In the process of segmentation, pixels are assigned to regions based on features that distinguishes them from the rest of the image. Value Similarity and Spatial Proximity, for example, are two important principles that assume that points in the same region will have pixels that are spatially close and have similar values.

In lots of situations this is true, but what about regions composed of pixels that are not similar in value? Consider the image below. The same "logical" region is composed of different elements that together represent something meaningful. In the same region there are trees with different sizes and shapes, with shadow over some of them etc. There are different things, with pixels that differ a lot in value, but I still need to group them together in the same region. From the image you can see that I don't care so much with differences in color. In this case texture is the most important attribute. What algorithms are used to do the segmentation and classification in problems like that?

I'm already looking for some algorithms and techniques that focus on texture, but some opinions from the experts will help me a lot. I think I need some orientation. Thanks!


  • $\begingroup$ Before I give some suggestions, I need some clarifications. Do you want the earth pixels to be classified as being the same crop as the surrounding crop? How different do images of different aerial crops look? $\endgroup$ – JahKnows May 29 '17 at 23:08
  • $\begingroup$ @JahKnows I've made some modifications in my question. I think it is clearer now. $\endgroup$ – user7369 May 31 '17 at 17:13
  • $\begingroup$ Have you simply googled "texture segmentation"? The results should keep you busy for some time. $\endgroup$ – Rethunk Jun 2 '17 at 12:39
  • $\begingroup$ @Rethunk As I said above: I'm already looking for some algorithms and techniques that focus on texture, but some opinions from the experts will help me a lot. $\endgroup$ – user7369 Jun 2 '17 at 13:36
  • $\begingroup$ @rrd You haven't made the problem specific enough. As your question stands, it's not easily distinguishable from a request for SO users to help you with the most basic information. Unless you mention some specific texture segmentation technique (e.g. Law's techniques), it doesn't look like you've made an effort. Also, it's not in the spirit of SO to provide general guidance; make your question as specific as possible. In this case, start with the names of specific texture segmentation methods and ask specific questions about how to compare them. $\endgroup$ – Rethunk Jun 4 '17 at 13:00

If you have enough data, you can train a segmentation net example for a specific group of data (i.e. trees). Single DNN with multiple output branches can easily segment & classify the data that you are searching for.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy