Fine vs Super Fine revisited

Started Jan 25, 2013 | Discussions thread
Don_Campbell
Senior MemberPosts: 1,092
Like?
Re: Fine vs Super Fine revisited-addendum
In reply to Don_Campbell, Jan 27, 2013

One thing I left out of my OP is the option in JPEG compression is "chroma subsampling." This is an optional step between the colorspace conversion from RGB to LAB and the application of the quantization tables to the 8x8 blocks of pixels that the discrete cosine transform has converted.

Chroma subsampling is reducing the number of pixels in the chroma data by combining neighboring pixels. The rationale is that the eye is less sensitive to slight changes in color than it is to changes in brightness. Typically when there is no subsampling it is referred to as "1x1" subsampling (1 vertical pixel for each horizontal pizel). This is also referred to in jpg engines as 4:4:4 subsampling. Also possible and used by many camera makers is "2x1" subsampling (2 vertical pixels per horizontal pixel) also called 4:2:2 subsampling. If the chroma subsampling has been applied in both directions it is called "2x2" or 2:2:2 subsampling.

With Photoshop's quality levels 7-12 there is no subsampling of the chroma data. Similarly GIMP and Irfanview default to no subsampling. These are thus 1x1 or 4:4:4 chroma subsampling. When there has been subsampling applied, the quantization tables are applied to that reduced amount of data. In recovering the original image the chroma values are interpolated to recreate the full pixel density.

Camera jpgs

You can learn about both the subsampling and the quality value indicated by the quantization tables of various camera brands/models on the jpegsnoop site. Its site navigation is a bit tricky to figure out but the information is pretty well presented.

Canon's P&S cameras pretty much have 2x1 chroma subsampling and quality settings for Fine and Super Fine that are pretty similar throughout the lines. For models that are not listed, or for an interesting peek into your camera's jpg files you can download jpegsnoop.exe and snoop inside to get the actual quantization tables and subsampling data as well as lots of other stuff.

One thing that is not completely clear to me is whether Fine and Super Fine jpgs have had different levels of noise reduction or sharpening applied. I suspect not but I'm not sure how to tell.

Simulating Fine and Super Fine jpgs

I am starting with an uncompressed raw image that I shot of my little still-life and tweaked slightly in DPP and converted to uncompressed tiff. For the SX50 HS (discovered using jpegsnoop to look at actual images out of the camera) and most of the Canon P&S cameras on the jpegsnoop site approximating the jpeg compression will involve setting the subsampling to 4:2:2 (2x1) and then applying a quality number of 97 for Super Fine and 93 for Fine. (It is less easy to apply different qualities for chrominance and luminances--I'm assuming that the luminance quality is more likely to be what your eye perceives since it is less sensitive to variations in color.)

The advantage of starting from a single image is that the alignment of the subsequent jpgs is exact making the layered comparison fair. From GIMP I exported the image as qualities 93 and 97 jpg with chroma subsampling as applied by the camera of 4:2:2. I then opened the two images as two layers in GIMP.



My still life scene. Most of the images are of crops of the clock face since those show the most visible artifacts.

Note about the images below

The galleries require jpg file format for images. Therefore after making screenshots I saved them as jpg quality 100 and no subsampling of chroma data. I convinced myself that the very small visible differences were still virtually identical to the lossless view I could have obtained saving as tiff. The screenshots of 100% and 1000% magnifications should have arrived in the galleries in a manner that would allow them to be put into layers and compared as I describe below. Likewise, the original of the difference file, although jpg, acurately reflects the differences and can be played with using the levels control to make those differences visible.

The image in the top layer is selected by the visibility icon and so it is visible, blocking the view of the lower image. The images are in perfect alignment because both were created by saving as jpg from the same original tiff.

By clicking visibility of the top layer off and on it is possible to see more clearly the differences in the two jpgs at qualities 93 and 97.



By clicking visibility of the top layer off and on it is possible to see more clearly the differences in the two jpgs. However, at 100% the differences are extremely difficult to see. If you try the same experiment at 100% with jpg quality 40 and jpg quality 97 the difference is more apparent.

The quality 93 jpg at a 1000% magnification (ticks and 5 min dots from clock face).



With the quality 93 jpg invisible, this is the quality 97 jpg below it.

At this 1000% magnification (extreme pixel peeping?) you can see a few more slightly off color pixels around the ticks and the dot. The slight halo around the tics and dot is probably due to sharpening applied when DPP "developed" the original raw image.

With the images loaded in layers and fully aligned, it is possible to take the difference between the top and bottom images pixel by pixel. The difference mode selection actually shows the absolute value of the differences in the RGB values for each pixel location. The absolute value gives the difference therefore as an unsigned magnitude. The result for the full image is shown below.

Black is zero, near zero is pretty black. At this scale it is extremely difficult to see the differences.



At 100% the differences are likewise extremely difficult to see.

By exporting the image of the differences as a tiff file, opening it again in GIMP it is possible to see the differences using the "levels" control. By expanding the pixel values for RG&B from 0-9 to 0-255, the very small differences are scaled up, making them visible:



So, small differences are there in the two jpg images but they need to be expanded in range in order to be seen clearly.

The differences are of course there between quality 93 and quality 97. Both quality levels in out of camera jpgs are also affected by the one dimensional chroma subsampling. I do think that careful control of the exposure would make differences seen in Fine and Super Fine at 100% very hard to detect in a blind test. Not that you couldn't see some difference but that picking which quality setting was used would be very difficult especially if your images were of flowers, leaves, waves, bricks people and birds.

Reply   Reply with quote   Complain
Keyboard shortcuts:
FForum PPrevious NNext WNext unread UUpvote SSubscribe RReply QQuote BBookmark post MMy threads
Color scheme? Blue / Yellow