[ptx] Re: Does enblend impairs quality?

[Andrew C Mihal]

On Fri, 24 Sep 2004, Pablo dAngelo wrote:

> Enblend doesn't blend areas that are not overlapping with other images.
> So differences can only be spotted inside the bounding box of the
> overlap between layers. Blending a single image will just produce a
> copy, I think.

If you just give Enblend a single file, it will not invoke the 
multiresolution spline algorithm - it will just copy the file.

In Enblend versions 1.3 and earlier, floating-point arithmetic is used for 
intermediate calculations, but the Laplacian and Gaussian pyramids are 
stored in integer format to conserve memory. This leads to quantization 
errors that introduce some noise into the output images. The worst noise I 
have seen is 2 pixel values or less. For 8-bit images, this turns out to 
be a problem especially in the green channel. People have posted example 
images with "banding" in areas of the sky. The bands are areas where the 
average green value changes by only one or two. Since the areas have 
well-defined borders, and since the eye is most sensitive to green, you 
can see them. I have written an article about this that I will post on the 
Enblend web page this weekend.

Second, Enblend _will_ change pixels that are outside of the overlap
region. It depends on how close they are to the overlap, how close they 
are to a region where only the other image contributes, and the spatial 
frequency of the region. This is necessary for panoramas because our input 
images have such irregular boundaries and irregularly-shaped overlap 
regions. I'll probably write up an article about this too in the future.

Andrew

---------------------------------
Andrew Mihal
www-cad.eecs.berkeley.edu/~mihal
mihal at eecs.berkeley.edu