As an Amazon Associate, Akermin earns from qualifying purchases. Product links on this page are affiliate links.
Each detector element reports one number for the entire cone of scene it looks at, and that number is a radiance weighted average of the hot part and everything around it. A part at 120 degrees C filling a quarter of the measurement spot on a 25 degree C background returns roughly 57 degrees C, and no amount of computing intermediate pixels from that average can put the missing 63 degrees back, because the arithmetic has only the averaged values to work from. Interpolation fills space between samples. It does not create samples.
A detector reports one number for everything in its cone
A microbolometer element is a small suspended membrane that warms up when infrared energy lands on it. The optics in front of the array map a specific cone of directions onto each membrane, and every photon arriving from anywhere inside that cone contributes to the same heating. The element has no way to say where within its cone the energy came from. It reports a single value, and that value is set by the total radiance collected.
Radiance does not scale with temperature, it scales with roughly the fourth power of absolute temperature for a surface behaving as a blackbody. That is why the averaging works out the way it does. If a fraction of the cone is filled by a hot surface and the rest by a cool one, the total energy is the area weighted sum of two fourth powers, and the apparent temperature reported is the fourth root of that sum. The hot fraction is pulled down hard, and the result sits much closer to the background than a simple average of the two temperatures would suggest for small fractions and above it for large ones.
This is also the reason manufacturers quote a measurement spot that is larger than the angle of a single element. Alignment is imperfect, optics scatter a little energy into neighbouring elements, and the stated accuracy only holds when the target comfortably fills the specified spot. A single element exactly covering a target is not a measurement, it is a coincidence waiting to be disturbed by a millimetre of pointing error.
New pixels are arithmetic on old samples
Once the array has produced its grid of averaged values, resampling to a larger grid is a purely mathematical operation. A bicubic or bilinear scheme fits a smooth surface through the existing values and evaluates it at intermediate positions. Nothing in that process has access to the scene. It has access only to the numbers the array already produced, which are the averages described above.
The consequence follows immediately. If the peak temperature of a small part never appeared in any element because it was always diluted by its surroundings, then it is not in the input, and no function fitted through the input will produce it. The output grid will contain many smoothly varying values none of which exceeds the diluted maximum. The picture becomes larger and smoother. The highest number in it does not move.
The same constraint appears in every sampled measurement. A logger that records a value every minute cannot report a spike that lasted five seconds, and plotting the log at one second intervals through a smooth curve produces a graph with sixty times as many points and the spike still missing. A gas instrument that draws a sample into a chamber reports the concentration of what arrived, so a small escape dispersed into a large volume reads as a small number regardless of how finely that number is displayed. That dilution problem is the direct analogue of the one in a thermal image, and the discussion of whether gas detectors find small leaks works through it in the chemical case. The instrument in that discussion, the sort of multi channel unit represented by the Forensics Detectors four gas monitor with its oxygen, carbon monoxide, hydrogen sulphide and combustible channels and its adjustable audible, visual and vibrating alarms, faces exactly the same structure: the reading is an average over whatever the sensor was exposed to.
What a small hot part does to the reading
Here is the arithmetic in full, for a part at 120 degrees C against a background at 25 degrees C, with emissivity taken as one and atmospheric effects ignored so the only variable is how much of the measurement spot the part fills. The apparent temperature is the fourth root of the area weighted sum of the two fourth powers, converted back to degrees C.
| Fraction of the spot filled by the hot part | Apparent temperature reported | Simple area weighted average for comparison | Shortfall against the true 120 degrees C |
|---|---|---|---|
| 100 percent | 120.0 degrees C | 120.0 degrees C | None |
| 50 percent | 81.9 degrees C | 72.5 degrees C | 38 degrees C low |
| 25 percent | 57.1 degrees C | 48.8 degrees C | 63 degrees C low |
| 10 percent | 39.1 degrees C | 34.5 degrees C | 81 degrees C low |
| 5 percent | 32.3 degrees C | 29.8 degrees C | 88 degrees C low |
| 2 percent | 28.0 degrees C | 26.9 degrees C | 92 degrees C low |
Read the last two rows carefully, because they describe the common case. A part hot enough to fail, sitting in a dense assembly where it occupies a small part of the spot, raises the reading by three degrees above its surroundings. On a screen with a wide temperature span across the palette, three degrees is a faint warm tint that is easy to pass over, and an alarm threshold set at any sensible margin above ambient will not trigger on it. The failure is not that the camera is wrong. The camera correctly reported the average energy arriving from that direction.
