Dear 123jack,
Thank you very much for your kindly reply.
I tried the sobel edge detection by matlab but don't know how to get the region I am interested. The results by edge detection are some small non-connected segments.
As my know, the common features of all pictures provided to me are:
1) the ellipse;
2) the line at the right side in the ellipse region.
Except these two features, all other features are different, include the background inside or outside the ellipse region, the disturbs inside the ellipse region.
As you said, I am trying to mask out everything outside the ellipse region. However, I still can't find a general method that can seperate the ellipse region when the background changed. For example, there are another three pictures that I can't seperate them.
For the 'cymk seperation', I only know RGB to CYMK convert, could you please tell me some detailed information about CYMK seperation method, or some link?
Thanks,
Tang Laoya
Hi Tang
Look for sobel edge detection as a starting point (its very much a staple).
As someone mentioned above - talk to whoever generates the image to ask them if they can standardise every aspect of it as much
as they can. This is always the first step as it reduces complexity of coding the most. They may not help but do ask.
I know the thread I pointed to is a bit long but the problem discussed is in effect the same as your own.
The solutions and methods are just the same but your problem is far easier than a random fracture.
What I'd do is:
Start by defining the range of different results and identifying what factors will always be present - those are what you look for.
Then think about how you could mask out everything you don't want
Then think about how software could find what you do want.
Off the top of my head looking at the couple of samples provided the first thing I'd do is mask out the right half of the image
then (because colour is undefined you say) either do a simple colour seperation or just turn it into black and white
(a cymk seperation is easy to write) Look at the resulting images of a few of your worst samples. Ideas often come from those.
Then do a simple edge detection and see if it approximates a top to bottom connection in the image and choose a probability
if you dont get a definite answer (your curved example might need an approximation if the background noise is bad and grey
and your line is whiteish grey for example)
You can get smarter but its seldom needed on this type of image. (Take a look at a cheap digital camera these days
and how easily they spot faces or smiles or even eye-blinks) - its all done by building up these methods.
It is a very simple image as these things go.