Re: Classification problem question


"lpm" <lpmickens@xxxxxxxxx> wrote in message
news:1172006838.317838.112770@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
I'm working on classification problem in which I want to recognize an
unknown image sample as belonging to one of several known classes.
.. and How do you know when you have enough training and testing
data to obtain a robust classifier? Or a better question may be: What
is the best way to determine how many training & testing samples are
needed to provide statistical significant testing results?

Interesting question, with a not immediately obvious answer. Clearly its
more than 1, and less than say several thousand.

I'd say it depends on the nature of the class, and the type of
classification algorithm you are using. You need to know what the class
represents in a known image, in terms of the number of pixels as a
percentage of the total image area. From the class definition you can also
get some idea of the range of pixel values in each vector of the image (ie
R, G and B channels, or more if you have them for multispectral images).

The comon iterative approach is to select some training areas (TA) , run the
classifier and see how many mis- or un-classiffied pixels you get according
to your result reuirements.

You will find that past a certain TA pixel count the classification results
don't improve. The simplest approach is to take a series of TAs that appear
visually similar, from different parts of the image, say in the order of
200 - 1000 pixels or so per class, or about 1 - 2% of the image pixels,
whichever is larger.

A simple parallelipiped classifier needs less TA pixels but is less accurate
than a max likelihood classifier (more compute intensive) that works better
with more TA pixels when the class signature is developed.

PiXCL 6 includes a multi-spectral parallelipiped classifier and
multi-spectral max likelihood classifier from signature functions.
--
regards,
Stewart DIBBS
============================================
Image Acquisition and Advanced Image Processing
PiXCL 6.1 is very affordable and available now !
www.pixcl.com Gatineau, Quebec, CANADA
========================================




.