Logo Text Detection Accuracy

asked 2014-03-13 13:17:04 -0500

Rolo gravatar image

updated 2015-10-08 10:19:28 -0500

Hey everyone,

I'm attempting to identify store logos using a haar cascade classifier I trained. It was trained using 60 positive images and ~600 negative images.

My question comes about because the images in which I am attempting to find the textual logo, there also exists other text.

OpenCV is determining this "other" text to in fact be the found trained logo. Sample image sizes were 90x30. Sample image used in detection is 300px wide by whatever length is needed to maintain aspect ratio.

So I'm wondering a few things:

  1. Is OpenCV distorting the text in my positive samples to the extent that other text is being detected?
  2. Should I be fine with the number of samples listed above? I followed the tutorial here, though I have heard from other tutorials that 1000's of samples is needed?
  3. Are my sample image sizes okay? I've read a few things about opencv preferring 25px by 25px ?
  4. When training the classifier, I limited it to 5 stages to save time. How much more accurate would the classifier be with say 20 stages? Or is the accuracy only relative to the number of positive samples?

Thanks so much for any input you can provide on this. I'm really looking forward to generating this custom haar file and having it work well.

Very much value your comments :)

Please let me know if I'm missing something trivial :)

Thanks

edit retag flag offensive close merge delete

Comments

Can anyone please clarify the above for me?

Rolo gravatar imageRolo ( 2014-03-15 09:51:04 -0500 )edit

Can you upload some example image of the logo you are trying to detect and the "other" text that is being detected?

Also, what exactly were the parameters you used in Haar training? Can you write the exact command?

GilLevi gravatar imageGilLevi ( 2014-03-16 17:12:43 -0500 )edit