4

Going through this paper on Robust text detection which talks about using MSER (Maximally stable external regions) to locate regions in an image as it is typically more robust than the usual contour-finding connected component analysis we'd been doing so far.

The MSER is mentioned in section 2.1 of the text-detection paper (in case you actually checked it out :D) from where I can't figure out how to apply/implement this in OpenCV. The documentation doesn't have any examples so its a bit hard to actually get any code running.

Has anyone tried this before?

The code from an implementation is there however there's some color conversions and stuff that's slightly confusing. Could someone help me figure out what exactly is done in MSER and how to put that in terms of OpenCV? Thanks in advance.

EDIT: Found a MATLAB implementation here.

AruniRC
  • 5,070
  • 7
  • 43
  • 73

2 Answers2

4

OpenCV already has an MSER implementation since 2.1

Here is the c++ documentation. The C function is cvMSER I think, but that could be wrong, haven't used that for a while.

Siarhei
  • 2,358
  • 3
  • 27
  • 63
Darcara
  • 1,598
  • 1
  • 13
  • 33
  • I've seen the documentation. Which is not exactly overflowing with information. :) – AruniRC May 27 '11 at 13:25
  • 1
    The OpenCV version is a true linear time implementation and licensed under BSD, whereas the VLFeat is GPL. – Doug Oct 12 '11 at 21:47
3

I frankly recommend the VLFeat MSER implementation that you found. I've used it in the past to good effect.

Note that VLFeat implementation has C/C++ for the core numerics with both a C and Matlab API. This allows you to do rapid prototyping with Matlab and later on switch to C easily.

For the purposes of experimenting with feature detection/description, I personally prefer that library to OpenCV.

peakxu
  • 6,667
  • 1
  • 28
  • 27