Self‐training approach for crack detection using synthesized crack images based on conditional generative adversarial network

Abstract

Urban infrastructure plays a crucial role in determining the quality of life for citizens. However, given the increasing number of aging infrastructures, regular inspections are essential to prevent accidents. Deep learning studies have been conducted to detect structural damage and ensure high accuracy and reliability of these inspections. However, these detection algorithms often face challenges due to scarcity of damage data. To overcome this issue, this paper proposes a method for synthesizing crack images and utilizing them for crack detection. Initially, crack images are synthesized from labeled images by using a conditional generative adversarial network. Subsequently, a new self‐training method is implemented wherein the synthesized crack images from the prediction images were incorporated into the learning process to enhance data diversity. The proposed approach yields promising results with a mean intersection over union of 80.34% and F1‐score of 76.31% on average. The proposed method can aid further research on virtual image generation for crack detection, seeking to reduce the reliance on extensive image collection.

An Integrated Framework for Image Acquisition, Processing, and Analysis Procedures for Automated Damage Evaluation of Concrete Surfaces

“…Nevertheless, there is a shortage of datasets containing the required images to train machine learning models for quantifying the extent of cracking (i.e., minimal annotated datasets) [16,17]. Moreover, most of the literature is focused on the crack detection task [18][19][20][21][22][23][24][25][26][27][28][29][30][31][32] where images are To perform a quantitative analysis of the damage extent to monitor volumetric expansion over time, researchers have introduced a surface crack mapping technique known as the cracking index (CI). This non-destructive quantitative tool assesses the degree of damage, estimating concrete expansion by measuring the widths of cracks observed on the ASR-affected concrete surface [6,7].…”

“…Nevertheless, there is a shortage of datasets containing the required images to train machine learning models for quantifying the extent of cracking (i.e., minimal annotated datasets) [16,17]. Moreover, most of the literature is focused on the crack detection task [18][19][20][21][22][23][24][25][26][27][28][29][30][31][32] where images are taken near the surface and without quantification of the damage. In a condition assessment, a quantitative value is required to inform the decision regarding the next steps and to monitor the increase in the damage over time to capture the rate of the damage.…”