White dots in raw

Software RawTherapee
#1

I have noticed nearly white dots in raw files from Honor 6A phone with opencamera, i don’t understand what they could be (hot pixels??? but there seems to be a structure in them, af points?)

Here is the raw and the jpeg: https://we.tl/YTKvEx7IXa

I attach a screenshot, i haven’t noticed them before, but now that i have zoomed in at 400% i almost always see them. With amaze are less visible, but in such a blue sky are visible too. Sorry for my left finger in the picture!

Thank you in advance for ideas or possible solutions!

Screenshot
Screenshot.jpg1366×768 93.8 KB

#2

I don’t know what causes them, but you can fight them easily by using Impulse Noise reduction in Detail Tab (second tab from left)

Edit: it seems that impulse noise reduction doesn’t remove all of them. Never saw such a pattern of bright pixels before

#3

Thank you, it already does something.

Seems there is a repeating pattern, i really think they are on phase detection autofocus pixels…

Dots
Dots.jpg1366×768 201 KB

and i think they are common with various phone’s sensors, see here:

http://www.dailytech.com/iPhone+6s+Focus+Pixels+Are+Already+Used+in+Galaxy+S5/article36587.htm

IV. Where Apple Shines: Thinness, Sensor Die Size, and OIS
Roughly half a year since Samsung first added it to the GS5, Apple’s new iPhone 6 and 6+ add the passive autofocus method, courtesy of Sony. Much like Samsung, Sony partially masks some of its green pixels for its autofocus. Sony, however, continues to use the much larger 1.5 pitch pixels, which it introduced with the iPhone 5S.

On https://raw.pixls.us/ there are some phones including this Honor, i’ll take a closer look tomorrow.

#4

These indeed look like PDAF pixels …

@heckflosse Ingo, there is an opcodelist2 tag in the DNG exif where these should be marked … with exitool we can use the -b option

http://u88.n24.queensu.ca/exiftool/forum/index.php/topic,8373.msg43012.html?PHPSESSID=91t02ns3brfj1dsubvdsf4ppu1#msg43012

Dng vignetting correction via opcodes?
#5

I think better is the “hot pixel filter” at RAW tab set at 20.

It would be interesting to investigate the behaviour of these pixels … at first sight they look to be all at Blue pixel places but behave as half shaded clear pixels (no blue filter in front, just an opaque screen covering half pixel’s area ) … if this is correct we can easily reconstruct the wrong value with better fidelity than the “hot pixel filter” or the bad pixel filter …

#6

Yes it seems it filters white dots nice, thank you!

I have tried with other raw samples from https://raw.pixls.us/ but they are too noisy to see if there is such a pattern, or it is unnoticeable except the sample from my Honor 6A (dli-l22) that is a 13 megapixel one (on the net there are various opinions about the image sensor, but i think who writes review for phones is mostly a cut and paster without any knowledge)

HUAWEI - DLI-L22 - 16bit (4_3)(2).dng -> shows signs

Apple - iPhone 6s Plus - 16bit (4_3).dng -> no signs
Apple - iPhone SE - 16bit (4_3).dng -> too noisy
HUAWEI - EVA-AL00 - 16bit (4_3).dng -> too noisy
HUAWEI - EVA-L09 - 16bit (4_3).dng -> too noisy
LG - D855 - 16bit (4_3).dng -> no signs
Samsung - SM-G950U - 16bit (16_9).dng -> no signs

#7

Here’s a screenshot which shows the pattern much better. It’s only on blue sensels btw:

grafik
grafik.png1920×1080 139 KB

Same phone, different raw... rgbw and rgb versions?
#8

Ingo, it’s as you guessed … pdaf pixels.

I think we can find the pattern compatible with the camconst.json data but best would be to decode the DNG opcodes …

BTW … these pixels look not exactly like “bad pixels” although the hot pixel filter works (almost) fine on them … I mean that their values are not exactly unreliable but needing a specil care (decrease )

I think (I have to take some measures to be sure) these are of a special kind of pdaf pixels with no color filter and a mask shading half of the pixel … Being clear permits 3X the photons to pass … divide by 2 for the shading mask …
So I think we could reconstruct a proper blue value by subtracting the average of the neighbouring R & G pixels … something like Bcorrect = Bwrong*2 - Gaverage - Raverage

I am a bit anxious about using the average of Blue pixels because they are 2 pixels_spaces away …

Best wishes … Merry Christmas