Re: wavelet for sharpening/enlargement: who does it?

"Brannon" <brannonking@xxxxxxxxx> wrote in news:1158290995.705748.9510
@d34g2000cwd.googlegroups.com:

It seems to me that sharpening an image using wavelets would require:

1. convert to the wavelet domain (one pass should be fine so you end up
with four quadrants)
2. erode the high freqency quad with a kernel the size of an "edge"
3. erode the two mid frequency quad with a kernel half the size used on
the HF quad
4. convert back

Anybody seen such an algorithm before? Mind posting a link to the
paper? It seems that an algorithm like this would reduce noise while
sharpening edges.

It seems this could be used to enlarge images while maintaining sharp
edges: you would know exactly how big to make the erosion kernel.
Anybody seen such a thing attempted?

Thanks for your time,
Brannon



Yes, your idea is right and has been used sucessfully in the past. It is a special version of what is
called wavelet histogram manipulation techniques: instead of processing a signal or image in the
spatial domain, do the same in the wavelet domain, only now you have the option to selectively tamper
with features of a specific morphological size. This way, you can enhance edges or shapes without
sharpening artifacts at much lower scales (e.g. noise).

You can search for references on histogram techniques for wavelet image processing, or review
papers on noise reduction and noise-resistant sharpening filters. However, noise is still a problem,
since only white noise rests solely at very small scales, while other types of noise (for example jpeg
artifacts) affect the image content at multiple scales/sizes.


--
Harris
.

Relevant Pages

  • wavelet for sharpening/enlargement: who does it?
    ... It seems to me that sharpening an image using wavelets would require: ... convert to the wavelet domain (one pass should be fine so you end up ... erode the two mid frequency quad with a kernel half the size used on ...
    (sci.image.processing)
  • Re: wavelet for sharpening/enlargement: who does it?
    ... convert to the wavelet domain (one pass should be fine so you end up ... erode the two mid frequency quad with a kernel half the size used on ... adapted to edge discontinuities in 2-D. ...
    (sci.image.processing)