underexspose or up iso

That was a very interesting review of Scottes' tests and Leo's tests.

My understanding of Scottes' test and Leo's test is that they are both very similar tests. Both tests illustrate that for a given ISO setting, exposing to the right provides greater possibilities for producing a better final image than when underexposing - especially in regard to recovery of shadow detail and in regard to noise. Both tests shot everything in RAW.

My test, on the other hand, was entirely different and should not be compared with either of the above-mentioned tests. My test compared shooting RAW versus shooting JPG with everything else being the same (ISO setting and exposure). The result of my test did not provide any indication of the virtues of exposing to the right - instead, it showed the virtue of shooting RAW versus shooting JPG. Specifically, it showed that in the cases of underexposing in the range of -1 to -2 stops, the RAW shooter always wins, but that at -2 stops underexposure it's more of a draw.

In other words, at 1 stop underexposure, RAW is a clear winner, but at 2 stops underexposure you are "out of gas".

There are two separate issues with underexposing vs. high ISO. The first one is noise levels which I can see has been discussed. The second relates to the way digital files work. In RAW the camera will record a total of 12 bits per colour channel which equates to 4096 different shades of Red, Green, and Blue. We all know that for each increase in stop we double the amount of light. What this means digitally is that the last stop the camera can record will contain 50% of all of the information (i.e. 2048 shades of each colour). The next stop down will contain 1024 shades, then 512 shades then 256 shades etc.

What you can see from this is that if you limit your shot to omit say the top three stops you will only end up recording 512 shades of each colour which is only 12.5% of the information that the camera is able to capture (or at least transmit in the RAW file). This is an issue when we try to correct the image as we will see posterisation of the file (as well as the noise). If the image is captured using the entire 4096 shades (i.e. exposed to the right) then the final image will have much greater tonal information than the underexposed shot.

In short, unless the noise is much greater using a higher ISO (which I don't believe it is) then always ensure proper exposure (or better yet expose to the right!).

Sorry Scottes, I didn't look at your link until after I posted this. This issue has been covered quite comprehensively there.

Even though I agree with "part" of your conclusions, I have to point out some errors in your analysis. First, in RAW there is not 12 bits per color channel - it's 12 bits per pixel. One fourth of those pixels are seeing red levels, one fourth of them are seeing blue levels and one half of them are seeing green levels (the infamous Bayer pattern). So, even though your preamble is somewhat flawed, the conclusion is correct in regard to loss of detail. No one has questioned that - in fact it has been noted explicitly that underexposure will cost us some detail in the shadows.

Your next point was that the noise level is not significantly greater as the ISO setting is increased. All measurements that I have made using my own methods and all measurements that I have seen reported on dpreview and the noise profiling that is provided in NeatImage all show clearly that the noise level measured (for example as a standard deviation) goes up almost double for each doubling of ISO setting. That's just a fact that I don't see anyone contesting - although the toe of that curve is different for different cameras (i.e., some cameras show a noise floor out to around ISO 400, whereas my G5 shows a significant increase just from ISO 50 to ISO 100).

Now, in regard to your conclusion, which is to give the proper exposure or better yet to expose to the right.

Of course, it is good to give the proper exposure, and of course it is even better to expose to the right. That was never in question. The OP was, shooting RAW, which is better, to shoot at ISO 100 and underexpose or up the ISO to 1600?

My test indicates that it is clearly better to up the ISO and shoot RAW rather than to underexpose, and also that underexposing in JPG would be at least as bad as underexposing in RAW and worse in most cases.

The other gentlemen's testing, if I interpret them correctly, says that when shooting RAW at a given ISO, it is better to expose to the right than than to expose normally or to under-expose

And that seems to be the same as what you have concluded.

Robert, I'm not sure if you misread my comments or I wasn't very clear in the way I presented them. What I was trying to point out is that the issue of noise in only half of what should be considered when deciding to under expose an image or increase the ISO in the camera (which is what the OP was asking). The other issue is the loss of data in the three colour channels which will affect the picture quality.

Your comments relating to the Bayer configuration of the sensor or true, however the image is outputted from the camera in 12 bits per colour channel. This is independent of the way in which the camera captures the data. Maybe I was a little vague in my description although I did try to state in my second paragraph that I was referring to the RAW file.

My statement that noise is not much greater in higher ISOs was a bit misleading. What I was trying to say is that the total noise generated by a higher ISO should be bellow the total noise generated by under exposing the image at a lower ISO and then correcting during RAW processing. Obviously with all things being equal, noise will increase as ISO is increased.

I don't think this is true. A lower ISO records less noise. Adding EC in a RAW editor will amplify the existing noise, but it will also amplify the existing signal equally. It does not create more noise. So the "lower ISO/+EC" shot should have less noise than a higher ISO shot.

