Fine vs Super Fine revisited
An issue that has come, gone and come back again is how to think about the jpg qualities "Fine" and "Super Fine" that are available in many Canon cameras. More controversial has been a several year practice of Canon to leave out Super Fine from point and shoot cameras and CHDK's ability, as add-in software, to put back a choice of Super Fine.
The discussion of these quality settings includes some misunderstandings that return to the conversation again and again. Chief among those misunderstandings is the "common sense" observation that a raw photo that has a file size of 13.5 MB, a Super Fine jpg of the same scene of 3.5 MB and a Fine jpg size of 2.5 MB will show a proportional drop in visible quality as the file size goes down because "obviously" detail is being lost in the processing.
One of the things that rarely enters into the discussion is what the process of jpg compression represents as a set of operations on the image file and what the meaning of "quality" numbers or names really is. I tried a brief description of this in 2009 when the cries of angst went up over the SX20 losing Super Fine as an option when it was an option in the SX10:
The first thing to say is that the underlying processing to turn a lossless image into a jpg is not a single crunching by a simple algorithm. It involves at least: 1) changing RGB colors to LAB, 2) using a discrete Fourier transform to decompose the image data into frequency components (various degrees of detail) and then using a pair of "quantization matrices" or "tables" to round off components to simplify the result. There are a lot more details I've left out but which also affect compression and the visual quality of the result.
You can read about that stuff at the International JPEG Group's (IJG) website, Wikipedia and on the very interesting jpegsnoop website if you want to dig into the details.
The main take home messages should be that the visual quality is not linearly related to either the jpg "quality" level or the file size. Some corollaries include that files of substantially different file sizes can be very hard to distinguish when examining them side by side.
What do quality numbers mean?
The IJG defined the algorithms used to create and decode jpg files. In the process they produced a demonstration program to illustrate the process. In that example there are example quantization tables for luminance and chrominance data and a numerical way to get those tables from a quality numbers ranging from 1 to 100. Many jpg creation programs use the IJG's quality routine with little or no change. For instance Irfanview and GIMP both use basically the same quality scheme as does the package of image manipulation routines collectively known as ImageMagick. Whenever you see jpg values expressed as a number that ranges to 100 you can expect that the quantization tables will resemble the IJG tables.
Photoshop uses a totally different scheme with numbers that range up to 12. These roughly translate to IJG numbers and you can read about that in more detail on the jpegsnoop site.
Another take home is that in spite of camera maker's in-camera jpg conversion algorithms, their jpgs can be analyzed and found to be roughly equal to an IJG quality value. Packaged into the jpg file are the quantization tables used for luminance and chrominance. They can be compared mathematically to the IJG tables and an approximate value of quality computed.
What about Fine and Super Fine?
Just as the quality numbers are somewhat arbitrary or completely arbitrary (Photoshot's numbers represent different levels but I'm not aware that there is any simple mathematical translation from the numbers to the actual quantization tables), so are designations of Fine and Super Fine. However, over a large number of Canon point and shoots, Fine represents quality levels of about 93 for luminance and 88 for chrominance. Overall that is about like 93 since visually luminace is more important to the perception of detail. Super Fine for most Canon point and shoots has quality levels of about 97 for luminance and 95 for chrominance or an overall quality level of about 97.
In my 2009 forum post I did a comparison of a lossless image I created by drawing a curved line on a white background and then making jpgs of 97 and 93. The point was that this kind of high contrast edge is a challenge for jpg compression and yet it is difficult to decide which of the images is better.
What is the best way of comparing images of different quality settings?
It's nice to start with images as identical as possible. That was achieved in my earlier curve test by starting with the same lossless image and saving it at two compression levels. With a camera, that can be approximated by using the same exposure of the same scene saved by the camera in the two quality settings. I submit that side by side comparisons of those images will be almost impossible to distinguish in a test where you choose without knowing the quality levels.
A better way is to put the two images into two layers of an image editing program (I use GIMP) and then click the "visibility" of the top image off and on to switch from seeing one or the other. Fine and Super Fine are so similar that it is important that the two images be in nearly exact registration. There's a registration alignment tool in GIMP that shifts and even rotates images to get virtually exact registration. The problem with that is that it is doing even a slight rotation by crunching pixels and you don't want another variable creeping in.
