Prepoznava zračnih žepov v kompozitnih strukturah na osnovi obdelave slik s konvolucijskimi nevronskimi mrežami

Možina, Nejc (author), Bračun, Drago (mentor)

Pri izdelavi kompozitnih izdelkov nastanejo zračni žepi, ki poslabšajo določene mehanske lastnosti kompozita. V nalogi je prikazan razvoj slikovnega sistema za pregledovanje kompozitnih izdelkov, ki vključuje osvetljevanje s strukturirano svetlobo, zajem slike osvetljene površine ter skeniranje izdelka. V nadaljevanju je zajet teoretični in eksperimentalni prikaz prepoznave zračnih žepov z uporabo konvolucijskih nevronskih mrež. Globoko učenje smo izvedli s pomočjo sistema YOLO, ki je namenjen prepoznavi objektov na slikah. Vsi pomožni programi za predobdelavo slik in ustvarjanje baze podatkov so bili razviti v programskem okolju Python. Končne uteži globokega učenja in delovanje zaznave zračnih žepov smo nato uspešno testirali na testnih slikah in ovrednotili dobljene rezultate.

• Language: Slovenian
• Keywords: kompozitne strukture, zračni žepi, zaznava napak, strojni vid, globoko učenje, konvolucijske nevronske mreže
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FS - Faculty of Mechanical Engineering
• Place of publishing: Ljubljana
• Publisher: [N. Možina ]
• Year: 2021
• Number of pages: XXI, 60 str.
• PID: 20.500.12556/RUL-127235
• UDC: 620.168:004.932:004.85(043.2)
• COBISS.SI-ID: 70023683
• Publication date in RUL: 27.05.2021

MOŽINA, Nejc, 2021, Prepoznava zračnih žepov v kompozitnih strukturah na osnovi obdelave slik s konvolucijskimi nevronskimi mrežami [online]. Master’s thesis. Ljubljana : N. Možina . [Accessed 31 March 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=127235

Secondary language

• Language: English
• Title: Detection of air voids in composite structures based on image processing with convolutional neural networks

Abstract

During the manufacture of composite products, air voids are formed which affect certain mechanical properties of the composite. This paper presents the development of an imaging system for viewing composite products, which includes structured light illumination, image acquisition of the illuminated surface, and scanning of the composite product. This is followed by a theoretical and experimental demonstration of air voids detection using convolutional neural networks. Deep learning was performed using the YOLO system, which is designed to recognize objects on images. All programs for image preprocessing and database generation were developed in the Python software environment. The final weights of the Deep Learning and the functioning of the air voids detection were then successfully tested on test images after the obtained results were evaluated.

• Keywords: composite structures, air pockets, detecting defects machine vision, deep learning, convolutional neural networks