[GRASS-user] Image segmentation to separate object from background

[Moritz Lennert]

On 26/06/17 15:01, Johannes Radinger wrote:
> Hi GRASS users,
>
> has anyone of you tried to use GRASS image tools (e.g. segmentation etc)
> to identify an object in a picture. For example I have multiple photos
> of fish and would like to separate the fish from its background in an
> automatized way. The images look like:
>
> http://fishbase.org/photos/PicturesSummary.php?ID=4730&what=species
> <http://fishbase.org/photos/PicturesSummary.php?ID=4730&what=species>
> http://fishbase.org/photos/PicturesSummary.php?StartRow=0&ID=4662&what=species&TotRec=5
> <http://fishbase.org/photos/PicturesSummary.php?StartRow=0&ID=4662&what=species&TotRec=5>
>
> So what I already know apriori: I would like two final classes (fish and
> background), the fish is more or less centred within each picture, and
> there is only one fish in each picture. I played already around with
> i.segment but the results are not yet satisfying.

You will certainly have to go through a classification phase: 
oversegment the image and then classify the resulting segments as fish 
or non-fish. Probably results will be better if you classify into more 
non-fish classes which you can then merge at the end.

Currently, we don't have semantic segmentation in GRASS GIS, i.e. 
segmentation that "recognizes" real objects and it is difficult to get 
complete real objects just through the segmentation methods we have 
(check i.segment, i.superpixels.slic, but also edge detection, such as 
i.zc).

For real semantic object recognition, neural network approaches seem the 
way to go today, so you might want to look into those.

Moritz