Use of Deep Encoder-Decoder Network for Sub-Surface Inspection and Evaluation of Bridge Decks

Abstract

The automation of various processes underlying maintenance and inspection of bridges using different robots have gained considerable attention in recent literature. For the development of effective methods to automate existing manual processes, a number of different solutions have been proposed. In this paper, the automation of rebar detection and localization will be discussed, which is one of the process for sub-surface health inspection of bridges. This study explores the utilization of Deep Encoder- Decoder Networks for the segmentation of GPR data in the form of B-scan images to extract parabolic rebar profiles. This research area is problematic, as the B-scan image data is fraught with noise, signal reflection and other artefacts that hinder the effective extraction of these rebar profiles. The data is collected in this study using Ground Penetrating Radar (GPR) sensor, which is employed in this study consist of data from 8 different bridges from different parts of the United States. A leave-one-out approach was used for the training and validation of the performance of the proposed system; the data from seven bridges was used for training and validation was performed on the remaining single bridge data. A number of different encoder modules have been trained and evaluated using SegNet as the backbone architecture. The performance of the proposed rebar detection and localization system has been evaluated in terms of different qualitative and quantitative metrics. On average, for the different encoder modules, the mean intersection-over-union (mIOU) values range between 60%-70%. The qualitative examination has highlighted the level of similarity between the ground truth and outputs from the different encoder modules within the SegNet framework.