Home // PATTERNS 2023, The Fifteenth International Conference on Pervasive Patterns and Applications // View article

Crack Detection Performance Using Nested U-Net Models

Authors:
Haifa Alhasson

Keywords: Convolutional Neural Networks, CRACK500 dataset, Deep Learning, Deep Neural Networks, Image Segmentation, Road Crack, Pavement Crack, Semantic Segmentation.

Abstract:
As smart cities have gained attention over the past few years, it has become an important research direction to find methods to continuously and automatically monitor the structure with the least amount of manpower. Pavement damage is the most significant factor and the most challenging feature of road maintenance. Many researchers employ deep convolutional neural networks to segment pavement cracks accurately. Specifically, U-Net with its extensions is one type of Fully Connected Convolutional Neural Network that can be applied to various road crack datasets. As a result, it is difficult to determine which method will be most effective for segmenting road cracks. Therefore, this study aims to evaluate different semantic segmentation models based on the frequency with which they are used to segment images in the real world. In this paper, we compare several models including U-net, U-net++, ResUnet++, and U-Net3+, and evaluate their results using publicly available dataset. The U-net+3 outperforms other architectures by (97%,0.56, 0.699, and 0.726 ) in accuracy and Jaccard index, F1, and F2 measures.

Pages: 18 to 23

Copyright: Copyright (c) IARIA, 2023

Publication date: June 26, 2023

Published in: conference

ISSN: 2308-3557

ISBN: 978-1-68558-049-0

Location: Nice, France

Dates: from June 26, 2023 to June 30, 2023