Newsgroups: sci.image.processing
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!MathWorks.Com!europa.eng.gtefsd.com!howland.reston.ans.net!pipex!uknet!pcl!palmerj
From: palmerj@westminster.ac.uk (Jake Palmer)
Subject: Re: CONTRAST ENHANCEMENT IN F
Message-ID: <CxK3Mu.JGp@westminster.ac.uk>
Organization: University of Westminster
X-Newsreader: TIN [version 1.2 PL2]
References: <941008120358624@lifescan.com>
Date: Wed, 12 Oct 1994 10:54:30 GMT
Lines: 62

Terry Moore (terry.moore@lifescan.com) wrote:
: References: <HURRELLG.6.000A326A@AgResearch.cri.nz> 

: GH> We are having difficulty getting sufficient video  colour separation
: GH> (contrast) between cereal crop plants, weeds and the soil surface.
: GH> You would think green plants and brown soil would separate OK; not
: GH> so.  Soil is

: I have been doing some work with pictographs and petroglyphs (paintings 
: and carvings made on rocks). These images are captured on 35mm slide 
: film and then scanned on a slide scanner, or transferred to Photo CD. I 
: then attempt to enhance them using an image editing program. I have 
: found that on *some* images, working in something other that RGB color 
: space is helpful. For instance, one image of red "paint" on a red\brown 
: rock was almost impossible to see, and working in RGB (using 
: PhotoStyler) was of no help. I found that by changing to HSB color 
: space, copying the saturation (S) channel, enhancing the copy of the S 
: channel, and the substituting the enhanced copy of the S channel for the 
: brightness (B) channel I was able to get a very nice image. What were 
: indistinct smears are now distinct forms. Perhaps this (or a variation 
: of this) may work for you? 
Even better than this is a transformation based on the principal
components transform (assuming that it's the contrast enhancement
you want and don't care about accurate colour rendering).  Basically
you define two regions in the image that are typical of the two
objects you want to have contrast between; then you map the pixel
values as three dimensional scatter plots in your colour space
(Some people say that HSI colour space is better for this than
RGB, but the difference is slight); then you draw a vector joining
the centre of mass of the two scatter plots; finally you map every
pixel in the image onto this vector (perpendicularly) to end up
with (usually) the best contrast your data can give you.

Sometimes (but very rarely) the HSI alone will give you better contrast
but usually this specific version of the principal components transform
will be better.  For a comparison between the effects of starting
with different colour spaces look up some references by Scott E Umbaugh,
or look out for JRPalmer et al. Effectiveness of the principal components
transform for enhancing contrast between closely similar colours, to be     
published in Intl. J. Remote Sensing, or JRPalmer et al. Enhancing 
minimal colour differences, to be published in J. Photogr. Sci.

Finally, what I have said applies where you have already captured
the image as RGB and makes the best of what you've got, by FAR the
best way of producing amazing contrast between your objects of
interest is to slightly modify the imaging system in a very simple
way, as outlined in my last post.

If anyone would like more details, please email me

Jake Palmer     palmerj@wmin.ac.uk

: I have found that different procedures are necessary for different 
: images.


: TDM

: terry.moore@lifescan.com                           Compuserve 70711,1014

: Motorcycles..Fish Tacos..Sword Swallowing..Hot Rods..Festoons..Petroglyphs
:  * RM 1.3 01881 * I tried watersking, but I couldn't find a lake will a hill i
