Newsgroups: sci.image.processing
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!gatech!sdd.hp.com!hp-pcd!hpcvsnz!briand
From: briand@cv.hp.com (Brian Dixon)
Subject: Re: object-searching in an image
Sender: news@hpcvsnz.cv.hp.com (News )
Message-ID: <D9IB3C.AIK@hpcvsnz.cv.hp.com>
Date: Thu, 1 Jun 1995 18:23:36 GMT
References: <3q1to7$rfe@hammer.msfc.nasa.gov>
Nntp-Posting-Host: hpcvsgen.cv.hp.com
Organization: Hewlett-Packard
X-Newsreader: TIN [version 1.1 PL9.4]
Lines: 62

Not sure what you mean by 'enhance radiographic images by using a template'.
Are you saying that you have a priori knowledge about what is in the image
and you want to use that to remove the noise signal?  If your template is
full greyscale, e.g. not just an edge-transform image, you can remove the
object of interest from the noisy image through the use of frequency space
(Fourier transform)...but your template has to match exactly in size and
orientation etc.  In other words, the template should be as close as possible
to an exact reproduction of the original image, but without noise.  The
object-less image can then be removed from the noisy image.  None of this
sounds to practical though, now does it?  A better way is removing the noise
directly, but you need a priori knowledge on the format/source of the
noise.  The better it can be defined, the more successful you can be when
you remove it.  Another technique would be to use convolution-based methods
such as one of the low-pass filters to smooth the image (and blur), followed
by a high-pass filter to sharpen it.  If you choose linear first order
filters, then you don't get the shifting of edge locations that other 
methods will give  you.  There are other noise removal approaches also.
You can design any number of kernels of varying sizes designed to enhance
values in some regions while un-enhancing others.  All of this is very
difficult to guess at without actually seeing your images, your templates,
and having more knowledge about what you are calling 'noise'.  By far, the
easiest thing to do is to try to improve the source of the images so that
the noise, distortion, and any other corruptions are minimized.  When 
people come to me and ask what I can do for crummy images, I almost always
send them back to the drawing board and try to get better ones to begin
with by improving optics, lighting, fixturing of parts etc.  If all of that
is taken care of first, *then* I take a closer look at the image processing
(which quite often falls out easily once all of the above is taken care of.)

So, maybe someone else out there is more inspired than me and can come up
with more ideas.  If it were up to me the first thing I'd do (and I'm happy
to do it) is to take a look at example images and templates and to get
a detailed explanation of your lighting/optics/camera/acquisition hardware
setup.

Brian

In article <3q1to7$rfe@hammer.msfc.nasa.gov> you wrote:
: This was a good explaination of the general method.  Your experience may   
: provide an answer to a similar question.  Rather than detect similar   
: features to a pattern, I'd like to enhance radiographic images by using a   
: template, much like you described, but to improve the appearance of the   
: image which is often noisy, without compromising the scientific   
: information in the image.  The pattern for the template would be an edge   
: enhancement, but of a features of similar size and orientation where one   
: end is round, the remainder is straight sided.  These are shaped like the   
: outline of the ends of your fingers on one hand. 
: It's one thing to say you can see the features with your trained eye, but   
: another to print these images out (with the attendant losses) and have a   
: layperson still see the features. 
:  
: --  
: William F. Kaukler, PhD., Met. & Mat. Sci.          phone 205-544-0693 
: The University of Alabama in Huntsville 
: Center for Microgravity and Materials Research; Huntsville, AL 35899 
: and at NASA, Marshall Space Flight Center, Alabama 35812 



--
Brian Dixon, Machine Vision Engineer, Hewlett Packard (Corvallis, Oregon)
503-715-3143 (wk), briand@cv.hp.com (email). "Opinions & attitudes are mine!"
