Document Type

Thesis - Open Access

Award Date

2020

Degree Name

Master of Science (MS)

Department

Electrical Engineering and Computer Science

First Advisor

Kwanghee Won

Second Advisor

Sung Shin

Abstract

The visual inspection of a concrete crack is essential to maintaining its good condition during the service life of the bridge. The visual inspection has been done manually by inspectors, but unfortunately, the results are subjective. On the other hand, automated visual inspection approaches are faster and less subjective. Concrete crack is an important deficiency type that is assessed by inspectors. Recently, various Convolutional Neural Networks (CNNs) have become a prominent strategy to spot concrete cracks mechanically. The CNNs outperforms the traditional image processing approaches in accuracy for the high-level recognition task. Of them, U-Net, a CNN based semantic segmentation method, has been one of the most popular in the deep learning because of its excellent performance in open-source crack classification. Although the results of the trained U-Net look good for some dataset, the model still requires further improvement for the set of hard examples of concrete crack that contains the stain, waterspot, and small width crack. In this paper, we address the challenging problem of accurately detecting a thin concrete crack. We designed a U-Net like structure that has a contracting path and an expansive path to overcome this challenge and compared it to current models, including original U-Net and pyramid pooling module network. The proposed architecture utilizes multiple feature maps in a down-sampling path to obtain a higher pixel-level segmentation precision. The down-sampled feature is then up-sampled from the output of the pyramid pooling module [13], giving a binary crack and non-crack semantic segmentation. In the experiment, we have collected hard examples and evaluated the approach. Experimental results demonstrate that the proposed network outperforms the U-Net and a pyramid pooling module network in detecting a thin crack in a noisy environment.

Library of Congress Subject Headings

Concrete -- Cracking.
Neural networks (Computer science)
Computer vision.
Computer architecture.

Format

application/pdf

Number of Pages

38

Publisher

South Dakota State University

Rights

In Copyright - Non-Commercial Use Permitted
http://rightsstatements.org/vocab/InC-NC/1.0/

Share

COinS