Approaches to Training Multi-Class Semantic Image Segmentation of Damage in Concrete

Published

May 14, 2020

Author(s)

Peter Bajcsy, Steven B. Feldman, Michael P. Majurski, Kenneth A. Snyder, Mary C. Brady

Abstract

This paper addresses the problem of creating a large quantity of high-quality training image segmentation masks from scanning electron microscopy (SEM) images of concrete samples that exhibit progressive amounts of degradation resulting from alkali-silica reaction (ASR), a leading cause of deterioration, cracking, and loss of capacity in much of the nation’s infrastructure. We approached the SEM segmentation problem by applying Convolutional Neural Network (CNN) based methods to predict the damage classes due to ASR for each image pixel. The target microstructural classes were defined as paste damage, aggregate damage, air voids, and no damage. The challenges in using the CNN-based methods lie in preparing large numbers of high-quality training labeled images while having limited human resources. To address these challenges, we designed damage-assisted and context-assisted approaches for lowering the requirements on human resources. In a damage-assisted approach, we automate a creation of a damage super-class which lowers the manual annotation labor to drawing bounding regions with a computer mouse to subset the damage class into target sub-classes. In a context-assisted approach, we automatically detect a damage super-class together with surrounding contextual classes and then sub-divide the damage super-class into target classes. This approach minimizes the manual annotation labor to image level labeling via a web interactive interface. The evaluations of these two training data sets consists of expert’s time per annotation, pixel-level accuracy and standard deviation of trained CNN models, and image-level accuracy based on visual inspections. We disseminate 4594 raw fields of views (FOVs) with their damage-assisted, context-assisted, and U-Net predicted target microstructural masks within an interactive web-based validation system at https://isg.nist.gov/data.

Citation

Journal of Microscopy-Oxford

Volume

279

Issue

2

Pub Type

Journals

Download Paper

DOI Link

Keywords

artificial neural network, training data, concrete, scanning electron microscopy

Citation

Bajcsy, P. , Feldman, S. , Majurski, M. , Snyder, K. and Brady, M. (2020), Approaches to Training Multi-Class Semantic Image Segmentation of Damage in Concrete, Journal of Microscopy-Oxford, [online], https://doi.org/10.1111/jmi.12906 (Accessed November 21, 2024)