DNG lossy compression test = strange results?

[Redcrown]

I'm evaluating the DNG lossy compression and discovered something strange. Would appreciate it if someone would duplicate my test and tell me if they get the same result, or if there is something flawed in my test procedure. It appears to me that "Lens Profile Corrections" are applied differently between lossy and lossless DNG files.

I take a normal, uncompressed DNG file, load it in ACR (7.1) and save a copy with "Use Lossy Compression/Preserve Pixel Count." I then load both of these in Photoshop, and layer one on top the other. Usually there is no visible difference, so I put the top layer in Difference blend mode, display the histogram, and read the StdDev value. I get StdDev values between 0.6 and 2.5.

That's quite a range, but I assumed variables like noise would account for some difference. A high ISO noisy shot would likely compress differently that a clean, low ISO shot. But with more tests and closer inspection I noticed that my wide angle shots were showing the larger differences (StdDev > 2.0). So on some wide angle shots I put the layers back in normal blend mode, set the display to 500% or greater and toggled the layers to look for visible differences.

To my surprise, it looked like the images had different registration. Portions of the image appeared to be shifted 1 to 3 pixels. And the shift appeared to be greater near the edges than in the center. So I converted some wide angle images with the Lens Profile Correction off. Without lens profile correction the visible shift in registration disappeared and the StdDev value dropped significantly. With lens profile correction on, I get StdDev values around 2.1. When it's off I get StdDev values around 1.2 on the same images.

I have not tested other variables yet, but I'm confused why lens profile correction would be different on a lossy vs. lossless DNG file.

My tests were on Canon 5D3 images with the 24-105 lens, using ACR 7.1 in Photoshop CS6 on Win7.

[madmanchan]

Hi Redcrown,

I've reproduced the discrepancy and will try to figure out what's going on internally.

Thanks,
Eric

[Redcrown]

Waking on old topic....

I just re-tested this issue with the latest Photoshop CC and ACR 8.3.

Looks like it's still there with normal lens correction, but it's also there in a much bigger way with the new "Upright" feature under manual lens correction.

While the normal lens correction shows registration differences of less than 5 pixels, the Upright corrections are giving me differences of 50 to 100 pixels. Significant differences.

Something definitely strange about the way ACR handles lossey DNG files.

[madmanchan]

redcrown, upright is a little different in that it analyzes the picture itself to determine the correction.  It is sensitive to image resolution and pixel values, so enabling upright for a lossy compressed DNG is expected to produce somewhat different results as enabling upright for the original.  This is assuming that you're having ACR/Lr perform its upright analysis "from scratch" on each image independently.  If you're syncing (copy/pasting) upright results from one image to another (e.g., from the original to the lossy compressed DNG, or vice versa) then the results should be well aligned.

[Redcrown]

Thanks for the reply, Eric.

Unfortuately, your suggestion does not seem to hold true in my tests.

I don't have any ultra wide angle shots. Every image I've tested with is from a Canon 24-105L lens at 24mm. So the lens profile corrections are fairly mild.

For the Upright tests I used a couple architectural shots made from a close and low angle perspective at 24mm. The vertical lines in the buildings are slanted at about 20 to 30 degrees, so any Upright correction is pretty strong.

No matter how I set or sync the same Upright correction on a lossy vs lossless DNG, the results are significantly different. And regardless of how I set the Upright correction on either version, the result is the same for that version.

In other words, I can set the Upright correction directly on the lossy version or set it on the lossless version and then copy/paste (or sync within ACR) to the lossy version. In any method, the lossy version is the same (yet always different from the lossless).

I realize this is probably a low priority issue for Adobe, and not likely to be fixed soon. But, for me, it means avoiding Lossy DNG. It also provides more fuel to the anti-DNG crowd.