However, the critical thing to me is the loss of detail in the low ISO shot. Take a look at my example rock shot above and you'll notice that the left shot - the underexposed image - has much less detail. It's very clearly visible in the darker areas on the middle left. This loss of detail to me is far, far more important to me than additional noise. Noise can be removed quite easily with noise reduction software, and with little loss of detail if ones uses a mask to protect the edges. But details can not be added to an image. They're gone for good.

This is simply incorrect Bob. Each color channel is recorded with 12-bits of information stored in 16-bits of data, or 2 bytes. This is 2 bytes per color channel per pixel, or 6 bytes per pixel, which is why my 8 megapixel 20D produces a 48-megabyte TIFF after processing.

There must be some confusion about whether we are talking about the RAW data file coming out of the camera (which then is the input to RAW processing), or the processed data coming out of RAW processing as some other format such as TIF. In my statement I was talking about the data coming out of the camera as a RAW file - not the TIF file coming out of the conversion process.

For example, with my 5.1 MP camera, the RAW files typically run 6 to 7 MB in size, and that is for 5.1 million data points - each of which is either R, G, or B - not all three.

For your 8 MP camera, I assume that your RAW files are proportionately higher, at around 8 to 10 MB, correct?

Sorry, but in the RAW file there is no 3-color output, it is just grayscale values per pixel. The color information is assigned as part of the demosaicing process, which uses the data from the surrounding pixels to interpolate the R, G, and B values for each pixel from its surrounding pixels.

Regarding your contention that "the total noise generated by a higher ISO should be bellow the total noise generated by under exposing the image at a lower ISO and then correcting during RAW processing", cannot be supported. I just ran a simple test as follows:
Image #1 Shooting RAW at ISO 50 and under-exposing by 3 stops.
Image #2 Shooting RAW at ISO 400 and exposing normal.
(that is consistent with your description of the two cases.)

Processing in ACR to obtain identical images and outputting them to a JPG yields two images with exactly the same amount of noise. Based on that test, I don't think your statement can be verified.

OK, I see where you're coming from. In RAW format the data is stored in 36 bits and compressed. Note that no two shots are ever exactly the same size. Also note that high ISO shots are almost always larger RAW files. Both are tip-offs to the fact that compression is being used. Different pictures are always slightly different, thus they compress to different sizes, and the random noise of a high ISO shot makes compression much more difficult, thus the high ISO shots are larger. If the RAW files were not compressed they would *always* be *exactly* the same size - X megapixels of data plus header and EXIF. Oops, EXIF would change things *slightly* but probably just several bytes.

I can assure you that the camera is recording 12 bits per color channel per pixel, and the RAW is storing this.


Edit: At least I'm pretty darned sure about this, but now you have me thinking...

Edit2: Yes, I'm still sure it's 12-bits per color channel per pixel. Canon's page says "12-bit" and it's common to describe an image format by the number of bits per channel, not per pixel. And if it were 12 bits per pixel one would only get a total of 4096 possible unique colors in an entire image. I'm absolutely positive that I have recorded images with more than 4096 unique colors. A simple program that I wrote counted over 330,000 unique colors in one particular image. So I have no doubt that it's 12 bits per color channel per pixel.

Dear Scottes and JasonW,

A very brief but accurate description of what is in a RAW file (as it comes out of the camera and as an input to Adobe Camera RAW) can be found in several references, but the one that is most likely available to members of this forum is "Real World Camera Raw with Adobe Photoshop CS by Bruce Fraser. See page 4.

I do agree that file sizes vary with information content and that lossless compression is likely being employed.

Well I just pulled up an 8-bit TIFF of a processed RAW file into Paint Shop Pro and checked it's "Count Colors Used" tool. It listed 353,742 unique colors. This is simply impossible unless the image was captured and stored as 12 bits per color channel per pixel. Again, 12 bits per pixel yields a max of 4096 unique colors.

OK, I just read up on Bayer Interpolation and NOW I understand Bob.

You're right. I apologize.

CORRECTION:

I did make a mistake in that test - I let the ACR processing do its default sharpening and noise reduction, and in that case the noise was the same in the two images.
I just did the test over and set the sharpening and noise reduction sliders to zero. In that circumstance, you are right, Jason - the noise level is slightly higher for the case of ISO 50, underexposed and then boosted in ACR. I apologize for that error. I think we are in agreement then in regard to the issues of noise.

No apologies necessary - we could still all be wrong - including Bruce

So long as we can give and take and come back to contribute something on the next day, that's good!

In case anyone is interested in my "subjective" noise test here are the two images. They include the EXIF, so you won't have any problem in identifying which is at ISO 50 and which is at ISO 400. They are 100% crops.
So far as my eye can judge, the ISO 50, underexposed shot is the noiser of the two.

HOSTED PHOTO
please log in to view hosted photos in full size.

HOSTED PHOTO
please log in to view hosted photos in full size.