High vs low resolution thermal cameras

[Benchratt]

How much difference is there in the real world between 1280×1024, 640×480, and 384×288 thermal cameras?
And are those “super resolution” modes basically just fancy filtering?

[IR_Geek]

Like in most things in IR, it depends.   If you have the same pixel size, F/#, and FOV then yes you can tell a difference in your raw data as your IFOV (sampling) will be much better for larger array.   It all comes down to the optic used with the focal plane and the specific application/need.

There are some low res systems that will 'upscale' the video to some standard.  That will also depend on the monitor being used and your viewing distance.

[Fraser]

People often approach thermal imaging from the wrong direction, looking at the resolution as the primary measure of a thermal cameras performance when comparing models. There is no doubt that higher resolution is ‘better’ in terms of a cameras potential capabilities, but there is an almost exponential increase in cost when moving from low resolution microbolometers to the highest resolutions now available.

When considering the suitability of a thermal cameras performance for a task, or creating a technical requirements document the following should be considered…

1. Target size
2. Target distance
3. Target Delta T to background
4. Requirement for ‘detection’, ‘identification’ or greater detail (if so, what size of detail)

These basic parameters will then be used to determine the required IFOV and pixel size on the target at the specified distance. From this the microbolometer resolution, lens FOV and system MRTD may be determined. The imaging task requirements effectively dictate the microbolometer resolution, lens selection and system MRTD.

Most generic thermal imaging system designers will try to create a thermal cameras performance that balances the cost of components with desired capabilities and budget. This leads to generic thermal cameras with common resolutions, FOV and MRTD. Examples might be 50 Degree FOV lens illuminating a 320x240 pixel microbolometer of NETD 50mK or a 25 Degree FOV lens illuminating  a 160x120 pixel microbolometer of NETD 35mK. In the latter case, the FOV is half that of ten 320x240 pixel design, but the IFOV is unchanged and the 160x120 microbolometer is offering a better NETD. The down side of ten second system is that at a given distance it will only see of quarter of the scene that he 320x240 pixel system can image at any one time. This may, or may not be an issue, depending upon the use scenario. If best thermal sensitivity is a significant requirement then ten lower resolution system is actually the better option in this fictitious example.

So we have seen that, as IR_Geek stated…. It all depends !

If we take a look at professional thermal cameras designed for preventative maintenance over the years since 1995, it will be seen that many thermal cameras had 320x240 pixels to meet their users needs. A lens FOV of around 25 degrees was common as this gave a good IFOV at the cost of a relatively narrow field of view. In building survey work. The same resolution was used. But the FOV was doubled to 50 degrees at the cost of IFOV and image detail, for a given comparison distance with the preventative maintenance camera. When the level of detail provided by a preventative maintenance camera was required, but over a 50 degree, or greater field of view, users had to consider getting closer to the target or buying a 640x480 pixel thermal camera and suffer the associated significant financial investment associated with such. Thermal CCTV cameras need to be chosen to match the customer requirement. The FOV, IFOV and resolution will dictate the capabilities of the thermal CCTV camera so a 256x192 pixel camera with 30 degree FOV might be adequate for targeted observations but the scene may require a 640x480 pixel camera with 90 degree FOV for detection of intruders over a wide area rather than using two of more 256x192 pixel 30 degree FOV cameras.

The OP has mentioned “Super Resolution. This almost deserves it own thread but I have the following comments……

In thermal imaging systems you cannot beat the best performance ‘front end’ by using a lower performance ‘front end’ and trying to artificially improve performance through digital enhancement techniques. Quality of ‘front end’ data dictates what quality of output may be obtained in professional usage.

So what are the “Super Resolution” modes and how do they compare ?

1.Traditional Upscaling using interpolation. This has been around for a long time in the thermal imaging industry and is a legitimate means of improving a thermal images appearance and resizing the data set to fit a specific resolution of display. The usual approach in interpolation is to create artificial pixel between real physical pixels. This is done by analysing the values of the pixels surrounding the artificial pixel and setting the values of the artificial pixel to match those surrounding it. It becomes a new data point but it remains artificial when considering radiometric measurements. It may be considered a ‘fake’ pixel but it improves the thermal Image presented to the user and can make scene interpretation easier. Interpolation may be more complex and I have seen 16x16 pixel thermal images interpolated up to 128x128 pixels and the image is greatly improved in terms of human interpretation. The Radiometric usefulness of the image remains 16x16 pixels (256 data points) though. Using artificial/fake pixel values created by interpolation for radiometric measurements is fraught with risk and not something that is recommended in exact science.

