7

In the process of trying to estimate whether a ROI contains [Japanese] text, I need to get a rough count of edges, hopefully binned by length and/or orientation... I think this would give me strong-enough features to judge if the ROI is likely to contain text or not.

However, I can't seem to figure out an efficient (this eventually has to work on a video stream) way to get such a histogram (or even just a count of edges for certain length/orientation threshold): Hough transforms will tend to focus on certain shapes (lines etc.) while Canny and Sobel will detect any type, but their output is not particularly useful for counting...

I would appreciate any suggestion or hint at some algorithm I can use, to quickly extract a list of edges from an image...

Dave
  • 173
  • 1
  • 5

1 Answers1

3

I don't know if OpenCV has some implementation - but i can provide you with a pointer of algorithm for this.

MPEG 7 provides a fairly well defined (and well thought out) edge histogram descriptor [EHD] which essentially serves your purpose.

You can refer to this: Efficient Use of MPEG-7 Edge Histogram Descriptor which i think is an excellent paper on this.

If you want to see some implementation of MPEG-7 you can find it on JOANNEUM RESEARCH MPEG 7 page.

For more information on MPEG-7: List of possible image features for content based image retrieval

Community
  • 1
Dipan Mehta
  • 5,587
  • 2
  • 30
  • 53
  • Thanks for the answer. Although it is not quite what I need (the MPEG7 algorithm only gives edge bins for small area of the screen, since it is based on a rather simple subdivision of each area), it is close enough, so I'll accept it... – Dave Feb 20 '12 at 07:48
  • Incidentally, it turned out that the exact algorithm was described (with more detail than the above papers) on SO: http://stackoverflow.com/questions/909542/opencv-edge-extraction – Dave Feb 20 '12 at 07:49