Prepoznavanje objektov iz satelitskih slik z metodami globokega učenja na vgrajeni napravi : diplomsko delo

Domajnko, Martin

| | SLO | ENG | Cookies and privacy

Bigger font | Smaller font

First page > Show document

Show document

Title:	Prepoznavanje objektov iz satelitskih slik z metodami globokega učenja na vgrajeni napravi : diplomsko delo
Authors:	ID Domajnko, Martin (Author) ID Potočnik, Božidar (Mentor) More about this mentor... ID Haložan, Jernej (Comentor)
Files:	UN_Domajnko_Martin_2021.pdf (34,13 MB) MD5: 9B42ED11A910DED7947D4352C3465ED0 PID: 20.500.12556/dkum/11827250-dd72-45f3-8f61-97dbdb6049cb
Language:	Slovenian
Work type:	Bachelor thesis/paper
Typology:	2.11 - Undergraduate Thesis
Organization:	FERI - Faculty of Electrical Engineering and Computer Science
Abstract:	V diplomskem delu rešujemo problem prepoznavanja prometa iz satelitskih slik. Cilj je bil uporabiti metode globokega učenja, pognati modele na izbranih vgrajenih napravah in doseči povprečno natančnost vsaj 75 % pri hitrosti izvajanja 5 sličic na sekundo. Za eksperiment uporabimo modela Faster R-CNN in SSD iz knjižnic Detectron2 ter TensorFlow Object Detection API. Fazi učenja in testiranja izvedemo na satelitskih slikah baze podatkov xView, katere predhodno razdelimo na učno in testno množico. Na učni množici izvedemo tudi bogatenje slik. Naučene modele preizkusimo na grafičnih karticah Nvidia GeForce GTX 970 ter Nvidia Titan X Pascal, na procesorju Intel Core i7-4790 in na vgrajenih napravah Intel Neural Compute Stick 2 ter Nvidia Jetson TX2. Preizkuse izvedemo s pomočjo skript napisanih v programskem jeziku Python3. Te izvozijo modele v posebno zamrznjeno stanje, jih optimizirajo za izvajanje na izbrani napravi in izmerijo njegovo hitrost ter natančnost. Najvišjo povprečno natančnost 37,33 % dosežemo z modelom Faster R-CNN iz knjižnice Detectron2. Z modelom SSD iz knjižnice TensorFlow Object Detection API na grafični kartici Nvidia GeForce GTX 970 dosežemo povprečno hitrost izvajanja 84,5 sličic na sekundo. Demonstrirana rešitev v diplomskem delu je primerna za izvajanje na vgrajenih napravah, a žal ni dovolj natančna. Za doseganje boljših rezultatov moramo našo rešitev izvajati na hitrejši strojni opremi, ki podpira večje ter s tem natančnejše modele.
Keywords:	strojno učenje, globoko učenje, vgrajene naprave, prepoznavanje objektov, satelitske slike, računalniški vid
Place of publishing:	Maribor
Place of performance:	Maribor
Publisher:	[M. Domajnko]
Year of publishing:	2021
Number of pages:	X, 40 str.
PID:	20.500.12556/DKUM-80246
UDC:	004.85:004.932.75\'1(043.2)
COBISS.SI-ID:	87069187
Publication date in DKUM:	18.10.2021
Views:	1240
Downloads:	136
Metadata:
Categories:	KTFMB - FERI
:	Copy citation

Average score:	(0 votes)
Your score:	Voting is allowed only for logged in users.
Share:

Hover the mouse pointer over a document title to show the abstract or click on the title to get all document metadata.

Licences

License:	CC BY-NC-ND 4.0, Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International

Link:	http://creativecommons.org/licenses/by-nc-nd/4.0/
Description:	The most restrictive Creative Commons license. This only allows people to download and share the work for no commercial gain and for no other purposes.
Licensing start date:	03.09.2021

Secondary language

Language:	English
Title:	Object detection in satellite images with deep learning methods on embedded system
Abstract:	The problem of detecting traffic in satellite images is being solved in this thesis. The goal was to use deep learning methods, run the models on selected embedded systems and achieve a mean average precision of at least 75% at a processing speed of 5 frames per second. Models Faster R-CNN and SSD implemented in Detectron2 and TensorFlow Object Detection API libraries were used. The training and testing phases were done on satellite images from the xView dataset, which were previously separated into training and testing splits. Image augmentation was also done on the training split. Trained models were tested on Nvidia GeForce GTX 970 and Nvidia Titan X Pascal GPUs, on the Intel Core i7-4790 CPU and on Intel Neural Compute Stick 2 and Nvidia Jetson TX2 embedded systems. The tests were done with scripts written in the programming language Python3. The models were exported to a frozen state, optimized for the chosen hardware, and their speed and accuracy was measured by the scripts. The highest mean average precision 37.33% was achieved by the Faster R-CNN model implemented in the Detectron2. Also, a processing speed of 84.5 frames per second was achieved on the Nvidia GeForce GTX 970 GPU with the model SSD implemented in the TensorFlow Object Detection API library. The solution demonstrated in the thesis is suitable for running on embedded systems, but unfortunately it is not accurate enough. To achieve better results, our solution needs to be run on faster hardware that supports larger and thus more accurate models.
Keywords:	machine learning, deep learning, embedded systems, object detection, satellite images, computer vision

Comments

Leave comment

You must log in to leave a comment.

Comments (0)

0 - 0 / 0

There are no comments!

Back