Researchers in the FRISSBE department at ZAG explore how thermal image quality is assessed, where current methods fall short and what improvement could look like
Thermal imaging cameras (TICs) are one of the essential tools used by firefighters.
Using sensors that operate in the long-wave infrared (IR) spectrum, these devices convert heat into visible images, enabling first responders to see through smoke, navigate through burning buildings, detect hotspots, and locate people quickly, even in zero-visibility conditions.
But how good is the current technology? And, more importantly, how do we make sure it serves well the purpose of the people using it?
Why image quality matters
The effectiveness of a TIC can mean the difference between life and death.
A good image quality is not just about sharpness or resolution; it is about a firefighter’s ability to distinguish critical details: is that shape a person or debris? Is that glow a smouldering ember or a doorway?
To ensure that TICs, often referred to as thermal imagers, meet the high standards required for emergency and rescue uses, the U.S.
follows the NFPA 1801 Standard on Thermal Imagers for the Fire Service (NFPA, 2021).
Developed by the National Fire Protection Association (NFPA), this standard lays out requirements for durability and image performance.
A comprehensive report by the Department for Fire-safe Sustainable Built Environment (FRISSBE) at the Slovenian National Building and Civil Engineering Institute (ZAG) and Ghent University, in collaboration with NFPA, the Fire Protection Research Foundation (FPRF), and the Electronic Safety Equipment Technical Committee, raised some concerns about the current framework.
The findings indicate that the current testing methods can be inconsistent, the image quality observed in the field often does not match lab results, and the testing process itself is cumbersome and outdated.
After nearly two decades since the standard was introduced, it is time to revisit the framework and utilise the technological advances that the recent decades have brought.
How can we better assess the quality of thermal images and what quality is good enough?
Rethinking image quality in the fire service
In simple terms, image quality is how clearly a person can perceive useful details from visualising an image.
That could be impacted by blurriness, noise, or compression, any of which could obscure vital information in a firefighting scenario, where time matters and fast decisions must be made frequently.
To evaluate image quality, two main approaches are used:
- Full-reference image quality assessment: comparing an image to a pristine version of an image;
- No-reference image quality assessment: judging an image without any reference image.
In real-world conditions of fireground use, full-reference image quality assessment is not an option, as there is no image to compare with.
As a result, researchers have looked into no-reference methods; training models to predict image quality based on how people have scored similar images in the past.
Multiple approaches exist:
- Natural Scene Statistics (NSS), which relies on statistical distributions of the images to extract features and create models;
- Machine Learning models, that rely on complex architectures and data to make predictions;
- saliency-based models, which predict where people look at on images.
Common to each of these approaches is that most models need a large amount of data of previous scores from research studies on human perception to make reasonable predictions on images that have not been seen before.
Previous research, outlined in the report, has shown the potential for the use of these technologies on TIC images.
Models created for the visible spectrum have shown an extraordinary ability to score images similarly to how a human would score the same image.
While little research has been done on the IR spectrum, some have successfully used the Natural Scene Statistics (NSS) framework and trained it on thermal images.
There is therefore immense potential to utilise these models for thermal images in the fire service.
The challenge? There’s currently no publicly available dataset of thermal images in firefighting scenarios with human-scored quality labels.
The roadblock and the opportunity
Today’s models are trained mostly on images in the visible spectrum, whereas TICs operate in the IR spectrum.
Without a dedicated dataset, it is impossible to develop reliable no-reference image quality models for TICs.
That is unfortunate, as such a dataset could be used to improve testing standards, as well as open the door to future innovations.
Imagine TICs that use AI to highlight people automatically in heavy smoke, or training tools that simulate realistic rescue scenarios based on real-world TIC data.
A comprehensive, high-quality thermal image dataset could spark major advances in safety, certification, and even product development.
The good news is that the hard part, figuring out how to collect this kind of data, has already been solved for images curated and scored in the visible spectrum.
However, obtaining images in the IR spectrum, with firefighting scenarios in mind, does introduce some obstacles and challenges.
Are the same distortions that were previously applied in the visible spectrum still relevant? Are there other distortions, which were not considered for the visible spectrum, significant for images in the IR spectrum? Should the scoring of the images be carried out in a similar manner to the previous studies? These questions persist, but the recently published report tries to address some of these questions to bridge the gap and make the next steps more manageable, so such a dataset can come to fruition.
Areas of focus
The technical report, ‘Measuring Thermal Image Quality for Fire Service Applications’, funded by the NFPA Fire Protection Research Foundation, outlines three core areas of focus.
First, it reviews existing methods and emerging approaches for assessing thermal image quality, explaining that current metrics can be strengthened by incorporating established models that already predict image quality with a high degree of accuracy.
Second, the report highlights the need to curate a dedicated thermal image quality dataset.
This dataset should be developed specifically with the operational and environmental challenges of structural firefighting in mind, where thermal imaging is affected by heat, smoke, and dynamic conditions.
Finally, the report emphasises the importance of rigorous model evaluation and testing.
Identified models should be trained using thermal imagery and systematically assessed to determine which approach is most suitable for real-world firefighting applications.
What’s next?
Creating a reliable dataset might seem like a daunting task, but its potential impact is massive.
It could revolutionise how TICs are tested, certified, and ultimately used in the field, empowering firefighters with better tools and more confidence in their gear.
Have ideas, data, or feedback to share? Want to be part of the conversation? Reach out to Martin Veit, lead researcher of the project, at martin.veit@zag.si.
The authors of this article are: Martin Veit, Researcher in the FRISSBE department at ZAG, Slovenia, Andrea Lucherini, Senior Researcher in the FRISSBE department at ZAG, Slovenia, Grunde Jomaas, Era Chair Holder in the FRISSBE department at ZAG, Slovenia, and Bart Merci, Professor at Ghent University.