Visual Inspection of Precast Concrete Bridge Using UAS Technologies
Junwon Seo, Euiseok Jeong, and James P. Wacker
Appears on pages(s):
bridge, inspection, image processing, image enhancement, machine learning, pretrained denoising neural network
This paper proposes that Unmanned Aerial System (UAS) technologies integrated with image visibility enhancement algorithms and machine learning are an efficient yet supplementary concrete bridge inspection tool. Two different image enhancement algorithms, i.e., denoise algorithm and image property adjustment, were considered in this study. To assess the adequacy of the proposed UAS technologies in the bridge inspections, the technologies were applied to identify and quantify defects on an existing concrete double-tee bridge located in the state of South Dakota using a Matrice 210 unit. During the inspections, Matrice 210 recorded videos to extract numerous UAS inspection images throughout the bridge. Machine learning was applied to categorize each of the UAS inspection images into certain defect types such as rust and spalling. The denoise algorithm was used to reduce the noise on the categorized defect images based on the pretrained denoising neural network, while the image property adjustment algorithm was employed to improve the visibility of the images by filtering the images’ brightness, contrast, and sharpness. Through these algorithms, defects on the filtered images initially presented with low visibility, were detected. Furthermore, quantification of the defects was able to be completed using pixel-based image analysis with the filtered images. From the UAS-assisted inspections, concrete spalling and rust on railings of the bridge were observed, detected, and quantified successfully. The quantification of spalling showed only a 6.00% difference compared against the inspection report data provided by the South Dakota Department of Transportation (SDDOT).