Deep learning hits the road: a novel architecture for pavement crack detection and severity assessment
Researchers at the United Arab Emirates University are exploring the potential for deep learning to revolutionise how we manage and maintain civil infrastructure, engineering a system that identifies pavement cracks and categorises them by severity
Sponsored by United Arab Emirates University
Sponsored by
Sponsored by
Deep learning allows specialists in many sectors to process data at scale, empowering them to make better, quicker decisions. At the United Arab Emirates University (UAEU), researchers are applying such techniques to infrastructure management, allowing them to monitor and maintain road networks to promote public safety and sustainability. The team has adapted a Residual Sharp U-Net architecture to detect and assess the severity of pavement cracks. The researchers train the model with a large, high-quality data set and configure the system to process the images and set parameters for what they are looking for.
Hamad Al Jassmi, director of the Emirates Centre for Mobility Research (ECMR) and professor in the Department of Civil and Environmental Engineering at UAEU, highlights the complexity of detecting pavement cracks compared to other image segmentation tasks. “Tuning is definitely a challenging task,” Al Jassmi explains. “Let’s suppose you are training your network, and you are just segmenting a dog and a cat in an image – the features vary. In our case, with pavement crack detection, if you see the cracks, the properties are almost similar to the colour of the roads. If you see the texture of the crack, it’s black. If you are dealing with concrete cracks on a painted wall, the crack is really visible. In terms of the pavement, it is very challenging.”
“A major issue is the lack of local crack data due to the UAE’s relatively new infrastructure,” explains Fady Alnajjar, professor at the College of Information Technology at UAEU. “To address this, we used two publicly available datasets containing over 20,000 high-quality annotated images to ensure fair benchmarking and generalisability across different pavement types and lighting conditions.”
Luqman Ali, a research associate at the ECMR, stresses the need for fair model comparisons. “Using your own dataset doesn’t allow for a fair comparison when altering a model’s architecture,” he says. “You must use publicly available datasets that others have worked on. Both datasets we used had good image quality, but we chose them primarily to benchmark our work against existing research for a fair comparison.”
The next step for this research is to train a deep learning system to look for other defects, such as potholes in the road. In the future, the ECMR team wants to develop a way of applying these machine learning applications in the real world. Deep learning requires abundant computational power, which UAEU has on campus. But how much power would be needed to install this deep learning technology on an edge device, or on a drone that could capture images and assess them? The answer to that question could revolutionise infrastructure management.