Co-author Sylvain Paris discusses "Local Laplacian Filters: Edge-aware Image Processing with a Laplacian Pyramid," the Research Highlights article published in the March 2015 Communications of the ACM (cacm.acm.org/magazines/2015/3/183587).
00:00 In a flower, what is beauty? To a biologist, it might be this. To a romantic, it might be this. But whoever you are, it's probably not this.
00:18 That ugly halo effect is common to image-processing programs. The problem is that your computer sees edges and details the same. Boost the details, and edges take the hit. Focus on edges, and details disappear.
00:34 Now researchers have reapplied a long-neglected technique to process edges and details separately, with unprecedented results. Join us as we talk with Sylvain Paris about "Local Laplacian Filters".
00:50 [Intro graphics/music]
01:00 Perhaps you've heard of Gaussian blur, a often-used tool for image processing. It works by first reducing the resolution of an image. Details disappear, while sharp edges remain.
01:15 The difference between the blurred and the original image looks something like this. The details are clear, but edges seem distorted. This is the Laplacian filter.
01:25 A stack of Gaussian transformations is a Gaussian pyramid, so called because each level is half the height and width of the previous one. A Laplacian pyramid is similar, but using Laplacian transformations.
01:38 Researchers have combined the two for years. This "Laplacian of Gaussian" filter gives good results -- but still, those darn halos remain.
01:48 Now Dr. Paris and his colleagues found a solution. Rather than simply creating a Laplacian pyramid over the whole image, they get better results when they built those pyramids over small sections -- that is, "local" Laplacian pyramids.
02:03 DR. PARIS: That's how it works is that this idea is that making something look good locally is easy. And finding the one pixel in the pyramid that corresponds to that information is easy. And then once we have that, we just repeat that many many times, and build pixel by pixel, we build a new pyramid, that only the good part of each representation, of each of these local manipulations.
02:25 This technique is especially useful for processing high-dynamic range images, which contain information beyond what display devices can show. These have traditionally been handled by simply compressing brights and darks together.
02:40 DR. PARIS: Here what I've done is say, "O.K., it's too large, let me divide" -- now I can display. But when I divide, I divide everything in the image, and I get this very flat rendition.
02:50 The application of local Laplacian pyramids lets us compress only those places where light and dark come together most dramatically -- the edges.
03:00 DR. PARIS: So there's this one edge which is probably the worst in all the image. But some other images don't have such a strong edge. But then all together, the edges often are a challenge... So if we collectively take all the edges and compress them, it's a good strategy overall.
03:14 The parameters for adjusting these filters have been simplified and recombined in the photography software "Adobe Lightroom", where they appear as Shadows, Highlights, and Clarity.
03:25 The most amazing part is that Dr. Paris seized on local Laplacian pyramids when he figured out that they were partly responsible for the success of another algorithm -- even though nobody knew it at the time.
03:37 DR. PARIS: We were interested in a technique that one of my co-authors published before that's called exposure fusion, which is very good at dealing with these high dynamic range images... And that technique also used the Laplacian pyramid. It just was presented as a kind of detail at the end. "Oh, and by the way, we have this different result, and we do a Laplacian pyramid on them... And just by digging through more and more and more, we found that actually this last step was a lot more important than we initially thought.
04:16 Find out more in this month's Communications of the ACM, in the Research Highlights article, "Local Laplacian Filters: Edge-aware Image Processing with a Laplacian Pyramid".
04:26 [Outro and credits]