On Jun 12, 6:06=A0pm, Memo <gamali...@[EMAIL PROTECTED]
> wrote:
> On Jun 12, 11:15=A0am, aruzinsky <aruzin...@[EMAIL PROTECTED]
> wrote:
>
> > On Jun 11, 7:23=A0pm, Memo <gamali...@[EMAIL PROTECTED]
> wrote:
>
> > > =A0I have a low resolution image and I have high resolution images
of
> > > some parts of that low resolution image. =A0What is the best way to
> > > combine that information so I can improve the resolution in areas of
> > > the image where I don't have the high resolution information?
>
> > Why isn't it obvious to you? =A0You enlarge the low resolution image
to
> > match the high resolution images and seamlessly clone the high
> > resolution parts into the enlargement. =A0I have done this with a wide
> > angle picture of a cemetery, cloning a closeup of the face of a
> > tombstone so that the inscription can be read.
>
> I think there is a way to get higher resolution in the parts of the
> image for which you don't have the high resolution parts if you have
> some of the high resolution parts. =A0I think you can do more than just
> paste in high res parts. =A0Consider if you only had one line instead of
> an image with some missing points but in other parts there were extra
> points or frequencies. =A0You could set up a fourier transform matrix
> such as
>
> f(t)eiKw1t1 + f(t)eiKw1t2 + ... f(t)eiKw1tn =3D f(w1)
> f(t)eiKw2t1 + f(t)eiKw2t2 + ... f(t)eiKw2tn =3D f(w2)
> .
> .
> f(t)eiKwmt1 + f(t)eiKwmt2 + ... f(t)eiKwmtn =3D f(wm)
>
> lets say you want a fourier transform of 256 points and you do have
> 256 points of f(w) but they're not evenly spaced. =A0You could fit the
> matrix on the left to those points and solve for f(t) and then try and
> get the missing f(ws) from them.
>
> Another approach is to use a sum of integrals of f(t) where each
> integral is over a deltat centered at f(t) and you assume that over
> deltaT f(t) is a constant value. =A0The smoother f(w) is the better that
> works.
>
> I just wonder if there are other approaches. =A0There are problems with
> both the above approaches.
You have probably heard of fractal enlargement, e.g., Genuine
Fractals. Google,
"Iterated Function Systems" "image enlargement"
Roughly speaking, "range blocks" are matched approximately to
reductions of larger size "domain blocks." In an iterative process,
an enlarged image is reconstructed from pieces of itself; the range
blocks of one iteration are formed from the domain blocks of the
previous iteration.
Now comes my idea. Paste in high resolution parts after each
iteration. I expect that will improve the image quality in other
areas of the IFS enlargement, but I have no idea how much. Maybe, it
won't even converge.
(Don't anyone dare try to patent this idea after this post because, if
not previously patented or patent pending, this idea is hereby put
into the public domain.)
|
