Detekcija napak na odlitkih z globokim učenjem

magistrsko delo

Tomo Pšeničnik (Author), Simon Klančnik (Mentor), Božidar Bratina (Mentor)

Abstract

Cilj magistrske naloge je preučiti detekcijo napak na odlitkih z uporabo konvolucijskih nevronskih mrež. Predstavljena je klasifikacija slik dobrih in slabih odlitkov, ki temelji na globokem učenju. Za učenje nevronske mreže smo uporabili obstoječo zbirko podatkov, ki vsebuje več kot 7000 slik. Za izdelavo programa smo uporabili okolje Matlab s pomočjo Deep learning toolbox vmesnika. Izdelali smo model konvolucijske nevronske mreže, izvedli učenje in prikazali rezultate. V drugem delu smo rezultate želeli izboljšati, zato smo se poslužili tehnike s prenosnim učenjem. Našim potrebam smo prilagodili obstoječo AlexNet arhitekturo, naložili zbirko podatkov in izvedli učenje nevronske mreže. Na koncu prikažemo rezultate kot je klasifikacijska točnost modela. Delovanje modela preizkusimo še na testni množici slik, katere niso bile vključene v proces učenja.

Keywords

Globoko učenje;detekcija napak;klasifikacija;konvolucijska nevronska mreža;odlitek;magistrske naloge;

Data

Language:	Slovenian
Year of publishing:	2022
Typology:	2.09 - Master's Thesis
Organization:	UM FS - Faculty of Mechanical Engineering
Publisher:	[T. Pšeničnik]
UDC:	004.85:621.747.019(043.2)
COBISS:	151766019
Views:	206
Downloads:	27
Average score:	0 (0 votes)
Metadata:

Other data

Secondary language:	English
Secondary title:	Castings defect detection using deep learning
Secondary abstract:	The aim of the master's thesis is to study defect detection on castings with the use of convolutional neural networks. Classification of good and bad castings that works on the principle of deep learning is presented. We use an existing large database that consists of more than 7000 pictures to train the neural network. Our convolutional neural network model was designed in Matlab with the help of Deep learning toolbox. We designed our convolutional neural network model, trained it and displayed the results. We wanted to improve the results, so we tried the transfer learning technique. We modify an existing AlexNet model to fit our application, load the dataset and train the new model. At the end we show the results such as classification error of our model. We test the models accuracy on some pictures that were not included in the process of training.
Secondary keywords:	Deep learning;defect detection;classification;convolutional neural network;casting;
Type (COBISS):	Master's thesis/paper
Thesis comment:	Univ. v Mariboru, Fak. za elektrotehniko, računalništvo in informatiko, Mehatronika
Pages:	1 spletni vir (1 datoteka PDF (X, 36 f.))
ID:	16833777