I'm sorry, something completely wrong happens with image attachment on this forum. I had to delete and rewrite this message to fix images...
Earlier I showed some images, shooted in good conditions, when the scene tempretature difference is high. As you could see, there is no noticeable noise.
Now, let's look at abilities of this camera in poor conditions.
First, so that you can better understand what's going on, here is a photo of bicycle at the wall with a cup of warm water. The room temperature is +25.5 °С , the water in cup is +29.5 °С (measured by external thermometer). Temperature difference of the coldest and hottest objects is about 4-5°С. As you can see, the image is quite good, further we will analyze this scene without a cup of warm water.
The same scene without a cup of warm water. This gif is a sequence of images shooted with a small time gap in a raw mode (no any kind of filtering):
As you can see, the noise gets out, as temperature range is probably about ~1°С. The noise pattern is complex and quite specific. After reading tons of literature and some analysis and I came to think that there are two main noise components:
1. Row noise. This is a temporal noise, i.e. random fluctuations of the whole pixel line, that randomly changes every moment.
2. Column noise. This is a FPN (fixed pattern noise) that changes quite slowly and cannot be fully removed even by FFC (flat field correction) procedure.
Earlier I have developed a special averaging module that helps to average a sequence of multiple frames comming out of the sensor. Max possible number of averaged frames is 262144, though reasonable values are 2, 4, 8, 16, 32, 64, 128. Signal averaging is a common technique, intended to increase the strength of a signal relative to noise. So, let's try to use this module with different averaging values:
We can see that starting from x8 value the row noise disapears, that confirms the nature of row noise. At the same time column noise is still present, averaging do not help to neutralize this type of noise, we need something more.
Few days ago I found a way to remove this column FPN in frequency domain by FFT (fast fourier transform) and a custom filter. As a test scene I decided to use some uniform object. Here you can see how it works. Image from camera (uniform body) is on the left, FFT of this image is on the right. Black lines on the FFT plot are the actual filter function, that "cuts out" the noise components. Here is the link to the online FFT tool that I used for experiments:
https://www.ejectamenta.com/Fourifier-fullscreen/
At last I decided to test both averaging and FFT filters at the complex scene with a lot of details:
Also keep in mind that this experiments are done on equalized 8-bit image. I think that final image quality will be a little better if I apply FFT filter to the raw 14-bit pixel frames.
I don't want to start developing this FFT filter right now, it will be easier to develop it for new hardware and new FPGA. Though averaging is a good filter, it reduces the frame rate, that's why I should implement something else for row noise, of maybe FFT will be enogh, the time will show. At least I know how to deal with this noise. If you have good ideas and suggestions for image filtering, please don't keep in secret
P.S. This is how lepton 80x60 "sees" this bicycle: