[mlpack] Basic advice needed

[Ryan Curtin]

On Sun, Jul 02, 2017 at 06:40:02PM +0800, Andy Green wrote:
> Hi -
> 
> I am interested in using mlpack for a task involving identifying
> characteristic areas of bitmap images.
> 
> What I hope to do is feed it the images and for teaching purposes, show it
> where the characteristic areas are with a second image highlighting them,
> and eventually have it use this information to be able to estimate by itself
> where the regions of interest are.
> 
> It's along the same lines to face recognition, but it's generic 'areas of
> text' recognition.  It's not trying to understand the text, it needs to
> group the areas with contiguous text with a polygon or 'heat map'.
> 
> I should be able to use the C++ library OK but I have no clue what algorithm
> or general gameplan to apply.
> 
> Any advice gratefully received :-)

Hi Andy,

It sounds like you might be able to cast this as a classification
problem, if you can say that the characteristic area is a class (i.e.
'lower left' is 0, 'upper right' is 1, etc., etc.).  In that case you
could just use some standard technique like logistic regression or
decision trees (perhaps) on the pixel values of the image.

That's of course not the only way to do it, but it's the easiest
direction I can think to go given the tools that mlpack offers.  You may
consider using OpenCV for some image filtering algorithms to use before
passing into a machine learning model, but you could also implement
those yourself too.

I hope this is helpful.

Thanks,

Ryan

-- 
Ryan Curtin    | "It is very cold... in space."
ryan at ratml.org |   - Khan