary to obtain images for a database with which we can start learning the machine for later predictions. With the help of my coworkers at INEA d.o.o. we captured these images from an X-Ray at Jožef Štefan Institute. First thing after using filters on these images which help us see it's details better and get better quality images, the masks had to be made, because we will be performing segmentation and it is necessary to indicate which area of the image is in our range of interest so that we can train the machine to predict it. Python was chosen as a programming language, and PyTorch as a library, for machine learning. The model used for this project is U-Net, because it is shown that it has one of the best accuracy of prediction for image segmentation. This was also seen in the results of training process and later shown when predictions were made. The results of predictions were quite successful in the end, but they could have been even better if a larger image database was used to train the model. Moreover due to the difficulty of images the results were not as accurate, but still in acceptable range. Loss was getting lower and lower with each iteration in training which lead us to knowing that the model is successfully training on the given data. The prediction time of around 0.01s is short and implementable in the real world.
Sekundarne ključne besede: image preprocessing;neural network model;predicting;PyTorch;
Vrsta dela (COBISS): Diplomsko delo/naloga
Študijski program: 1000315
Konec prepovedi (OpenAIRE): 1970-01-01
Komentar na gradivo: Univ. v Ljubljani, Fak. za elektrotehniko
Strani: XIV, 46 str.
ID: 16448512