Looking at Powershot Exposures

I first received my G1 by post in the late afternoon -- tore open the packaging, put in the battery, figured out quickly how to make it do anything, walked over to the office window and made the shot above, my first G1 photo.

As soon as I reviewed the shot, I was happily struck by how much detail had been preserved in the highlight and shadow areas, as shown in the detail areas here. The G1 clearly had a better range than the digicams and video gear I'd grown accustomed to.
I wondered: what are the real limits?
I searched around and shot lots of pictures over the next few days. What were the real limits? It wasn't in the manual, and it wasn't on the Canon website. Time to break out the sense of curiosity and do some tests!

The illustration below left shows a color chart (well, two charts) -- specifically the "Kodak Color Separation Guide and Gray Scale (Small)," which has twenty gray patches arranged as a brightness progression in 1/3-stop steps. Using a chart like this, combined with a series of over- and under-exposures of the same shot, we can chart a characteristic curve to describe the G1's CCD response to different amounts of light. The Powershot Pro90 has the same CCD, so the basic results are valid for either camera.

What's a "characteristic curve"? A simple chart showing how well a particular kind of film (and the Powershot's CCD is a kind of electronic "film") responds to variations in light intensity.

Every film or CCD should expose correctly in the range right around a "correct" exposure -- 18% gray or so. What a characteristic curve tells us is how well the CCD responds to exposures that are outside the "correct" range. How far can we go away from "correct" and still have a useable image? How far can we safely overexpose? How far can we underexpose?

The charactertistic curve I came up with is shown at right. It shows exposures for red, green, and blue, along with a composite "Luminance" curve (this curve represents my camera under a particular lighting situation -- your mileage may vary).

Like most such charts, the scale is logarithmic -- a change in brightness from one f/stop to the next is an increase by a factor of two, so the geometric relationships between intensities are easier to see on a logarithmic scale.

If the CCD were perfect, there would be a direct relationship between the amount of light going into the lens and the brightness of the resultant pixels -- increase the brightness by two, get pixels that are twice as bright. Reduce the exposure, see a matching shift in the pixel intensities. The logarithmic chart would be a straight line.

No CCD (or film) is perfect in the real world. Ultimately, the CCD is limited at one end by the size of a byte -- the Powershot's 8-bit JPEGs will always peak-out at 255, because no intensity brighter than that can ever be stored (the digital equivalent of the darkroom "DMax.")

At the lower end, there comes a point where the CCD just won't respond well to light. The response will either taper off to a tiny value (digital "DMin") or will be lost in low-level noise (a bit like film base fog).

The chart shows that the Powershot's response is very good -- close to optimal within the limits of an 8-bit format. The slight bowing of the curve may be due to glare, given the simplicity of this test.

An important thing to see here -- probably the important thing to see -- is that there's less room at the top of the curve than at the bottom. The "0" column shows us the exposure the Powershot gives to a flat subject -- a blank wall, a gray card, etc. It places these "middle gray" tones in the center of the 8-bit brightness range, and 128 or so (in this test, a little bit closer to 132, but that's less than a 1/6th-stop difference). After only about 3 stops overexposure, that middle-gray will have increased too bright for the CCD to record.

Compared to the highlights, dark areas have greater latitude. I was only able to overexpose the test chart by a stop or two before seeing major loss of detail -- but even at five stops underexposde, the camera still recorded enougfh to be useable in, say, photoshop. Amazingly, even the darkest areas of the chart showed at least a little difference at the darkest test level.

At the bottom of the scale, the overall levels are still balanced, but the colors begin to peel apart -- it looks like the red and blue have been partly inflated by the Powershot firmware to balance the overall gray levels. The color accuracy suffers slightly, but at those low levels the overall gray levels are generally more important anyway.

The lesson here, as with other digicams, is that the CCD should be treated like color slide film: expose with the highlights in mind. You'll only get about 2-1/3 stops of headroom above the base exposure before highlight detail is lost -- but you'll have at least twice that amount of range in the darker areas.

Put it another way: if you overexpose, the image will wash out very quickly. If you underexpose, you've got twice the chance of coming home with a useable image.

The image below is a composite showing the chart patches ranging from two stops over normal exposure to five stops underexposure. I've marked the "middle" exposure with a red dot. Even in this somewhat processed (resize only) image, there's not a lot of maneuvering room in the direction of the highlights -- but even the lowest regions of the shadows have a little detail left.

Details: How The Test Was Done

Very carefully... well not really. It's pretty simple, and I'm sure optical engineers would find my methods quaint and my quality control measures atrocious. Oh well!

I used a few simple tools to do this test. My G1, obviously, and the "Kodak Color Separation Guide and Gray Scale (Small)," which I picked up for about $10 a couple of years back. Rather than strobe, I chose to use a 600W 3200K "hot" light source and set the G1's color balance to tungsten. I don't have a copy stand, didn't feel like re-fitting my enlarger, so the side of the livingroom bookshelf and my thumb had to do. There is unfortunately some glare, and the tungsten light source creates more of the Infrared chromatic aberration than strobe does (in fact, I never seem to get this aberration with strobe at all -- compare the top image of the strobe page, where the strobe head is directly in front of the lens -- no "ca" at all!).

I used the G1's meter with a gray card to determine the basic exposure for ISO 50, then set the camera to manual with that same exposure and shot the tests. I shot a normally-exposed RAW image, then reduced the camera to Small JPEGs (no appreciable difference in the color of the patches could be seen) and ran all the other exposures. I also exposed tests using auto WB, low and high contrast, saturation, and sharpness, but the results here are the "plain vanilla" cases, which will probably account for 90+% of all Powershot shooting (I may do another page some time to describe the contrast results -- suffice to say that almost all the changes are in the shadow contrast, not the highlights. As for saturation, saturated gray is... gray).

The images were then all read into Image Browser, identified, renamed in the Mac Finder and read-into Photoshop 5.5. I used the eyedropper tool tuned to a 5x5 sample, grabbed samples from each patch, and wrote the RGB values down.

Next, I shifted to Microsoft Excel, and copied all those numbers back into the computer. The illustration here shows one of the small spreadsheets I made from the data in my normally-exposed RAW image (converted to TIF based on the white point of the lightest Kodak gray patch, which was apparently a bit bluer than some other patches).

The "L" values use the standard "YIQ" formula found in video, which leans heavily toward green and probably comes very close to the formula used by Canon in the "BW" mode.