The fix is geometric. The spot area shrinks with the square of distance, so halving the working distance quadruples the filled fraction, and a part that filled 5 percent of the spot at two metres fills 20 percent at one metre and reads much nearer the truth. Where you cannot get closer, a narrower field of view lens does the same job optically.
Why the upscaled image still looks convincing
The uncomfortable part is that the processed image looks more trustworthy than the raw one. Smooth gradients read as detail to the eye, so an interpolated frame presents what looks like a sharply defined warm region with a definite edge, and a cursor placed anywhere on it returns a number to one decimal place. All of that visual confidence is manufactured after the measurement.
A blocky display is, in this narrow sense, more honest, because the size of the blocks is the size of the sampling and the operator can see immediately how many independent measurements sit on the target. That is the instinct worth keeping when reading any processed image: ask how many real samples lie under the feature you are looking at, and if the answer is fewer than a handful, treat the temperature as an indication that something is warm rather than as a measurement of how warm.
The same caution applies to the other direction of processing. A strongly contrasted palette stretched over a narrow span makes a two degree variation look like a hotspot, and a wide span flattens a genuine one into nothing. Neither changes the sampled data, and both change the conclusion an operator reaches from it.
Measuring small parts by touching them instead
When the target is smaller than the spot and you cannot get closer, the answer is usually a different instrument rather than a different image processing mode. Contact and point sensing methods have no spatial averaging problem because the sensing element is physically placed on or beside the thing being measured, and their published dimensions tell you exactly what they are sampling.
A thermal anemometer makes the comparison concrete. The BTMETER hot wire anemometer publishes the physical size of its telescopic probe, 11.75 by 1.61 by 0.93 inches, alongside an air speed range of 0.2 to 67 miles per hour at plus or minus 5 percent and a temperature range of 32 to 113 degrees F. The size of the heated element itself, which is what actually sets the sampled volume, is Not stated in the listing bullets, but the probe envelope at least bounds it. You know where that sensor is and roughly what volume it is responding to, which is more than an interpolated pixel can offer. The same framing explains why the choice between a vane head and a hot wire element is a question about sampling volume rather than about accuracy.
For electronics specifically, the small part that a thermal camera cannot resolve can often be identified and characterised directly. A bench instrument such as the HiLetgo LCR-T4 component tester, which identifies transistors, field effect devices, diodes, resistors and capacitors automatically and returns a result in about 1.7 seconds on its 128 by 64 display, answers the question the image could not: not how hot the part looks from a distance, but what the part is and whether it is still behaving. Used together, the camera narrows the search to a region and the contact instrument settles what is happening inside it.
Frequently Asked Questions
Why does a small hot part read cold rather than just blurred?
Because the element is an integrator, not a sampler of a point. It collects radiant energy from everything inside its angular cone and reports one value, so a hot part occupying a fraction of that cone contributes only its fraction of the total energy and the cooler surroundings supply the rest. The output is a weighted average dominated by whatever fills most of the cone, and since that is usually the background, the reading sits close to the background.
Would a more sensitive detector solve it?
No, because sensitivity and spatial sampling are independent limits. Sensitivity describes the smallest temperature difference one element can distinguish from its own noise when the target fills its cone. The problem here is that the target does not fill the cone, so the energy from the hot part is diluted before it ever reaches the detector. Improving the noise floor makes the diluted reading cleaner without making it correct.
Can moving closer fix it?
Yes, and it is usually the only field fix available. The angular cone is fixed, so the area each element covers shrinks in proportion to distance. Halving the working distance halves the width of the spot and quarters its area, which quadruples the fraction of the spot the part fills. Focus limits and access set how far you can take this, and on live equipment the safe approach distance may set it for you.
What is the rule of thumb for how many elements a target needs?
Commonly three across, giving roughly nine elements on the target, so that at least one of them sees the target and nothing else. Some manufacturers quote a measurement spot several times larger than the single element angle for exactly this reason, and their stated accuracy applies only when the target fills that spot. One element exactly covering a target is not enough, because alignment is never that good and any pointing error puts background inside the cone.
Does upscaling ever make the reading worse rather than just no better?
It can, by encouraging a measurement that would not otherwise be attempted. A smooth image invites a spot cursor to be placed on an apparent small feature, and the value returned is computed from surrounding samples rather than measured at that location. A blocky image makes the sampling limit obvious and discourages the mistake. The processing has not changed the physics, but it has changed what the operator is likely to do.
Is there any processing that does recover small detail?
Only processing that adds measurements. Shifting the optics by a fraction of an element pitch between frames and combining the results genuinely samples points the array missed, and so does scanning the instrument across the target and combining frames with known registration. Both need several captures and a scene that holds still. Anything computed from a single frame is bounded by what that frame sampled, whatever it is called.