Ask Your Question

Background subtraction from single static image

asked 2017-09-23 03:01:26 -0500

Martian gravatar image

Are there methods that attempt to differentiate foreground and background in a single static image?

Standard OpenCV background subtraction works by correlating multiple images, but in this case, I have only a single image. I do realize that the methodology would be completely different, but I'm not sure what other terms to use other than foreground and background. (So appropriate keywords for a search would be great)

The background image in this case is fairly monotonous. Think 'wall with some noise.' The value (luminance) of some of the 'foreground' objects can be relatively close to the background, but I'm guessing that some relatively sophisticated algorithms have been developed to deal with this.

edit retag flag offensive close merge delete


keywords: image segmentation

(opencv has grabcut, watershed for this. also, neural networks can be trained for this purpose)

berak gravatar imageberak ( 2017-09-23 03:34:12 -0500 )edit

Thanks, berak. Do you know of any examples of neural nets being used for similar apps? I've been trying to figure out how to use a convo net to simply identify the foreground images, which would solve the problem. But there are too many variations in foreground objects to do that directly.

If you know of an approach to identifying background pixels with a neural net, I'd love to hear more about it.

Martian gravatar imageMartian ( 2017-09-23 11:55:43 -0500 )edit

this post may help you

LBerger gravatar imageLBerger ( 2017-09-24 11:11:20 -0500 )edit

1 answer

Sort by ยป oldest newest most voted

answered 2017-09-23 11:35:18 -0500

Tetragramm gravatar image

Berak's suggestions are good, but one other idea is kmeans clustering. The large cluster is the background, and everything else is foreground. Then morphological operations to clean up the map.

I don't know how effective it would be, but it has the advantage of being easy to try.

edit flag offensive delete link more


Thanks for the idea. I'll take another look at kmeans.

Martian gravatar imageMartian ( 2017-09-23 11:57:11 -0500 )edit

Question Tools



Asked: 2017-09-23 03:01:26 -0500

Seen: 4,309 times

Last updated: Sep 23 '17