Can photography and AI illuminate concrete inspection?
Concrete may symbolise strength, but it does not last forever. In the UK, the majority of our infrastructure was built during the post-war construction boom of the 1960s and 70s, including bridges, tunnels and public buildings, and much of it is now ageing out of its intended service life, raising urgent questions about safety and maintenance.
As concrete ages, signs of wear such as cracking and spalling often start to appear. Some cracks are harmless, caused by shrinkage or intentionally designed movement. However, others may point to more serious problems, including corrosion of the internal steel reinforcement or structural weakness.
When steel reinforcement embedded within concrete corrodes, it expands, causing the surrounding concrete to break away. This process, known as spalling, exposes more of the structure to the elements and accelerates deterioration. It also poses a safety risk if falling debris strikes people below.
With an ageing asset population, regular inspections and timely repairs are needed more than ever to help keep structures in good condition, extend their lifespan and, ultimately, ensure they are safe for use. However, when infrastructure maintenance and inspections do not keep pace with deterioration, challenges arise.
The 2018 collapse of the Morandi Bridge in Genoa, Italy, where 43 people were tragically killed, shocked the world. Closer to home, there have been 13 bridge collapses in the UK since 2015, as reported by the UK Bridge Owners Forum.
Did you know?
According to a report from the UK Bridge Owners Forum on Grand Challenges, there have been 13 bridge collapses in the UK from 2015-2024.
2015 – Tadcaster Bridge in North Yorkshire; Pooley Bridge in Cumbria; Laxey Shore Bridge on Isle of Man
2016 – Bell Bridge and Keswick Path Bridge in Cumbria; Eastham Bridge in Worcester; M20 Footbridge in Kent; Barrow-upon-Soar in Leicestershire
2018 – Skipton Bridge in North Yorkshire
2019 – Grinton Moor Bridges in North Yorkshire; Bishopsford Road Bridge in Mitcham, London
2021 – Llannerch Bridge in Denbighshire
2023 – Heyford Bridge in Oxon
2024 – Powick Bridge in Worcester
These incidents are stark reminders of the potentially catastrophic consequences when concrete infrastructure inspection and maintenance falls behind deterioration.
While major tragedies grab headlines, the everyday effects of asset management challenges on society are often indirect but still impactful. Many public assets are maintained reactively, meaning that little or no intervention occurs until a problem becomes serious. This can seem cost-effective in the short term, but when issues escalate, repairs tend to be urgent and expensive. Since most major infrastructure is publicly owned, these unexpected and high costs are ultimately shouldered by taxpayers.
In the worst cases when degradation has advanced too far, repair is no longer viable. Demolition and full replacement becomes the only option. This is disruptive and also environmentally and economically costly. In most cases, the best structure is the one that already exists, provided it is properly cared for. This is why proactive maintenance is so important.
A study published in the International Journal of Sustainable Engineering, titled A new methodology to inform maintenance decisions and budget requirements for bridges, demonstrates that proactive maintenance strategies significantly outperform reactive ones, in terms of both carbon footprint and repair bill.
However, as inspections are slow and labour- and resource-intensive, they are typically reserved for when necessary.
As a result, asset managers are often left with no option but to carry out maintenance reactively rather than proactively.
Traditionally, inspections have relied on experienced human professionals to identify visible signs of damage, such as cracks, corrosion and spalling.
But even skilled inspectors face limitations. Some areas are difficult or dangerous to access, and visual assessments can be subjective, leading to inconsistencies between inspectors.
Did you know?
The San Giorgio Bridge is a replacement hybrid concrete and steel bridge for the collapsed Morandi Bridge in Genoa, Italy. Equipped with four robotic arms that handle inspection and cleaning, two of these arms are fitted with cameras capturing over 25,000 images every eight hours along the structure’s beam spans.
Super vision
Recent advances in camera technology, robotics and drones are beginning to address these challenges. Remote image data capture improves inspection safety by reducing human exposure to hazardous conditions and enhances efficiency by minimising the need for disruptive closures during inspections. However, reviewing the collected image data still depends largely on human interpretation.
Artificial intelligence (AI) has the potential to transform this process. By analysing high-resolution images collected remotely by drones and robots, AI-powered systems can automatically detect, classify and measure defects. These systems can map surface deterioration and even forecast potential failure zones by analysing patterns in the data over time. And, when imaging conditions are right, they can do this consistently and objectively, reducing the variability that comes with human judgement.
AI computer vision models learn from large datasets of labelled images showing various types of structural defects. These models can be thought of as a ‘black box’ that take images as an input and produce defect classifications as an output. The interpretation from input to output is controlled by many adjustable variables within the black box. A model is ‘trained’ to learn the optimal variables. During training, it analyses large volumes of examples, such as cracks, surface corrosion, or spalling, by processing input images and iteratively adjusting its black-box variables.
After each round of adjustment, the model produces a score that reflects how well the current variables performed for the defect detection task. After many iterations, the model saves the variable settings that achieved the best performance score and these are then used by the model to accurately detect defects in new, unseen images. Bear in mind, this is a simplified explanation and the actual process is more complex in practice.
Casing the joint
AI and cameras can also build a digital and documented timeline of deterioration. Asset managers and owners can then make informed decisions about maintenance priorities and long-term investment. Theses benefits are increasingly being seen in real-world applications, where continuous monitoring allows for timely and preventative maintenance.
The San Giorgio Bridge is a replacement hybrid concrete and steel bridge for the previously mentioned Morandi Bridge that collapsed. Equipped with four robotic arms that handle inspection and cleaning, two of these arms are fitted with cameras capturing over 25,000 images every eight hours along the structure’s beam spans. These images are processed by a model to check for anomalies by comparing them to previous inspection data. Asset managers can monitor defects and take full advantage of the benefits of a preventative maintenance regime.
Other concrete assets, such as tunnels, are also seeing the benefits of AI-based inspection systems. Hong Kong’s Trunk Road T2 and Cha Kwo Ling Tunnel project uses a coordinated inspection system of an unmanned ground vehicle and an unmanned aerial vehicle.
The two devices work in tandem to remotely capture high-resolution, 360o images of the tunnel interior. AI processes this data to identify defects at the millimetre level. According to its creators, the system speeds up the inspection process by a factor of 23 and cuts costs by 50%.
These projects demonstrate the power of implementing AI into inspection regimes. However, variations in lighting, shadows, motion blur or even weather conditions can impact image quality and therefore accuracy. These are the same factors that affect how humans interpret inspection images.
It’s all about the lighting
'It’s all about the lighting' is a phrase many professional photographers swear by. Photographers regularly use specialist lights and reflectors to carefully control illumination. This eliminates unwanted shadows or, alternatively, use shadows strategically to highlight specific areas of an image. Lighting can make or break an image.
However, in concrete infrastructure inspections, this level of care is overlooked. Most image acquisition systems used for concrete structures ignore the importance of lighting. A few acknowledge it in the form of a flash or lighting system to remove shadows, but none use lighting in a strategic way to enhance image quality using shadows like a photographer would.
At the University of Strathclyde’s Department of Civil and Environmental Engineering, we are working to close that research gap. Our team is exploring how lighting can illuminate an image scene and actively improve inspection quality.
Our patented lighting-enhanced inspection approach (named ALICS) works by illuminating an image scene with lighting projected from different angles and directions to gather more information about the surface being inspected.
These varied lighting sources create shadows in defects such as cracks, helping to highlight them against the concrete surface and improving image consistency.
Better lighting does not just make the image clearer for the human eye, it also improves how AI can interpret and process inspection data. Providing an AI model with numerous images of a scene, each captured under different lighting conditions, gives the model more data to help its reasoning. Additionally, this approach can reveal defects that were previously not visible and ensures consistent lighting conditions across all inspection data.
