Automatic Delamination Segmentation for Bridge Deck Based on Encoder-Decoder Deep Learning Through UAV-based Thermography

Authors

ChongshengFollow
Zhexiong ShangFollow
Zhigang Shen, University of Nebraska - LincolnFollow

Date of this Version

6-19-2020

Citation

Chongsheng Cheng, Zhexiong Shang, Zhigang Shen, Automatic delamination segmentation for bridge deck based on encoder-decoder deep learning through UAV-based thermography, NDT & E International, Volume 116, 2020, 102341, ISSN 0963-8695, https://doi.org/10.1016/j.ndteint.2020.102341. (https://www.sciencedirect.com/science/article/pii/S0963869520303224)

Comments

Copyright © 2020 Chongsheng Cheng, Zhexiong Shang, and Zhigang Shen

Abstract

Concrete deck delamination often demonstrates strong variations in size, shape, and temperature distribution under the influences of outdoor weather conditions. The strong variations create challenges for pure analytical solutions in infrared image segmentation of delaminated areas. The recently developed supervised deep learning approach demonstrated the potentials in achieving automatic segmentation of RGB images. However, its effectiveness in segmenting thermal images remains under-explored. The main challenge lies in the development of specific models and the generation of a large range of labeled infrared images for training. To address this challenge, a customized deep learning model based on encoder-decoder architecture is proposed to segment the delaminated areas in thermal images at the pixel level. Data augmentation strategies were implemented in creating the training data set to improve the performance of the proposed model. The deep learning generated model was deployed in a real-world project to further evaluate the model’s applicability and robustness. The results of these experimental studies supported the effectiveness of the deep learning model in segmenting concrete delamination areas from infrared images. It also suggested that data augmentation is a helpful technique to address the small size issue of training samples. The field test with validation further demonstrated the generalizability of the proposed framework. Limitations of the proposed approach were also briefed at the end of the paper.