KammaGamma is a website about rather technical image enhancement topics. After almost a year (!) a new posting showed up with a heads-up for a scientific paper (Proc. ACM Siggraph 2008). It was published by an Israeli university in collaboration with Microsoft Research. The university’s web site has a short and fast-paced video introducing the work. It contains images and image sequences showing the result of the algorithm – so even if you can’t follow all of the nerdiness, it shows what they claim.
As far as I can tell, the authors want to be able to adjust different spatial frequencies (example: the overall image structure, course details and fine details) independently of each other. This requires splitting an image into layers that can be processed independently and later – if required – composed back into a single image.
Arguably this sounds like a 2D equivalent for a 1D audio equalizer: you want independent control over bass, mid-tones and high-frequencies. You can use this to rid of all high frequencies and show just the course structure of the landscape (making it, in extreme cases, look like a cartoon). Or you can alternatively sharpen the fine details. In their video, the authors actually also make this 1D analogy (1D graphs), but probably don’t use the audio analogy. Arguably the author want the equivalent of picking up the bass instruments (big drum) while suppressing the piccolo – but while having the drum sound like a hi-fi recording: big drums produce sharp transitions which also contain high frequencies.
In image terms, the challenge seems to decompose the image into large-scale objects and (one or more scales of) smaller details. But without making the large-scale objects look blurry: a building removed of its details should look like a bunch of polygons with sharp edges. Rather than an image taken in the fog (as with a standard low-pass filter). Similarly, enhancing the high frequencies should increase brick texture, without adding halos around the edges of the building. Examples of the latter (a beach scene) can be found here (the new algorithm is WLS, a competitor is BLF).
The paper apparently uses a least-squares fitting algorithm to decompose the image into different layers of detail. Potential applications (incomplete – I have only found time to partially read the original article ;-):
- halo-free sharpening
- creating a graphical version of a photographic image
- creation of High Dynamic Range images with high contrast details
- any image processing which wants to work on details without creating edge artifacts (blurring or ringing)