A study titled Comparison of directional and diffused lighting for pixel-level segmentation of concrete cracks, in the journal Infrastructures, shows how this technique can outperform regular lighting set-ups for crack detection and measurement.
Building on our research, we have translated this technique from the laboratory to real-world impact as part of National Highways’ Structures Moonshot Programme – designed to test emerging technologies that could help address the UK’s ageing infrastructure crisis. As part of the programme, our combined lighting and AI-based inspection approach was deployed on the A432 Badminton Road Bridge – a post-tensioned concrete motorway viaduct.
The bridge, which was closed for safety reasons, exhibited severe cracking across its deck that was difficult to document and record using traditional inspection methods. Using our approach, we generated a comprehensive ‘crack map’ of one of the bridge’s beams. This created a timestamped record in which every crack could be measured at sub-millimetre precision. In principle, repeated mapping over time would enable detailed monitoring of crack progression. However, in this case, only a single scan was undertaken before the bridge was demolished one week later.
By comparison, a conventional bridge inspection typically requires one-to-two days and usually produces a hand-drawn sketch of crack locations and paths, supplemented by limited manual measurements. This process risks omissions and lacks precision. Achieving sub-millimetre accuracy by hand across an entire structure would be impractical, and this is precisely where the ALICS system delivers value.
For the 60-metre beam surveyed at A432 Badminton Road Bridge, the system produced a full crack map in approximately five hours using two manually positioned devices. With robotic or drone-mounted deployment, the process could be automated and potentially conducted overnight, further reducing inspection time and disruption.
There are some downsides to our approach. For example, capturing images under multiple lighting conditions can increase the time needed for image acquisition. Also, effective artificial lighting often requires low ambient light or darkness to produce the desired effect. But these R&D challenges can be addressed with further innovation. Regardless, one thing is clear – the data given to an AI model needs to be high quality to avoid missing any defects.
An intelligent future?
More broadly, AI inspections come with their own difficulties. It is important to remember that AI models are not immune to mistakes. Training models also demands significant time and resources, and running AI models raises sustainability concerns due to their energy consumption.
For example, a 2024 white paper by the Electric Power Research Institute, on Understanding impacts of data centers on greenhouse gas emissions in the United States, estimates that a single ChatGPT query uses a staggering 2.9Wh of energy, equivalent to powering a 10W energy-saving bulb for around 17 minutes.
Applying an AI model from one concrete structure to another requires a process known as domain transfer and can be problematic due to surface textures and colours, sometimes requiring extra tuning or retraining.
This challenge is illustrated in a 2019 study from the University of California, Los Angeles, USA, which showed that a robust animal classification model failed when presented with glass replicas instead of real animals. In one bizarre example, the AI confidently misclassified a polar bear as a can opener.
The black-box nature of AI models adds further concern. Their inner workings are often not fully transparent, making it difficult to understand how decisions are made. This lack of transparency can lead to errors that are hard to trace or correct, which poses challenges for ensuring inspection reliability and accountability.
Despite these challenges, the benefits of AI in infrastructure inspection are significant and ongoing advances are likely to reduce these downsides over time.
So when will we see an army of AI-powered humanoid robot inspectors scanning M25 overpasses? Probably not anytime soon – at least not unless Arnold Schwarzenegger himself shows up to lead the charge.
AI is exactly that – artificial. It should be seen as a tool to support human inspectors, not replace them. AI can help speed up inspections, assist in safety improvements, process huge volumes of data and help focus human attention (otherwise known as non-artificial intelligence) on areas that matter most. In my opinion, it will not replace expert judgement, it will enhance it.
For the foreseeable future, the most effective approach will be a hybrid one, combining human insight with AI assistance. As the technology improves, this partnership will only become more capable and efficient.
Extracted from University of Strathclyde's website, read more here