image segmentation
in [Matlab]
|
From: Li on 16 Jun 2010 12:28 Hello, I need to use the "watershed" function to do image segmentation. The output of this function is a matrix labeling different regions. It is said that: The elements labeled 0 do not belong to a unique watershed region. These are called watershed pixels. My question is: how can I get rid of these watershed pixels? I want these pixels to be segmented to one of their neighboring regions, so that every pixel of the input image belongs to one segment. Thanks!
From: ImageAnalyst on 16 Jun 2010 13:05 Those are the black lines that define the boundaries between the regions. They do not belong to either region. Think about it. It's not like you have an image of objects with different intensities and you can say that this pixel belongs to that intensity blob and that pixel belongs to that other intensity blob. With watershed, you use it when you have big blobs that are basically all the same intensity and you need to cut them apart (separate them). That line of separation is not really considered to be in either one or the other blob. If you WERE to say that they belong to one of the two blobs, which one would you say should take it? See - it doesn't make sense. Chances are you can do what you want to do without assigning the dividing line pixels to any particular blob.
From: Li on 16 Jun 2010 13:39 Thank you so much for your reply! I agree for those boundaries, I cannot say which blobs they belong to. But as results of image segmentation, there are so many pixels left without any labels by the "watershed" function. I want to process each of these patches one by one, so that the entire image is covered. No matter which blobs those boundaries are assigned to, I am fine as long as every pixel has a label. Do you have any suggestions in dealing with this problem? Thanks a lot!
From: ImageAnalyst on 16 Jun 2010 14:13 My suggestion is to ignore them. WHY does every pixel need to have a label? This is usually not the case. In every case I can think of, you threshold an image and then call bwlabel(). Thresholding means that there will be foreground and background. Only the foreground pixels are labeled. The background pixels are never labeled - their value is zero and they are ignored by regionprops(). I see no need for you to try to include these. Even if you did, it would be arbitrary, so what's the point? Any area fraction, or whatever, that you calculate will be negligibly affected by whether or not you include these as I'm sure there is more variability inherent in your sample images than you could ever hope to recover by including these boundary pixels.
From: Li on 16 Jun 2010 15:57
Thanks! Okay, I will think about your suggestions... I also tried other image segmentation methods, which have every pixel labeled. In some cases, it is not a foreground vs background problem. It specifies which patch a pixel belongs to. Also, it seems there is nothing I can do to control the number of segments. Thank you. |