2. Super-Resolution through scene movement. This image resolution enhancement technique comes in two common varieties…

a) Hand shake - When holding a thermal camera, the users hand is not completely still, unless they are a robot ! The users hand moves ten camera slightly in all directions and this effectively changes the scene registration with the microbolometer pixels. Digital,processing is used to track the users hand shake and the resolution is improved using the real pixels as they effectively create new data points as they minutely scan the thermal scene in the horizontal and vertical planes. The new data points coming out of ten digital processing are real data points and not like those created by interpolation. These real data points become new pixels used for radiometric measurement and display purposes. Note that Hand Shake based Super-Resolution will not work if the thermal camera is mounted static on a tripod or similar. No hand shake means no scanning effect and no resolution enhancement.

b) Scanning prism Super Resolution - This is a resolution enhancement technique that increases the image resolution through mechanical and optical means. Similar to the Handshake approach where new physical data points are created, the mechanical super resolution system employs scanning of the microbolometer pixels over the thermal scene, but only by a very small amount. The scanning effect is achieved using multiple thin prisms mounted on a rotating wheel that is positioned in front of the microbolometer. The prisms are used to shift the beam output from the lens system by a very small, yet precise, amount. The microbolometer pixels effectively shift slightly whilst observing the scene and this creates a whole new set of physical pixel measurements and associated data points. This technique is used on high end thermal imaging systems to improve the true resolution of the system for radiometric activities without the exponential increase in cost of a higher resolution microbolometer array. Even very high resolution microbolometers can be equipped with this scanning system to further enhance their resolution for demanding applications. The mechanical scanner assembly is not an inexpensive component to add to a thermal camera so prices of such cameras remain relatively high. Note that, unlike hand shake based Super Resolution, the mechanical scanner equipped thermal camera may be static mounted on a tripod and still delivers the resolution increase due to ten action of the rotating prisms.

3) AI Driven Super-Resolution. …….. “AI” is a much over-used term these days and is often used as a marketing gimmick to suggest an appliance is in some way sentient. News flash…. Current AI is not sentient. With that out of the way, what is AI driven resolution enhancement as found in some thermal imaging systems ? Note that the systems are “AI driven” and not actual AI as that would take a lot of processing power and hot processors inside a thermal.y sensitive appliance ! Not a great idea and it would also be expensive. In thermal camera AI driven resolution enhancement we see algorithms acting on an image and trying to find identifiable patterns that have been learnt by an AI deep learning system and that was used to create the algorithms. In simplistic terms, the algorithm in the camera tries to find familiar patterns in the thermal image scene and then applies pre-programmed ‘corrections’ to the scene data to ‘improve’ it. The simplest example would be the edge of a house roof that is at around 45 degrees. With such a target, there is often a noticeable ‘stair step’ pattern along the edge of the roof due to the IFOV and resultant pixel size on target. The AI driven image enhancement algorithm recognises this image pattern and determines that it needs to be smoothed to remove the stair step pattern. The AI image enhancement changes the image so that the higher resolution displayed image presented to ten user no longer contains the stair step pattern and the roof edge looks sharper as a result. Note that he changes made to the image data are not from physical image data coming from the microbolometer, so, just like with Interpolation, the artificial pixels and pixel values should not be used for accurate radiometric measurements. AI driven resolution enhancement can be a blessing or a curse, depending upon how the Algorithm responds to a particular scene. The Algorithm needs to correctly ‘recognise’ the read of ten scene that it can enhance with its changes. Sometimes it gets it right, other times it make a bit of a mess of some ares of a thermal scene. In my experience AI driven resolution enhancement copes well with simple thermal scenes that include lots of straight edges. More complex thermal scenes with indistinct boundaries or low thermal contrast lead to the Algorithm making mistakes or effectively giving up and destroying scene detail by applying broad context flattening to areas that originally contained subtle details. AI driven resolution enhancement is still under development and it can improve the appearance of thermal images that it is able to recognise in terms of shapes within the scene but more development is needed and it still will not create useable data points for radiometric measurements that need to be accurate to the observed scene. It is basically a “Pretty picture” mode. Much like Marmite, some users will love its enhancements, and others will hate it and switch it off.

Fraser

[Fraser]

