>Steve and Sangthong,
>
>A real neat application of image recongnition involves holographic
>optical elements. Basically the diffraction of light results in the
>near field and far field spatial distributions being related through
>fourier transforms. One can use a lens to record a fourier transform
>hologram and there are even "lensless" fourier transform methods for
>making holograms. And a holdgram can be made as a multple exposure
>where each exposure represents a different "look" at the object (i.e.,
>rotation, scaling, and even perspective). Just use a unique direction
>for the reference light when each image is recorded. Then this
>holographic optical element is used to analyze the DUT. The HOE will
>recoinstruct the reference light associated with the matching image.
>This has even been used by the FBI for quick fingerprint
>identification. The issues they ran into were related to the poor
>quality of many of the older reference fingerprints on file.
>
>Clay
>
Hi Clay,
Interesting!  I've seen people try to market optical devices for image
cross-correlation before, but the results have always been disappointing. 
The fast processing they hope to achieve always seems to loose out to
problems such as readout bandwidth, resolution, and the like.   But it
sounds like what you are describing is a step above what I have seen
before.  Thanks for telling us about it.
Steve


>

On Apr 23, 11:52&#4294967295;am, "SteveSmith" <Steve.Smi...@SpectrumSDI.com>
wrote:
> >Steve, I've read through the brute fource you recommend, I still don't
> get
> >it in term of correlating with certain prominent data point, how would
> this
> >help with the scaling? Thank you very much
>
> Hi Sangthong,
> I wouldn't quite say that I recommend it. &#4294967295;This brute force method rapidly
> becomes overwhelming in computational requirements. I'm not sure what you
> are referring to, but my guess is that it is some algorithms to select the
> most likely convolutions to perform, based on key features of the data.
> This sounds like an interesting method, but I'm just guessing, I've never
> directly worked in this area. &#4294967295;Good luck!
> Steve

Steve and Sangthong,

A real neat application of image recongnition involves holographic
optical elements. Basically the diffraction of light results in the
near field and far field spatial distributions being related through
fourier transforms. One can use a lens to record a fourier transform
hologram and there are even "lensless" fourier transform methods for
making holograms. And a holdgram can be made as a multple exposure
where each exposure represents a different "look" at the object (i.e.,
rotation, scaling, and even perspective). Just use a unique direction
for the reference light when each image is recorded. Then this
holographic optical element is used to analyze the DUT. The HOE will
recoinstruct the reference light associated with the matching image.
This has even been used by the FBI for quick fingerprint
identification. The issues they ran into were related to the poor
quality of many of the older reference fingerprints on file.

Clay

>Steve, I've read through the brute fource you recommend, I still don't
get
>it in term of correlating with certain prominent data point, how would
this
>help with the scaling? Thank you very much
>

Hi Sangthong,
I wouldn't quite say that I recommend it.  This brute force method rapidly
becomes overwhelming in computational requirements. I'm not sure what you
are referring to, but my guess is that it is some algorithms to select the
most likely convolutions to perform, based on key features of the data.
This sounds like an interesting method, but I'm just guessing, I've never
directly worked in this area.  Good luck!
Steve 

>Hi Sangthong, 
>This is the primary limitation of using cross-correlation in image
>analysis.  The target pattern you are looking for can be scaled and
maybe
>even rotated in the image being examined.  Brute force would simply
>cross-correlate with different scalings and rotations of the target.  Of
>course, this would require a huge computation time.  
>
>The amazing thing is that the eye/brain can do this task very quickly. 
It
>is humbling to see that Nature can process a signal so well, and my DSP
>does it so poorly.
>Regards,
>Steve 
>
>
>
Thank you very much Clay and Steve,

Steve, I've read through the brute fource you recommend, I still don't get
it in term of correlating with certain prominent data point, how would this
help with the scaling? Thank you very much

Hi Sangthong, 
This is the primary limitation of using cross-correlation in image
analysis.  The target pattern you are looking for can be scaled and maybe
even rotated in the image being examined.  Brute force would simply
cross-correlate with different scalings and rotations of the target.  Of
course, this would require a huge computation time.  

The amazing thing is that the eye/brain can do this task very quickly.  It
is humbling to see that Nature can process a signal so well, and my DSP
does it so poorly.
Regards,
Steve 

On Apr 20, 1:23&#4294967295;pm, "sangthong" <sangthong2...@yahoo.com> wrote:
> Hi, guys
>
> I've got a question on photographic correlation for doing inspection by
> cross correlation. I can find sub-image inside the original image, however,
> I don't have an idea if the images have been taken at far different
> distances, so the resolution of the big image and sub-image are different.
> Can anybody tell me theoratically on this? thank you very much.
>
> Sangthong

Hello Sangthong,

Are the subjects being photographed flat (2-dimensional)? If so simple
scaling/translation/rotation will at least allow you to compare the
images - apart from lens distortions. One hopes that barrel/pincushion
distortion has already been compensated for. If the subjects are three
dimensional, then the photos taken from different distances will have
different perspectives unless the lenses are object space telecentric
- i.e. provide orthographic projections. do you know how the pictures
were captured?

Clay

Hi, guys

I've got a question on photographic correlation for doing inspection by
cross correlation. I can find sub-image inside the original image, however,
I don't have an idea if the images have been taken at far different
distances, so the resolution of the big image and sub-image are different.
Can anybody tell me theoratically on this? thank you very much.

Sangthong