Highlights

Above

Bayesian analysis is helping algorithms distinguish true cracks from false positives, even on complex surfaces like terrazzo.

© Shutterstock

Piecing the cracks together

31 May 2021

A hybrid approach combining deep learning with Bayesian inference has enabled more accurate, efficient and automatic crack detection.

As you walk around your neighborhood, you will likely see cracks in the concrete under your feet, in beams and even along some buildings. These cracks can be a sign that the structure may become unsafe, which is why inspections in many industries—from infrastructure to aeronautics—commonly include crack detection.

Manual visual inspection is still the main method used to detect cracks today, an approach that requires specialist knowledge and is labor-intensive, expensive and time-consuming. In the past three decades or so, researchers have made considerable strides in developing ways to automatically detect cracks, first using image processing methods and more recently using machine learning-based approaches.

The challenge with automatic detection is that cracks vary widely in length, shape and orientation, and often have low signal-to-background ratios, making it difficult to train deep learning technologies accurately. In this study, first author Fen Fang, a Research Scientist at A*STAR’s Institute for Infocomm Research (I2R) supervised by team leader Liyuan Li, describes a novel hybrid approach that combines deep learning with Bayesian analysis, a method of statistical inference, to more robustly and efficiently detect cracks from images.

The researchers first re-trained an existing algorithm to detect patches of cracks with sufficiently high signal-to-noise ratios, layering it information about real-world cracks that had been annotated for machine learning. To identify the true cracks, they next trained a deep learning model to estimate the orientation of the crack in each patch and developed a Bayesian algorithm to analyze the probability that the detected crack is real.

“Based on domain knowledge, true cracks are tiny, linked lines, while false positives are often isolated and separate patch detections,” Fang explained. “Hence, we developed a Bayesian integration approach based on spatial proximity, orientation consistency and alignment consistency to connect the potentially true patch detections and suppress false detections.”

The researchers tested their approach on a newly built dataset of 1,675 raw images of cracks found in over ten materials that were captured at different times of the day and under varying weather conditions. “Our dataset included not only concrete and asphalt road surfaces but also surfaces made of rock, terrazzo, marble, brick, tiles and so on. With this enhanced diversity, we achieved much better performance on real-world images,” Fang said.

The patent-pending technology has been commercially licensed by a multinational corporation for use in construction and building inspections, in addition to being used by a government agency for airplane inspections, she added.

The A*STAR-affiliated researchers contributing to this research are from the Institute for Infocomm Research (I2R).

Want to stay up to date with breakthroughs from A*STAR? Follow us on Twitter and LinkedIn!

References

Fang, F., Li, L., Gu, Y., Zhu, H., Lim, J.H. A novel hybrid approach to crack detection. Pattern Recognition 107, 107474 (2020) | article

About the Researcher

Fen Fang

Research Scientist

Institute for Infocomm Research
Fen Fang received her Ph.D. degree in computer science and engineering, majoring in Computer Graphics, from Nanyang Technological University, Singapore, in 2014. She is currently a Research Scientist with A*STAR’s Institute for Infocomm Research. Her research interests include defect detection, object detection and segmentation, scene recognition and 3D object reconstruction.

This article was made for A*STAR Research by Wildtype Media Group