1

I want to improve text recognition accuracy in somewhat complex image.

I'm currently using following sample code:

https://github.com/opencv/opencv_contrib/blob/master/modules/text/samples/textdetection.cpp

I'm trying to detect text from this image:

https://i.ytimg.com/vi/WFobUoRn6Ek/maxresdefault.jpg

(Note: It is similar kinds of footage I'm trying to handle)

Result:

There're many issues, but the biggest problem to me is small letters can be easily omitted or mis-recognized.

e.g.) 'i' became l
e.g.) 'in' became 'm'
e.g.) 'l' can be gone...

I think the problem is related to the resulted image from erFilter. As this pic shows some small parts are already omitted at this point.

enter image description here
Please let me know if there's a good way to avoid such small parts omission. Possibly some sort of preprocessing to the image?

Note: I already checked following post, but the my target footage is similar to example of 'failure' case in his paper.

Community
  • 1
  • 1
Aki24x
  • 1,058
  • 1
  • 13
  • 28
  • If you have large amount of data use deep learning. – Video Analysis Deep Learning Oct 11 '16 at 09:24
  • if possible, mis-recognition can be reduced by: 1. gather all potential ambiguous things like i=l, O=0, etc. 2. in your current image, spot all those ambiguous things. 3. find from a dictionary and context: try to find the most likely version. For example it is unlikely that the word is "TRALNING" but more likely to be "TRAINING". Yes, that's a lot of work... In special tasks (like ANPR-redetection) it might be enough to just replace all ambiguous things by a single fixed thing. – Micka Oct 11 '16 at 13:04
  • You can pre-process your image like in this post : https://stackoverflow.com/questions/47601528/how-to-split-noise-and-text-from-the-image-for-preprocessing-of-ocr – sixela Dec 15 '17 at 10:16

0 Answers0