Some Technical Questions About Thermal Imaging

[themystick]

Hello!
I am a newbie to thermal imaging. I have a couple of technical questions for my master thesis research. Actually, All questions related to Thermal Imaging Performance issues. I am sure some of the people are very well experienced with my questions. Here:

• I know NETD for detectors measuring at the factory and It's mainly causing the noises, non-uniformities, etc. But when the detector becomes the "Camera", There are too many things to affect the noises. My question is, with the image processing, if you reduce the noises with some filters, algorithms, etc., and increase the Signal-to-Noise Ratio, Will it cause to increase in the NETD performance? (I mean higher sensitivity)
• What is the technical difference between NUC and FFC? (As I know, NUC only controls the Fixed Pattern Noises but FFC controls the pixel to pixel sensitivity when the detector is faced with a shutter. Am I right?)
• As I know, the shutter is mainly used for the FFC. I've readen some articles about the "Shutterless IR Cameras." Without the shutter, Is it possible to do FFC?

Thanks!

[bap2703]

1) Of course you can alter the NETD. Just do a time average across multiple frames to get a better NETD (not more sensitivity, but less noise). You can't do that for an infinite time though because you'll start averaging low frequency noise. Since you're doing a master's thesis you might want to have a look at what's Allan variance. If the camera you have allows you to fiddle with the sensor, there are probably ways to tune the NETD for specific needs.

2) NUC and FFC are the same thing, they are both some form of offset and sensitivity corrections to remove spatial noise. The details might vary. Some might call NUC just getting an updated offset value.

3) Yes you can do a NUC/FFC without a shutter. For example shake/rotate your camera or what's in front of the camera so that it just sees an averaged uniform flux or IR light --> you just got yourself a virtual shutter/flag: the camera is watching a uniform target so the picture you end up getting is only caused by non uniformities in the camera.

[Bill W]

1 - Noise is usually just that, single pixel standard deviation of output.  The normal (and large) non-uniformity of the sensor is of no concern as that is 'the cameras' problem' to sensor manufacturers.
As noise is then divided by response, NETD can be altered two ways; by longer integration at the detector (less mV of noise) or by bias alterations to increase response (more mV per 1°C scene change).

Filters, scene averaging, gamma curves, different displays and colourisation do not change NETD, they change the camera system MDTD. (MDTD = minimum discernible temperature difference)   
MDTD is affected by any uncorrected fixed patterns from the camera system.
A smaller aperture lens does not change sensor NETD, as that is quoted at f/1, but it certainly changes camera MDTD.

2 - As bap notes, NUC can be used to mean a full two point calibration of gain and offset, or used loosely to mean for a single point offset only correction.  FFC is always just a single point offset only correction - 'flat field correction'.
You'll need to be specific and consistent when referring to these separate functions.

3 - yes several ways, with variable degrees of inadequacy depending on the intended use.  Suggest you look out the various patents in circulation.

[themystick]

1 - Noise is usually just that, single pixel standard deviation of output.  The normal (and large) non-uniformity of the sensor is of no concern as that is 'the cameras' problem' to sensor manufacturers.
As noise is then divided by response, NETD can be altered two ways; by longer integration at the detector (less mV of noise) or by bias alterations to increase response (more mV per 1°C scene change).

Filters, scene averaging, gamma curves, different displays and colourisation do not change NETD, they change the camera system MDTD. (MDTD = minimum discernible temperature difference)   
MDTD is affected by any uncorrected fixed patterns from the camera system.
A smaller aperture lens does not change sensor NETD, as that is quoted at f/1, but it certainly changes camera MDTD.

Thank you for your responses!
I clearly understood what you mean. Actually, I know the MDTD and, the difference from NETD. Why did I ask in this way? I noticed that some camera manufacturers specifically mention that in the camera datasheet, for NETD value "30mK<NETD". But when I take a look at the sensor specs used in the camera, the sensor manufacturer mentions that "50mK<NETD". I think It's a trick. I thought maybe they may improve the NETD by decreasing the noise density but as I understand from your statements, It's not true. Probably they are sharing the MDTD as NETD or They are doing a trick for the NETD test conditions.

[themystick]

1) Of course you can alter the NETD. Just do a time average across multiple frames to get a better NETD (not more sensitivity, but less noise). You can't do that for an infinite time though because you'll start averaging low frequency noise. Since you're doing a master's thesis you might want to have a look at what's Allan variance. If the camera you have allows you to fiddle with the sensor, there are probably ways to tune the NETD for specific needs.

Thank you for your responses!
Actually, as far as I know, these noises can occur for different reasons. For example, TEC performance, the temperature gradients of the PCB connected to detector pins or even the ADC performance of the detector. For my thesis, I am trying to find out raw infrared images, without any filter, fully noised digital uint8 values. But I couldn't find that yet. So, I decided to add some fixed pattern noises and white noises to the digital image data. After that, I am putting the NUC(CNN Based algorithm) and some filters like guided, Butterworth, and low/high pass. In that way, I've increased the PSNR, and also other quality measurements like SSI, QI, Entropy, etc. look very good. On the paper, it seems like working but I didn't try with a "real" detector. I am just trying to understand whether I am in the right way or not...

