Civil and Environmental Engineering
First Advisor
Chungwook Sim
Date of this Version
5-2024
Document Type
Article
Citation
A thesis presented to the faculty of the Graduate College at the University of Nebraska in partial fulfillment of requirements for the degree of Master of Science
Major: Civil Engineering
Under the Supervision of Professor Chungwook Sim
Lincoln, Nebraska, May 2024
Abstract
Concrete cracks and structural steel corrosion are two of the most common defects in bridges. Quantifying and classifying these defects provide bridge inspectors and engineers with valuable data for assessing deterioration levels. However, the bridge inspection process is typically a subjective, time intensive, and tedious task, as defects can be overlooked or in locations not easily accessible. Previous studies have investigated deep learning-based inspection methods, implementing popular models such as Mask R-CNN and U-Net. The architectures of these models offer certain advantages depending on the required task. This thesis aims to evaluate and compare Mask R-CNN and U-Net regarding their crack detection and segmentation performances. Both models utilize identical datasets for training and validation. These datasets are compiled from publicly available sources and from previous data collection efforts by a research team at the University of Nebraska-Lincoln. Detection accuracy is evaluated utilizing a labeled orthomosaic image of a pedestrian bridge located in Lincoln, Nebraska. Analyzing the effectiveness of the deep convolutional neural networks on a large-scale image, rather than local images, is more akin to how traditional inspections are completed. In addition, both models are evaluated for their capabilities through surface strain analysis on the same pedestrian bridge. The crack predictions of the models are comprised of binary pixel-wise classifications (crack or non-crack) where the Euclidean distance and centerline-to-centerline spacing were measured. These measurements were gathered to compute relative crack width and crack spacing for strain calculation. Average strain values for each of the three spans of the bridge were then calculated. The average strains obtained from U-Net mask prediction measurements for each span fell within the valid service stage, encompassing the range from cracking to yield stress. In contrast, Mask R-CNN yielded strain results indicated stress levels exceeding yield in those spans.
One of the largest publicly available structural steel corrosion datasets was also published as part of this thesis [at https://github.com/bennett-jackson/steel-corrosion-dataset]. Corrosion detection and evaluation has not been researched as extensively as concrete crack detection due to limitations in dataset accessibility. The published dataset consists of 3,423 images containing various levels of corrosion. These images were categorized into five condition states (good, fair, poor, severe, and weathering) based upon an existing corrosion annotation guideline. Classifying images in this manner aims to provide an organized dataset for future research.
Advisor: Chungwook Sim
Included in
Categorical Data Analysis Commons, Civil Engineering Commons, Digital Communications and Networking Commons, Other Civil and Environmental Engineering Commons, Structural Engineering Commons
Comments
Copyright 2024, Bennett Jackson. Used by permission