Segmentation of openings in non-crimp fabric using deep learning

Non-Crimp Fabrics (NCF) are widely used in industries like aerospace due to their lightweight and high-strength properties. Accurate characterization of openings in NCF is essential for maintaining material quality and performance. Traditional machine vision methods, such as histogram-based and edge...

Full description

Bibliographic Details
Main Authors: Md Ali, Syaril Azrad, Norhellme, Nornajme, Ahmad Rashidi, Syazwan, Mahmud Zuhudi, Nurul Zuhairah, Abd Aziz, Noor Zuhaira
Format: Article
Language:English
Published: Aeronautical and Astronautical Society of the Republic of China 2024
Online Access:http://psasir.upm.edu.my/id/eprint/117868/
http://psasir.upm.edu.my/id/eprint/117868/1/117868.pdf
Description
Summary:Non-Crimp Fabrics (NCF) are widely used in industries like aerospace due to their lightweight and high-strength properties. Accurate characterization of openings in NCF is essential for maintaining material quality and performance. Traditional machine vision methods, such as histogram-based and edge detection approaches, often struggle with complex NCF patterns and rely heavily on manual feature engineering. This study applies deep learning models U-Net, LinkNet, and DeepLabv3+ for the binary segmentation of openings in carbon fiber NCF. These models use a ResNet18 encoder, pre-trained on ImageNet, to automatically extract robust features and improve generalization. To address data limitations, we augmented the collected NCF dataset to increase variability. The results show that U-Net, LinkNet, and DeepLabv3+ achieved Intersection over Union (IoU) scores of 0.7419, 0.7228, and 0.7172, respectively, demonstrating their effectiveness in segmenting openings in NCF. This study presents a method that enhances segmentation accuracy and improves ability to generalize across varying data conditions, providing a more adaptable approach to NCF characterization with potential applications for other advanced materials.