• Advertisement
Sign in to follow this  

Neural network question/problem, not game but still AI.

This topic is 4030 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

Ok, here is the problem. I am attempting to write a program to look at an image and recognize black circle left over from scanning punched pages (to tiffs) and then remove the hole. While the quick answer would seem to be neural network, the problem is in the fact it will take 14400 inputs (approx 120x120 target area allowing for sloppy holes), 14401 outputs (to create a mask to remove “bad” area and one extra so I know that if a hole was removed) and that is saying nothing about a hidden layer. My concern is with the amount of inputs and outputs being so high, will it be possible to train such a monster? I have considered scaling the target image down to make target area smaller, but the problem with that is then in scaling the mask back up will not be that good and there are times that information runs right up to the holes (or into the holes at times). Also there is memory consideration with so many nodes. The only other solution that comes to mind is using a search tree to divide the search area and then based on the areas large enough to have a hole in them use sample masks that best fit that profile to test against and then use the mask that has the best percentage of match. Any thoughts, ideas, or suggestions?

Share this post


Link to post
Share on other sites
Advertisement
Don't use neural networks for such a simple problem.

I'd first run an edge trace on the image and then use a Hough transform to find your circles. Google for it.

If your circle radius is invariant in your image set the problem becomes easier.

Share this post


Link to post
Share on other sites
Thank you for the point in the right direction Anonymous P.

I had a feeling that I was starting with the hardest method first. Unfortunately the radius does change, but the found a few papers based on your suggested search that address that along with ellipses that should get the results that I want and be simpler (relatively speaking) than what I had in mind.

Share this post


Link to post
Share on other sites
You might also consider whether a histogram analysis would improve your assessment of whether black pixels were present. If you expect your image to have significant dark content then this obviously wont work though.

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement