Industries / Industrial equipment and services / Industrial services / Thermography
Reading and reporting images
This work turns a heat picture into a temperature, a finding and a report: the sensors, the corrections that make a reading true, the training of the people who read the images, and the software and labelled image sets that now help them.
6
Milestones achieved, from 1992 to 2020.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1992
Cameras lose their coolers
Until 1987, when Mitsubishi brought out a Stirling cycle cooler (a small sealed refrigerator) for camera detectors, cooled thermal cameras ran on liquid nitrogen, and thermographers had to keep track of how much they had left, a FLIR history recalls. Honeywell began developing the microbolometer in the late 1970s under a classified US Department of Defense contract. It is a grid of tiny sensors, each a resistor whose resistance changes as it warms, that works at room temperature. The technology was declassified in 1992, according to a patent analysis by KnowMade, after which most of Honeywell's patent families on it were published. Uncooled microbolometers are the sensors in today's low-cost thermal cameras, among them FLIR's Lepton.
Credit
Honeywell, for the US Department of Defense
1997
Rules for correcting readings
A thermal camera measures the heat radiation that reaches it, not temperature. To turn that into a temperature it must be told the surface's emissivity, how strongly the surface gives off heat radiation compared with a perfect radiator, on a scale of 0 to 1 (polished metal is low, most paint is high), and the reflected temperature, the heat from the surroundings that bounces off the surface into the lens. ASTM's Committee E07 published practices for measuring and correcting for reflected temperature (E1862, first edition 1997) and for emissivity (E1933, 1999), and a guide to examining electrical and mechanical equipment with thermography (E1934, 1999), which it reapproved most recently in 2024.
Credit
ASTM Committee E07
2008
Thermographers get three categories
ISO 18436-7, first published on 15 October 2008, sets what people who monitor machines with thermography must know to be certified. Category I applies a set technique, runs the camera safely, avoids poor data and makes basic fault and severity calls by established procedures; Category II works to established and recognized procedures and supervises Category I; Category III supervises both. Minimum training is 32, 64 and 96 hours, and candidates must also show documented experience. A second edition replaced it on 1 April 2014.
Credit
International Organization for Standardization (ISO)
2014
A heat camera for phones
In July 2014 FLIR Systems launched the FLIR ONE, a $350 case that snaps onto an iPhone 5 or 5S and shows heat images on the phone's screen. It is built around FLIR's Lepton, an uncooled microbolometer of 80 by 60 pixels with a lens made at wafer scale, sensitive to wavelengths from 8 to 14 micrometres (µm).
LessonIn a 2015 test that measured skin temperature against a blackbody (a reference surface held at a known temperature), Allen Curran and colleagues found that calibration narrowed the gap between a FLIR E60 professional camera and a spot thermometer from 1.0 °C to 0.48 °C but did not help the FLIR ONE, which they judged unsuitable for absolute temperatures; a cheap camera shows where heat is more reliably than how hot it is.
Credit
FLIR Systems
2015
Free code reads the files
A radiometric image file stores a raw value for every pixel and the constants needed to turn it into a temperature, in a format each camera maker sets. On 26 March 2015 Glenn J. Tattersall of Brock University in Canada published Thermimage, a free package for the R statistics system that reads FLIR files and converts the raw values to temperatures. In his checks on one image, three FLIR programs given the same settings returned temperatures that differed by 0.07 to 0.13 °C, and Thermimage came within about 0.03 to 0.04 °C of one of them, about the resolution of the camera used.
LessonVersion 4.0, released on 31 October 2019, fixed an error in the atmospheric constants the conversion uses, which mattered at long distances; a reader of thermal files needs testing just as a camera does.
Credit
Glenn J. Tattersall
2020
Solar fault labels go public
Raptor Maps, a US company that analyses drone surveys of solar farms, released InfraredSolarModules in February 2020: 20,000 thermal images of single solar modules, each 24 by 40 pixels, labelled in 12 classes such as hot cells, cracking, an activated bypass diode (a part that lets current skirt a failed section), shading, soiling, plant growth and offline modules, with 10,000 showing no fault. It was presented at a workshop on AI for earth sciences at the ICLR 2020 machine learning conference and is free under the MIT license. Outside teams train and compare fault classifiers on it, for example JEFFNet, a model described in a July 2026 preprint.
Credit
Raptor Maps
Watch for
Software writes the findings
Chauvin Arnoux, a French instrument maker, says its CAmReport AI software uses artificial intelligence trained to identify thermal defects to write observations and recommendations for a thermal image "in just one click", and builds reports in three steps for export to PDF or Word, for electrical, mechanical, solar and building work. Its first leaflet lists the company's CA 1950 and CA 1954 cameras as compatible.
Achieved whenan outside party runs the software on thermal images whose findings certified thermographers have already agreed, and publishes how often its observations match and what it misses.
Claimed by
Chauvin Arnoux, February 2026
Open
01
One file, many readers
A public scorecard of the programs that turn radiometric files into temperatures. Photograph one indoor bench scene, with plates held at known temperatures and checked by contact probes, using cameras from at least five makers. Run every file through each maker's own software and through free readers, first at the camera's default settings and then with emissivity, reflected temperature and distance changed, and publish each reader's temperature for each plate beside the reference and beside the maker's own program.
The briefCredit · unclaimed02
Emissivity of plant surfaces
An open table of emissivity in the 8 to 14 µm band that most inspection cameras use, measured on about 40 surfaces as a thermographer meets them in electrical rooms and plants: painted, galvanized and rusted steel; clean and oxidized copper and aluminium busbar; tinned and bolted lugs; PVC, rubber and cloth cable jackets; aluminium and stainless pipe cladding; glazed porcelain insulators. Each sample is measured at three temperatures by two methods, against a contact thermocouple and against a patch of high-emissivity tape of known value, with a photo of the surface, the camera used and an uncertainty for every value.
The briefCredit · unclaimed
Open 01 · the brief
One file, many readers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public scorecard of the programs that turn radiometric files into temperatures. Photograph one indoor bench scene, with plates held at known temperatures and checked by contact probes, using cameras from at least five makers. Run every file through each maker's own software and through free readers, first at the camera's default settings and then with emissivity, reflected temperature and distance changed, and publish each reader's temperature for each plate beside the reference and beside the maker's own program.
Why it is still open
Tattersall's checks covered one image from one FLIR camera read by FLIR programs, and he wrote that the differences "suggest caution, however in assuming all software is operating on the same principles". Jonathan Klingspon and colleagues at the University of California, San Diego, report (August 2026) that they found no public dataset that combines raw radiometric data, an independent temperature reference and a matching colour image, and that a reader which skips the window setting stored in FLIR One Pro files is off by +2.1 K at 15 °C and -2.9 K at 40 °C. Časlav Livada and colleagues (Journal of Imaging, July 2023) showed, with one FLIR E60bx camera and a blackbody at 60 °C, how brightness, contrast and JPEG compression change the temperatures read back from a file, but tested one maker's software. STR Drones' free sample sets let developers compare their software with values recorded in the files' own metadata, not with a measured reference. A GitHub search in November 2026 found many readers and no scorecard across them.
github.comarxiv.orgstrdrones-readme.static.hf.spacepmc.ncbi.nlm.nih.gov
Built with what exists
Borrowed or rented cameras from makers such as FLIR, DJI, Hikmicro and Testo; aluminium plates coated with high-emissivity paint on laboratory hot plates at 30 to 120 °C, with calibrated contact thermometers; each maker's desktop software; Thermimage, ExifTool (which pulls the conversion constants out of a file), flirpy and DJI's Thermal SDK; Python to build the tables.
First step
Photograph two heated plates of known temperature with one FLIR and one DJI camera and run both files through every reader that opens them.
Done looks like
Files from at least five makers and ten camera models, each read by at least four programs at three settings; a public table of every reader's error against the reference plates, with the largest disagreement between readers for each maker; all files and scripts released.
Open 02 · the brief
Emissivity of plant surfaces
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open table of emissivity in the 8 to 14 µm band that most inspection cameras use, measured on about 40 surfaces as a thermographer meets them in electrical rooms and plants: painted, galvanized and rusted steel; clean and oxidized copper and aluminium busbar; tinned and bolted lugs; PVC, rubber and cloth cable jackets; aluminium and stainless pipe cladding; glazed porcelain insulators. Each sample is measured at three temperatures by two methods, against a contact thermocouple and against a patch of high-emissivity tape of known value, with a photo of the surface, the camera used and an uncertainty for every value.
Why it is still open
The 2020 emissivity database from the University of Strathclyde's Advanced Forming Research Centre covers metals and furnace linings at forming temperatures, such as oxidized aluminium, electrical steels, Inconel 718 and titanium alloy. Klingspon and colleagues (August 2026) built a heritage-materials table from the literature, found no published direct measurement for some named stones, report that weathered metal reads 0.85 to 0.98 where polished metal reads 0.15 to 0.35, and say their low-emissivity entries "remain externally unvalidated"; an outside spectral library from the HADAR project matched only their stone and plaster entries. They note that the FLIR app they used writes a preset emissivity of 0.95 into every image. Building materials are better served: Eva Barreira, Ricardo Almeida and Maria Simões (Sensors, 2021) measured nine of them by emissometer and by camera with reference tape and found values different enough from published tables to shift surface temperatures by up to 7 °C, and the Spectral Library of Impervious Urban Materials (SLUM) gives emissivity spectra for 74 urban surfaces. Searches in November 2026 found no such table for in-service electrical and plant surfaces.
sensorsportal.comarxiv.orgpmc.ncbi.nlm.nih.govmicromet.reading.ac.uk
Built with what exists
A radiometric thermal camera; a laboratory hot plate; type K thermocouples with a calibrated reader; high-emissivity tape or paint; samples salvaged from scrapyards and electrical contractors' offcuts, none of them connected to power; ASTM E1933 as the reference for the method; Zenodo to publish the table and images.
First step
Measure five surfaces (painted steel, galvanized steel, oxidized copper, a PVC jacket and glazed porcelain) on a hot plate at 50 °C by both methods.
Done looks like
40 surfaces measured at three temperatures by both methods, with the agreement between methods and the uncertainty of each value published, and a count of the makers' published table values that differ from the measured ones by more than 0.05.