Monday, January 23, 2012
Noise Reduction RT vs 'the others'
Two months or so back I had this post all planned out The 'other' was to be my antiquated version of Adobe Camera RAW that came with Photoshop CS3. When I compared the noise on two of my images RT clearly lost with one image but tied with the second. Slightly confusing.
So I checked the RT forum to see if there was some info I had missed. There I learnt that the noise reduction routines were being rewritten. I decided to wait for the revision. Then life interfered with blogging and I didn't come back to this until a few day ago.
mbod had already done the comparison where his 'other' was NeatImage. The link for the discussion is http://www.rawtherapee.com/forum/viewtopic.php?f=1&t=3708 He provided this screenshot; I measured the noise in the sky with ImageJ
The interesting thing about these measurements is the noise profiles are nearly identical even though the RT processed image on the right looks much noisier. (click to view original size) That's caused by the relatively large chromatic blotches emil talk about in the forum post. They are what you see but since a line profile only measures noise that is a pixel wide, it makes the Signal /Noise look better than it really is.
Normally I do most of my photographic work on my 32 bit desktop machine. While I have 32 bit memory management problems, I also have a semi ancient Sun workstation monitor that was built to last.
And has lasted. With a large 1600 pixel screen, the monitor can be calibrated and color managed so the colors and values on the screen are identical to those I see on my prints. Just as important when I raise or lower my head, the image intensities don't change like they do on my laptop screen.
But this time I used the 64 machine after I downloaded mbod's RAW file. And I was surprised by what I saw after I ran it through RAW therapee.
The noise profile of the 64 bit jpg (left) was much less noisy than I expected. More important it didn't have the chromatic blotches that I have in my 32bit jpg (right) and that mbod had in his conversion. (click the screen captures to view their original size). Big shock to discover there was that much difference between the two versions of RT.
Both jpgs were converted using identical settings. The formula used was exposure compensation to correct for underexposure followed by RL sharpening with a radius of .50 and amount of .24. This was followed by impulse noise reduction of 75 and luminous noise reduction of 49. All other settings were left at the default values.
glascort also posted a NeatImage conversion which I compared to my 64bit jpg. While my noise is 4 times lower I lost fine detail like the guy wire on the tower (circled in red). Larger detail like that on the side of the building (circled in blue) wasn't affected. But there is no question that my version is less crisp than the glacort's version.
That turned out to be rather easy to fix. I loaded the 64bit jpg into RT and upped the Lab mode contrast to around 30. That brought out more noise so I added a bit more impulse noise reduction. As you can see in my reworked jpg on the left, the noise is still lower than the NeatImage jpg, the overall image is as crisp and I recovered the lost guy wire (circled in red).
While I don't have any of the 'others' installed on my 64 bit machine yet, I feel the current 184.108.40.206 64 bit noise reduction is quite good. So we might be far closer to overtaking ACR than some believe.