Improve the selection tools to state-of-the-art image segmentation
I would like to discuss the actual behaviour of the selection tool. I have some image processing skills, and I am often disappointed by the result of the selection tool.
In practice, I frequently have to select with the lasso tool to get a good result, which is time-consuming.
I noticed Photoshop implemented advanced tools for this purpose. In particular, neural network, which is probably the right answer to perform state-of-the-art image segmentation.
Gimp has two tools
- one selects pixels by contiguous area
- the second selects by colour
- and two more advanced tools that work well or not
Frequently, the selection by contiguous area features unwanted areas because there is a connected pixel that feature a value which permit to flood an inner area. Then we have to try several threshold values to try to find a better result. This process is painful in Gimp.
Of course, it would be challenging to achieve the human vision and the purpose of Gimp is not to be an image segmentation tool.
But firstly, I think it could be improved by a more clever water flooding algorithm.
Also, we could apply some Mathematical Morphology algorithms to filter the shape of the selected area. We can for example filter the selection area with a circle of radius R px to remove unwanted area which cannot contains this circle. This would easily make a smoother selection area. Such features only requires to implement open/close operations.
It would be interesting to know if the G'MIC tool could work on this.