Instead of automated registration I shifted the top image in 2 dimensions while it was 50% transparent. That alignment is good enough to compare. Then I made the top 100% visible (0 transparent) and I could compare the top from the bottom by clicking on the visibility icon of the top image. All this may be gobbledegook of you aren't familiar with GIMP but hopefully you have access to a layering image editing program that allows you do the same thing if you're interested.
What did I find in pixel-peeping?
When I first got my SX50 HS, it was raining hard outside and night was falling. I set up a small shelf top still life and a tripod and shot some images including one Fine and one Super Fine at the same exposure. I looked at them at the time and decided they were not very different but I didn't compare them carefully as described above. This morning I did that comparison when thinking about the issue (it's raining outside again).
Getting the alignment of the upper and lower images was pretty easy using some high contrast text on a clock face. Switching back and forth by clicking on the visibility icon of the top image I could definitely see differences at 400% magnification. At ISO 80 these images still have a little residual noise and that shifted around as I clicked. In all visually interesting areas of the image (a stuffed toy, a couple of labeled packages, a Dilbert doll) the differences were there, but it was not possible to tell which image was which clicking the upper one visible and invisible.
Aha! The numbers on the clock were high contrast black on white. I went to that area of the image and repeated my off and on clicking. There! I found one had slightly less fuzz around the high contrast line between white and black (at 400% again). At 100% the difference was back to being indistinguishable.
But I had my answer for at least these two images.
I had images 0008 and 0009 aligned in two layers of GIMP but I had totally forgotten which was which. But I did know that 0009 when viewed at 400% magnification was slightly less noisy along the high contrast edge of the numeral "8" from the clock in the scene. I looked back in the file folder and 0009 was the Fine version. What??? It turns out that some other text on a side of a bottle was colored on white and on contrasting colored backgrounds. There the noise (at 400%) at the high contrast edges was slightly more noisy in the Fine image. Again, at 100% it was not possible for me to tell which image was which.
My test was of two photos taken under controlled conditions from a tripod, indoors and of a pretty bland scene of a still life in front of a textured off-white wall. The difference between Fine and Super Fine was only noticeable at very high magnification of high contrast text. Interestingly the Fine was slightly better for black on a white background and Super Fine was slightly better for colored text on white or colored backgrounds. It is not clear why--but it might have something to do with the slight differences in quantization tables for chrominance and luminance. Basically it makes me have no concern whatever that some modes of the camera may produce only Fine and not Super Fine.
I will also repeat my opnion that for a "busier" scene with more detailed components of varying color and luminance a blind test will probably yield a correct guesses between Fine and Super Fine that statistically are indistinguishable from chance. I say that not withstanding the knowledge that lots of folks here are convinced that they see a difference. I think that if they fully control the situation and do the test blind they may come up with another answer.
|Post (hide subjects)||Posted by||When|
|Jan 25, 2013||3|
|Jan 25, 2013|
|Jan 25, 2013|
|Jan 25, 2013|
|Jan 25, 2013|
|Jan 25, 2013|
|Jan 26, 2013|
|Jan 26, 2013|
|Jan 26, 2013|
|Jan 26, 2013|
|Jan 27, 2013|
|Jan 28, 2013|
- Canon EOS M58.8%
- Panasonic G85/G803.3%
- Panasonic FZ2500/FZ20001.9%
- Panasonic LX10/LX151.2%
- Panasonic GH5 development3.6%
- Sony a99 II15.9%
- Nikon KeyMission 170 and 801.0%
- Fujifilm GFX 50S development28.3%
- Olympus E-M1 II development18.7%
- Olympus E-PL80.1%
- Olympus 25mm F1.2 Pro1.5%
- Olympus 12-100mm F4 IS Pro1.9%
- Olympus 30mm F3.5 Macro0.1%
- Sigma 85mm F1.4 Art3.6%
- Sigma 12-24mm F4 Art2.6%
- Sigma 500mm F4 DG OS HSM Sport2.4%
- YI M12.2%
- GoPro Hero50.8%
- GoPro Karma drone2.2%