How To Magically Improve Your Camera's Dynamic Range


Why DPReviews Dynamic Range Tests Often Don't Test Dynamic Range!


October 27, 2007

I won't hold you in suspense. If you want to really boost the dynamic range of your camera, here's all you need to do. Load an image a JPEG image from your camera into Photoshop and apply a gamma adjustment of 1.6 using the "Levels" tool. That's it. Bingo. You just got a bunch more dynamic range. Now that was easy wasn't it?

What's that? You don't think its that simple? OK, then I'll prove it to you.

The demonstration images below demonstrate this quite clearly. The bottom image is from an Olympus E-330 taken at normal contrast. Using the DPReview method of measuring dynamic range (DR) you can see that the bottom image has 8 stops. The DPReview method pick the step wedge value that is closest to middle gray (127 on the 8 bit scale) and marks that. It then counts the steps up to and including the first step wedge that has a value of 255. In this case, we see nine one-third steps or three stops. Next, DPReview counts downward until it reaches a step wedge that has a value that value that drops below a "black point" about equal to 2% of 255 or until it reaches some noise threshold that DPReview simply doesn't publish or describe. That happens at step number 28. My example test matches what DPReview found almost exactly.

The top step wedge is the magically improved version. As you can see, applying a gamma correction of 1.6 gains us 2/3 of a stop in "highlight range" and a full stop in the shadow range. Wow!! The E-330 is actually a dynamic range champ now with 9.6 stops of DR. Nifty!!

But wait, we can go futher. I did my own "best case" development of the raw file. In the graph below, the blue and purple curves are graphs of the normal and enhanced E-330 tests shown above. The red curve is my "super-duper DR booster" version. I developed the raw in dcraw exporting a linear 16-bit tif file. I then brought that file into Photoshop 7.0 and tortur.... er ... modified it to enhance its DR. Viola! I've got thirteen and 1/3 stops of DR. Amazing! That's as good as a FujiFilm DSLR sensor!!!

Hmmm.... I get the vibe that some of my readers might be a bit skeptical. No way that you can get more than 12 stops out of 12-bit raw file. Right? Well, not really, but that's a subject for another time. I'll give you some strong hints though. The response of the sensor isn't entirely linear and so-called "highlight recover" takes advantage of the fact that the red channel is typically half as sensitive as the green channel. So typical CFA sensors really have a double exposure that allows you to add one extra stop of DR to the "theoretical" DR using so-called "highlight recovery".

But some other readers probably still smell something rotten and if so, they have good noses. They may be wondering how much noise is sitting down there at that last step wedge. The answer is that there's a whole lot! This image shows what the "super-duper DR booster" image looks like. It ain't pretty - even when resized by a whole bunch.

So, I guess I'll take my tongue out of my cheek now. Of course there isn't over thirteen stops of photographically useful DR coming from this camera. That's ridiculous. Just because you can measure it doesn't mean you want what you measured to show up in a photo. But what about my "magical" 9.6 stops? Is that wrong too? Well, I don't think so. Based on comparing the DPReview noise DR test for the E-330 at ISO 1600 which shows the tone curve hitting the DPReview noise threshold, it looks like the DPReview's threshold is about equal to Imatest's 1.0 (Low) threshold. At 9.6 stops, the E-330 is still inside that limit. So 9.6 stops probably isn't an exaggeration base on DPReview's noise threshold. And that means that 8 stops is under-stating what the sensor delivers.

Here are the problems as I see it. First, DPReview does not consistently apply a noise threshold. So when a DR test hits the 2% luminance threshold before it hits the noise threshold, there is actually more DR available from the sensor and you can actually use it if you bring the in-camera JPEG into Photoshop. In these cases, the reader doesn't get a DR test from DPReview. The reader gets a nice accurate report on the camera's tone curve instead. If the test hits the noise threshold and that is used instead, you do have a somewhat reasonable standard of comparison. But that only seems to happen on the high ISO tests, and you have to look carefully at the graphs to determine which tests hit a noise threshold and which ones don't. The noise threshold is a "standard" that isn't always applied. Probably all the DSLR based ISO tests are under-reporting the actual DR because they all hit a tone curve dictated luminance threshold before they hit the noise threshold.

Second, virtually every camera applies some noise reduction and some sharpening to its in-camera JPEGs. These things affect how much image noise there is and can affect DR test results that properly use a noise (or more correctly, a signal to noise) threshold. So testing in-camera JPEGs is really only sorta valid if you are only concerned with in-camera JPEG performance.

So testing JPEGs can be misleading as heck. And testing them in a situation where the threshold you use (noise level or luminance level) changes from situation to situation just makes a poor idea even worse. I've been picking on DPReview in this article, but they are hardly alone. This same basic criticism applies to just about every other online version of DR testing that I've seen. They all typically use in-camera JPEGS. Even online tests that use Imatest, which does use a noise based threshold, still have the JPEG related issues that go unaccounted for. There will always be the issue of noise reduction, sharpening, and even how JPEG compression affects noise. These things can potentially affects the DR test results. And to be fair, to their credit DPReview does do a couple of raw developments with each DSLR test. These are the best indicators of DR that DPReview provides. Look there to get a better idea of what a camera's sensor limitations might be. But again it isn't always clear whether the DR given is the result of a noise or luminance threshold being used.

If you want a good indication of DR, you should develop the camera's raw files using no sharpening, no noise reduction, and preferably no gamma correction. Then you can analyze the image using a program that will measure the response and measure the noise. There is no need to concern yourself with luminance levels as a threshold. Those values are completely manipulatable. Testing using a basic linear develope raw like this isn't a perfect solution. Everybody doesn't have the same idea of what is or isn't an acceptable level of noise. Furthermore, noise that measures the same doesn't always look the same. But doing DR tests this way will be far better and have far fewer variables than anything being done by online review sites right now. What we have now is a jumbled mess of sorta-sometimes-maybe-right test results.

Here's what the type of test I'm talking about can look like. This report is from Imatest, but you could derive the same kind of results a number of different ways. For instance, you could use the free program ImageJ to get the same basic results. It will just take a lot longer with that program. I used ImageJ and OpenOffice to create the nifty "Tone Curves" graph above.

Imatest gives three noise thresholds. None of them is necessarily the "right" threshold. Though as I mentioned earlier, I think the 1.0 (Low) threshold is probably about the same as the threshold that DPReview is using (when they use a noise threshold). Based on that, you can see that the E-330 has about 10 stops of DR. That makes sense when you consider that the four stop faster ISO 1600 test done by DPReview (that did use a noise threshold) came in at 6 stops of DR. You should lose about one stop for every ISO increase above the sensor's "native" ISO. The number in the test I ran isn't affected by noise reduction, sharpening, or the failure to apply a consistent standard for the threshold in the darker tones. Tests done this way give a much better idea of what a camera can ultimately deliver.

The final point I'd make is that making good pictorial use of nine to ten stops of DR means you will probably have to use, burning, dodging, layering different versions of the image and/or so-called zonal development. If you don't take such steps, you'll probably end up with a dull mess. There's a reason the camera companies aren't typically giving users tone curves that span 9 or 10 stops. That can easily result in flat looking photos. 10 or 11 stops of DR can certainly be useful, but it almost always requires extra effort from the photographer to put it to good use. So don't get too wrapped up in the idea that tons of DR is the be-all end-all of photography. It isn't. Great photographs have been made for decades with far less DR.


Revisions
10-28-2007 Some minor wording changes and editorial cleanup.

You are visitor # since 10-27-2007


If you have questions, comments, suggestions etc. please feel free to write me at:
jay@studio522.com


Copyright © 2007 by Jay Turberville