Saturday, February 23, 2019

Enhanced details and what kind of images it helps with

The heavens open
The Heavens open

Adobe has added a cool new geeky feature in the latest release of Lightroom and ACR. These are Classic 8.2, CC 2.2, and ACR 11.2. The feature uses machine learning to analyze and learn from your raw file to eke out the last very bit of detail. For some cameras that have non-Bayer mosaic sensors such as Fuji's X-trans sensors, this is enormously impactful and improves almost every image. However for regular Bayer sensors, you will rarely see any improvement. I estimate that perhaps 1% of my images show any improvement and then you will only see it in humongous prints. The problems manifest in artifacts visible in the standard demosaic and are solved by the enhanced detail feature. For the fall color image below that shows definite improvement when looked at at 1:1, I have a 4 feet high print in my office where you will have a hard time seeing it. So this feature is really for the most extreme pixelpeepers out there. Nevertheless since I am a geek and appreciate the imaging science behind this, I thought I'd show you a few instances where you can expect improvement. These all have to do with the very specific characteristics of Bayer array sensors. A Bayer array sensor as is common in almost every DSLR and also cellphones and compact cameras, has its pixels laid in a way that you have twice as many green sensitive pixels as blue or red. It looks like the below where every square is a single pixel on the sensor and the color indicates different color filters in front of the pixel:

By Amada44 - Own work, Public Domain, Link

The job of a raw converter is to try and intelligently interpolate between the different pixels to generate a full color image. The existing algorithms for this are not always very smart and even Nikon's own raw converter can generate artifacts under certain circumstances. However Adobe has done something extraordinary and uses artificial intelligence to get around this problem. Also, the above figure immediately tells you that detail that only has blue or red will be prone to artifacts. This is precisely where you might find improvement from this feature. Below is a zoomed in view of an image I found in my library that had these exact issues. These are images from a Nikon D600 which is a 24 MP camera. Remember that yellow is really green and red combined. You should click or tap it to see it bigger:

You can clearly see that in the regular image as well as the image converted using Nikon's own software there are very ugly artifacts around the leaves. The enhanced detail algorithm is able to get around the problems caused by the Bayer mosaic sensor lack of detail in blue and red. To make it more obvious here is the two even more zoomed-in views:

Standard Demosaic
Geometric

Enhanced Details:
Geometric

Note that at the scale you are watching this, if you are on a typical desktop computer, it would correspond to a print that is 10 feet high. Again, I want to be clear that enhanced details only matters on ginormous prints. It does not matter for any reasonable size print and definitely not for online images.
Another type of feature where you might see improvement is sharp colored edges against blue skies. Again most visible around red objects. Here is an example from the temple of the sun picture I opened this post with that was taken in Cathedral Valley in Capitol Reef National park at sunrise. A beautiful and not too well known location. Again click/tap for bigger

You can clearly see on the diagonal feature that the old demosaic (right hand side) causes stair stepping artifacts while the enhanced details version (left) has a smooth edge. Again you will only see this in absolutely gigantic prints. What you see me do here is taking pixel peeping to the absolute extreme. Another thing to take into account is that the process of doing an enhanced details creates a new dng file that is about 4x the size of the original file. For the tiny enhancements you might find in a tiny subset of images, it is highly unlikely to be worth it to do this except if you find egregious problems like in the fall leaf image. Even then you would need to print at gigantic sizes.

To conclude this post here is the full version of the leaf image above. You can see that the detail that I used above is tiny. See if you can find where I took the sample.

Geometric
Geometric