As an Amazon Associate, Akermin earns from qualifying purchases. Product links on this page are affiliate links.
A total pixel count is an area and the thing it is being used to suggest is a length, which is why the two diverge so sharply. Going from a 160 by 120 array to a 320 by 240 one multiplies the pixel count from 19,200 to 76,800, a factor of four, while the number of samples across the scene only doubles. Any claim presented as a total count should be converted back to the linear figure before it is compared with anything, and then converted again into the angle one element covers, because that angle is the only quantity that decides what you can measure.
Multiplying two sides makes a bigger number, not a finer one
Grids are described by two dimensions and sold by one. The product of the two sides is always the more impressive figure, and it is also the figure least connected to what an instrument can distinguish. Resolving two features as separate requires samples between them, which is a question about spacing along a line, not about how many cells exist in total.
| Array | Total elements | Samples across the width | Relative to the smallest entry, by count | Relative by linear sampling |
|---|---|---|---|---|
| 32 by 24 | 768 | 32 | One | One |
| 160 by 120 | 19,200 | 160 | Twenty five times | Five times |
| 240 by 180 | 43,200 | 240 | Fifty six times | Seven and a half times |
| 320 by 240 | 76,800 | 320 | One hundred times | Ten times |
The final two columns are the whole argument. A hundredfold difference in headline count is a tenfold difference in the dimension that governs whether a feature is sampled. When a listing offers the first of those numbers, the square root is the honest translation, and the square root of a large number is a much smaller one.
An array that states its own size plainly
Bare sensor modules make a useful contrast with finished instruments, because they are sold to people who are going to write the firmware themselves and therefore need the real numbers. The XICOOLEE MLX90640 infrared array states a 32 by 24 pixel infrared array, a 110 degree field of view, a two wire interface running at up to 1 MHz, operation at either 3.3 V or 5 V logic levels, and a noise equivalent temperature difference of 0.1 K at a 1 Hz refresh rate.
Every one of those is a hardware figure. There is no enhanced output grid, no effective resolution and no processed equivalent, because the module hands over raw values and the processing is somebody else to worry about. That is a 768 element device described as a 768 element device. The detector pitch and the absolute temperature accuracy are Not stated in the listing bullets, which is a real gap for a module sold to be built into something, but nothing in the listing pretends the array is larger than it is. It is a modest sensor honestly labelled, and a listing like it is the reference point against which a finished instrument quoting a single large number with no field of view should be read.
Note too what the sensitivity figure is tied to. The 0.1 K is quoted at a 1 Hz refresh rate, which is a reminder that noise performance and speed trade against each other: averaging over a longer interval reduces noise, so a sensitivity figure without a refresh rate beside it is incomplete. A finished camera quoting a sensitivity with no frame rate has left out the condition that makes it meaningful.
Working the field of view back to millimetres on the target
The conversion is short enough to do in your head once you have seen it. Divide the field of view in degrees by the number of elements across that dimension. Multiply by 17.45 to get milliradians. A milliradian covers one millimetre at one metre, so the milliradian figure is also the width in millimetres that one element covers at a working distance of one metre, and it scales linearly with distance.
For the 32 by 24 array described above: 110 degrees across 32 elements is 3.44 degrees per element, which is 60 milliradians, which is 60 mm at one metre. One element of that array, at arm length, covers a patch about the size of a fist. Three elements across a target, the usual conservative requirement for a trustworthy spot reading, means the target has to be about 180 mm across. For its intended job of noticing where a warm body is in a room that is entirely adequate, and for finding a hot terminal in a panel it is not remotely enough.
Run the same arithmetic on a 320 element row across a 40 degree field and you get 0.125 degrees, 2.18 milliradians, 2.18 mm at one metre. The two devices differ by a factor of about 27 in what one element covers, which is a far more informative comparison than the hundredfold difference in their pixel counts. The calculation also exposes the second variable people forget: a wide field of view spreads the same elements over more scene, so field of view and array size trade directly against each other and neither number means anything alone.
Screen pixels, image pixels and detector pixels
A single listing routinely contains three different grids, and only one of them is a sensor. The display has a pixel count, set by the panel the designer bought, and it is usually the largest of the three because small colour panels are inexpensive and plentiful. The saved image has a pixel count, set by the output format, which may match the display, the array or neither. The detector has an element count, which is the one that was never going to be the largest and is therefore the one most likely to be left out.
Three defences work. Check what noun the number is attached to, since a figure attached to the words screen, display or LCD is a panel specification. Check whether a single listing gives more than one grid, because when it does, the smallest is the best candidate for the array. Check whether the figure comes with a field of view, a detector pitch or an optics description, because those only make sense beside a sensor figure and never beside a display one.
Asking what the published count is a count of
The habit that protects you here generalises well beyond thermal imaging. Instrument listings are full of counts, and the useful question is always what each one is counting and over what. A particle instrument reports counts per unit volume, so the sampled volume is part of the figure and a number without it is not comparable. A logging instrument reports a number of stored records, which depends on the interval as much as on the memory. A meter reports a display count, which bounds the resolution of the reading and says nothing about its accuracy.
The same discipline applies when a published reading depends on conditions nobody mentioned. A review such as the PurpleAir Pixel review has to spend its length separating what the instrument reports from what it infers, and a question like whether cooking moves a carbon dioxide reading is really a question about what the sensor is responding to rather than about the number on the screen. Pixel counts are one instance of a general pattern: when a figure is large and flattering, find out what it is a count of, and convert it into the quantity you actually needed before you compare anything.
Frequently Asked Questions
Why is a total pixel count misleading rather than simply large?
Because the quantity that determines what you can distinguish is how many samples lie across the target, which is a length, while the headline figure is a count over an area. Area grows as the square of length, so a number that is four times larger represents a grid only twice as fine in each direction. The arithmetic is correct and the impression it creates is not, which is what makes it an effective way to present a modest sensor.
Does the same problem affect visible light cameras?
The arithmetic is identical, but the stakes are different. A visible camera with a large pixel count is usually limited by its lens and its noise rather than by sampling, and nobody is using it to assign a number to a small object. A thermal instrument is a measuring device, so a sampling limit translates directly into a wrong reading rather than into a slightly soft picture.
How do I convert a grid size into something useful on a target?
Divide the field of view in degrees by the number of elements across that dimension to get degrees per element, convert to milliradians by multiplying by about 17.45, and then the result in milliradians is also the covered width in millimetres at one metre. Multiply by the working distance in metres for any other distance, and by three for a conservative measurement spot.
Why do some sensors have very wide fields of view and tiny arrays?
Because they are built for presence and coverage rather than for measurement of small objects. A wide angle array watching a room for the position of a warm body needs very few elements to do that job, and the wide field is the feature. The design becomes a problem only when the same part is marketed into an application that needs fine sampling, where the element angle rules it out.
Is a stated sensitivity figure a substitute for array size?
No, they limit different things and both have to be satisfied. Sensitivity sets the smallest temperature difference distinguishable from noise once the target fills the element cone. Array size and field of view set whether the target fills the cone. An excellent sensitivity figure on a coarse array means small temperature differences on large surfaces are detectable, and small objects of any temperature are not.
What is the quickest sanity check on a resolution claim?
Ask what the number counts. If it counts elements across a dimension and comes with a field of view, it is usable. If it counts total pixels, take the square root to recover the linear figure before comparing it with anything. If it counts anything else, such as display pixels or saved file dimensions, it is not a sensor figure at all and should be set aside.