[Bill W]

Thank you for your responses!
I clearly understood what you mean. Actually, I know the MDTD and, the difference from NETD. Why did I ask in this way? I noticed that some camera manufacturers specifically mention that in the camera datasheet, for NETD value "30mK<NETD". But when I take a look at the sensor specs used in the camera, the sensor manufacturer mentions that "50mK<NETD". I think It's a trick. I thought maybe they may improve the NETD by decreasing the noise density but as I understand from your statements, It's not true. Probably they are sharing the MDTD as NETD or They are doing a trick for the NETD test conditions.

There are several explanations.....

1 - Just plain lying

2 - Measuring noise at the camera output after processing and with a 'well chosen' scene and 'correcting' for viewing the screen (eg human eye integration time).

3 - Guessing the noise at the camera output given (some of) their processing.  Most are very adept at 'forgetting' to correct for the lens that they are using.

4 - The sensor manufacturer will specify the guaranteed factory test limit.  The camera builder may use the average value of the sensors that they have had delivered recently and maybe 'accidentally' forget about the '<'.  Most marketing types think '<' means 'typical' or 'at least 51%'
As sensors evolve through product life, NETD may reduce a lot but the spec never gets changed.

5 - The sensor is being run at such different conditions to the factory test that this really is the NETD as set up in the camera.  Unlikely as otherwise the sensor builder would use those conditions for testing.


All of these come down to two basic problems, that actual testing of a camera is quite difficult and that most tenders are written and evaluated by accountants who will simply discard the worst NeTD's.

[Bill W]

Thank you for your responses!
Actually, as far as I know, these noises can occur for different reasons. For example, TEC performance, the temperature gradients of the PCB connected to detector pins or even the ADC performance of the detector. For my thesis, I am trying to find out raw infrared images, without any filter, fully noised digital uint8 values. But I couldn't find that yet. So, I decided to add some fixed pattern noises and white noises to the digital image data. After that, I am putting the NUC(CNN Based algorithm) and some filters like guided, Butterworth, and low/high pass. In that way, I've increased the PSNR, and also other quality measurements like SSI, QI, Entropy, etc. look very good. On the paper, it seems like working but I didn't try with a "real" detector. I am just trying to understand whether I am in the right way or not...

The following will affect (true) NETD:
TEC temperature setting (ie detector temperature)
ADC quality if internal and used for the NETD measurement

These will only affect camera MDTD, if that:
ADC quality if external.  The NETD testers i have seen have a 19" electronics rack for the low noise biases and ADC's.
Temperature gradients on the sensor.

As for raw images, do the images shown in my post here:
https://www.eevblog.com/forum/thermal-imaging/deer-photos-and-therm-app-pro-data/msg3633615/?topicseen#msg3633615
Seem useful if I dug out the 14 bit sensor data ?

[AkiTaiyo]

This might help..
Heres 202 sequential raw data (ADC reading) images from a very noisy unspecified thermal camera at 7.5Hz.
Half of them are with the shutter closed, half are with the shutter open.  All in CSV files..  Its 35MB so I've put it on some dubious looking cloud host site.

https://easyupload.io/dznocz

[IR_Geek]

NETD (or sometimes NEDT) is a good 'rule of thumb' benchmark for IR systems.  Unfortunately it is mis-used for marketing.  No offense to sales reps, but they are usually not very technical (nor is the general public).  Your school should have free access to standards organization ... ?? at least I hope they do.  If so then look up "ASTM E1543-14(2018) Standard Practice for Noise Equivalent Temperature Difference of Thermal Imaging Systems"    I don't have handy access to my copy, but if memory serves it states NETD is intended for older scanning systems.  It has been adopted (poorly) by modern staring focal plane arrays (FPA).   Modern FPA's are better served by 3D Noise measurements which break the sensor down into "7 orthogonal components including both temporal-varying and fixed-pattern noises." (see 1 below) .  This is better served to understand where the noise of the system is coming from.  The paper below used a term called 'monomania' ... basically reducing something very complex into a single value.  Rarely will vendors provide exact details on how the measurement was performed, so NETD is really only good if you measure all devices the same way ... usually with the same equipment.


(1)  Patrick O'Shea and Stephen Sousk "Practical issues with 3D noise measurements and application to modern infrared sensors", Proc. SPIE 5784, Infrared Imaging Systems: Design, Analysis, Modeling, and Testing XVI, (12 May 2005); https://doi.org/10.1117/12.604588

[IR_Geek]

Fraser has some good info in the post below

https://www.eevblog.com/forum/thermal-imaging/thermal-camera-noise-and-noise-equivalent-temperature-difference-(netd)/msg3550489/#msg3550489

[CatalinaWOW]

I would suggest that all of the above answers are substantially correct.  But differences in detail reflect differences in different application areas or regions.

For a master's thesis I would recommend carefully defining the metrics you will use, and why.  That will require you to understand the nuances of these variations to the problem areas of your area of interest.

[thermalengineer]

High-purity single crystal germanium has a high refractive index, is transparent to infrared light, and does not transmit visible light and ultraviolet light.