Further to my last comments, the OP asked about resolutions and the difference in the images produced. As IR_Geek and I have explained, the quality of a thermal image is very dependant upon the optical section specifications. A very high resolution microbolometer will not provide impressive image detail if it is married to a very wide angle lens. That said, such a combination is not illogical if the end user needs a single camera to observe a very large thermal scene and is accepting of the reduced image detail. If we think about the more common microbolometer resolutions of, for example, 80x60, 92x92, 120x90, 160x120, 256x192, 320x240, 384x288 and 640x480 pixels, there are limits to what may be achieved with the lower resolution microbolometers and reducing FOV to improve IFOV cannot really compensate for the low resolution as the FOV becomes annoyingly narrow for the user. Once a microbolometer has 160x120 pixels, it becomes a far more useful imaging device. The ‘new’ resolution on the thermal imaging scene is 256x192 pixels and this is a sort of “sweet spot” of microbolometer resolution. It provides decent image detail if married to a sensible FOV lens, yet its production cost is reduced through its small 12um pixels and 256x192 pixel resolution and die size. The 256x192 pixel microbolometers are now in mass production and prices have dropped significantly. Is 256x192 pixels an excellent resolution ? IMHO, no it is not. It can be made to work well with narrow field of view lenses but many manufacturers choose a 50 degree HFOV for general use cameras. For near field imaging, this is OK, but the limits of the microbolometer-50 degree HFOV men’s become obvious when trying to pull details out of a target some distance away from the camera. The IFOV and “pixel size on target” calculations tell, the full story, which is why I now include them in my reviews. A 256x192 pixel microbolometer married to a 20 degree HFOV lens would provide better scene detail, but at the cost of scene area. 320x240 pixels was an older standard resolution that is still used today. For me, this is the minimum standard for any serious thermal imaging camera used in industry. Once again choice of lens dictates the available scene detail. When we move up to 640x480 pixels, users immediately notice the improvement in resolution between cameras using 320x240 pixels Vs those using 640x480 pixels with the same FOV. That said, what users are actually seeing is less steps in edge detail and maybe finer details within the scene. Surprisingly, 640x480 is not a resolution that greatly impresses as it is still relative low Web Cam resolution and image enhancement is still required to make it look decent. Once we get up to 1280x1024, things change dramatically. Those users used to viewing 320x240 pixels microbolometer images suddenly see an image that looks like monochrome CCTV with few image artifacts to give away the fact that it is a microbolometer sourced image. The resolution is not “HD” but it is good enough for many users to consider it a significant step up in thermal imaging quality. All this, of course, presumes that a sensible lens FOV is selected for the camera. There are higher resolution thermal cameras used by the military and some wildlife programs, but they are specialist tools and often offer such enhancements as continuous zoom lenses, low image noise and amazingly crisp imagery. The stuff of dreams fir thermal imaging geeks. Such performance comes at a price though and, in my recent experience, the consumer market is in a race to the bottom on thermal camera prices. Not something that is compatible with high performance, higher resolution microbolometers. A little like the VHS vs Betamax video recorder war of long ago… the Betamax image quality was superior, but the consumers chose to buy the lower image quality VHS recorders as they cost less ! I see that happening in the thermal imaging realm. People seem to tolerate marginal quality provided it is a good price. This is a bit of a nightmare for manufacturers wishing to produce higher quality thermal imaging equipment that has an understandably higher production cost and retail price. Companies like FLIR have commercial and Military markets in which to sell higher quality thermal end thermal imaging equipment. Smaller, less well known, thermal imaging equipment manufacturers are not so fortunate and have to bow to public demand for cheap and cheerful products.
Often consumers still complain about the high cost of thermal cameras and low resolution compared to modern mass produced visible light digital compact cameras 

So yes, the consumer will notice a significant difference between a 256x192, 640x480 and 1280x1024 pixel thermal cameras that have similar FOV, but they will likely still opt for buying the much more affordable 256x192 pixel camera and live with its lower image detail. This is why we are seeing China mass producing thermal cameras with 256x192 pixel imaging cores and not pushing the resolution upwards as that increases the price.

Fraser

[Fraser]

For anyone wishing to see how bad things are with the race to the bottom on price in the thermal imaging market, take a look at my review of the FNIRSI TDM-120P. A decent idea ruined by the use of a 120x90 pixel microbolometer (with 50 degree HFOV lens !) in order to sell the unit cheap. A 256x192 pixel microbolometer core was what the unit needed, and deserved, to be taken seriously in the current market.

https://www.eevblog.com/forum/thermal-imaging/fnirsi-tdm-120p-thermal-camera-multimeter-hybrid-review/

Fraser

[RRtheone]

In Avio R550, there is a SR mode.

Super Resolution (SR) Mode:
Recording Pixels: 1280 x 960 pixel
Spatial Resolution: equivalent to 0.58mrad *1
Standard Mode:
Recording Pixels: 640 x 480 pixel
Spatial Resolution: 0.87mrad

There is a note for this SR mode:
This increased resolution results from detecting characteristic within all frames and aligning them to compensate for small vibrations such as those introduced by hand movement.

https://www.infrared.avio.co.jp/en/products/ir-thermo/lineup/r550/

In my R550, it takes several seconds to process the SR image. You can't record video or live streaming in SR mode.

I cannot upload csv files. Here's the file sizes comparision:
640 IR.CSV: 1.76mb
1280 IR.CSV: 7.04mb