Smart segmentation super pixels module

This module allows to classify super pixels on a grey level or color image.
To achieve that, the user can first define the super pixels parameters.
Then he can draw on the image, in order to manually classify some of the super pixels.
A random forest model will then be trained using those super pixels.
Once the model is trained, it can be applied on the complete image, or on similar images.

Here we have the graphical interface on this module :

Class selection

The class selection box contains all the information concerning the classes of the model.
The user can add or delete a class, and can change the color used to visualize a class.
If there is 2 classes on the model, the output image will be binary, otherwise it will be a labeled image.
It is possible to change the selected class by clicking on it.
Once a class is selected, the user can add super pixels to this class by drawing on the image.
If the eraser button is activated, drawing on the image will remove super pixels from the selected class.
The number on the right of each class indicates the number of super pixels contained by this class.

Super pixels parameters

In order to adjust the super pixels grid, the user can change differents parameters.
Size : The mean size of the super pixels.
Compactness : Define how much the super pixels will fit the contours of the objects. With a small value they fit better, with a high value they are more regular.
Size ratio : The minimal size a super pixel, as a fraction of their supposed size.

Features

Then the user can choose the features for the model.
Those features are measurments of the super pixels, and are used in the random forest model to classify them.
There are four categories : geometry, basic intensity, texture and histogram.
Concerning the histogram, each class of the histogram will count as a feature, the it can impact the model much more than the other features.
The list of features box shows all the features used in the model.
Once the model is computed, each features will have an associated percentage, which shows how much a feature impacts the model.
The functionality "Remove features below" allows the user to remove the less meaningful features.
To do so, the user choose a threshold in percentage, and click on the trash button.
The features selection widget will be updated automatically.

Model parameters

On the main window, the user can also change the number of trees used in the random forest model.
With more trees, the model will be more precise, but the computation time will be longer.
It is also possible to activate or deactivate the 'live update' mode.
If this mode is activated, the model will be actualized every time a pixel is added to a class, or removed from a class.
If this mode is deactivated, the user can actualize the model manually.

Example

Here we have an example of a model with 3 classes. The features setting used to compute this model is the one presented on the previous image.

Once the user is satisfied with his model, he can save it and use it in the Explorer interface.