UAV-aided bridge inspection protocol through machine learning with improved visibility images. (1st July 2022)

• Record Type: Journal Article
• Authors: Jeong, Euiseok; Seo, Junwon; Wacker, James P.
• Is Part Of: Expert systems with applications. Volume 197(2022)
• Publication Date: 2022-07-01
• Subjects: Drone -- Bridge inspection -- Damage quantification -- Image processing -- Visibility improvement -- Computer vision -- Machine learning -- Convolutional Neural Network (CNN); Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2022.116791
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: A bridge inspection protocol with UAV and machine learning technologies was proposed. Indoor and field inspections were conducted to validate the efficiency of the proposed protocol. Various types of damage for both inspections were efficiently classified through machine learning. The indoor inspections revealed that the protocol was translatable to a bridge structure. The field inspections demonstrated that the protocol was capable of damage quantification accurately.a Abstract: This paper aims to introduce a new bridge inspection protocol using Convolutional Neural Network (CNN)-based machine learning in conjunction with improved visibility images acquired by Unmanned Aerial Vehicles (UAVs). With two UAVs, separate inspections following the proposed protocol were initially performed indoor to quantify the damage state of three concrete columns and four Cross-Laminated Timber (CLT) beams. The protocol using the two UAVs was also adopted to inspect an in-service four-span timber bridge in Pipestone, Minnesota in the United States. During damage identification, various types of visually detected damage were classified through CNN-based machine learning. For image visibility improvement, each image with damage was processed with appropriate adjustment of brightness, contrast, and sharpness to identify and measure the damage in an efficient way. The proposed protocol was found to be capable of bridge damage identification and measurement with an average error of 9.12% when compared to the direct measurements.