Noise, Dynamic Range and Bit Depth in Digital SLRs | Hacker News Reader

Noise, Dynamic Range and Bit Depth in Digital SLRs

theory.uchicago.edu

46 pointsakkartik12 comments

Oooh, thanks for that. I had read it a long time ago and lost track of it. Now I've added it to my Evernote notebook :-)

Back in the stone age I was a lowly systems guy at the Image Processing Institute at USC and we spent a lot of resources computing all of the noise contributors to a digital camera system. The two uses were creating better cameras (that could compensate for their electronics) and identifying which images came from which camera (which could be used in authenticating an images veracity). These days there is so much post processing done in the camera body I doubt the pattern noise signature can be extracted from an image.

I've been using digital cameras since the first wave of them was introduced in the 90's.

Since the start I was making hot pixel maps and that helped fix images using them as a mask in photoshop to reverse the errors caused by hot pixels and non-random noise.

A hot pixel map is simply a long-exposure photo taken that is completely black - with the lens cap on, or the lens otherwise blocked from all light.

I've found that as the sensor ages, the "hot" pixels get "hotter". I take a new pixel map every six months to a year, and have observed the increase in non-random noise.

One of the interesting things that's changed some since that article was written has been the advent of truly itsy-bitsy sensors, for instance on cell phone cameras. Sony, OmniVision, and Aptina are all making sensors on the scale of 1/3", with pixel sizes as small as 1.1µm; the upshot of this is that the electrons wells for these pixels are fantastically tiny, and saturate well below 2,000 electrons.

This means that even on a bright, sunny day, with brilliant opportunities to pour light into the sensor while maintaining short exposure times, we /still/ can't get as much light in as we want, because if we do, we'll saturate the pixels -- and no amount of lowering the analog gain in the readout circuitry will recover the electrons that spilled out of the wells. So, we end up with shot noise as a dominating factor for cell phone images, even in "perfect" conditions.

As we cram more pixels into sensors, noise reduction -- in many different modes -- will become an important battleground in which image sensor and SoC vendors will have to compete. Glad to see it getting some attention :-)

The TL;DR is a little surprising:

First, try to maximize the number of photons collected, by using the largest aperture and slowest shutter that shooting conditions permit. (This isn't the surprising part.)

In photon-rich conditions, sensitivity ("ISO") doesn't matter much. In photon-poor conditions, however -- perhaps you want a small aperture to expand depth-of-field, or a fast shutter to freeze motion -- you get the best noise performance by using the highest sensitivity that doesn't saturate highlights. (This is the surprising part.)

Oh, and there's no point to 14-bit raw formats -- quantization noise at 12 bits is already less than noise from other sources.

Widest aperture is tricky. Most lenses are soft until you stop down a couple times. And with regard to highlights, that sounds like sort of a modern day restatement of Ansel Adams' classic "expose for the shadows and develop for the highlights." You never want to blow your highlights since you can't recover detail from them.

How is that possible? If you expose for shadows, you will blow out highlights before you develop.

Not with Ansel Adams' setup! Shooting straight to JPEG will do that, though.

yep, negatives have a "shoulder" at the top of the exposure curve which might be poor quality, but there is still information there.

Coversely, extreme underexposure means there is basically no information on the negative[

Well, yes. He was just talking about noise performance in the digitization process; he didn't discuss lenses at all.

Interesting, but I wish they would've given more information on how they metered the cityscape. "Underexposed" can be a very relative term, as some parts of a non-uniform image will need more exposure than others to be properly exposed, hence the reason why HDR is a thing.

Interesting article. Not just for the photographic analysis but also how quickly fancy things on the Web date.

"Mouseover to compare" and Mouseover menus all killed off by touchscreens. I'm glad about drop down menus, they are a UI mistake, even on desktop applications; Wirth and Raskin warned us about that.

last update: May 22, 2008

> worth noting (2008) in the title for reference

Reply on news.ycombinator.com