Segmentation_2_component_image_514_V2


Aim:
Manual or automatic segmentation of 2-component image, via the feature space (scatterplot).

History:
June 2005: First developments (in Java) by Laetitia Pasquet and No�l Bonnet.
June 2006: Version 2 by Cedric Gillet and No�l Bonnet


Method:
Load the 2 images:


1) Building the scatterplot:


            
                                                                         Scatterplot: linear scale                      logarithmic scale

Only pixels of both images with a grey level value between the minimum and the maximum threshold are considered.
In addition, a coefficient of reduction can be applied. If the coefficient equals 2, the size of the scatterplot is 128 x 128, for instance.
Two scatterplots are produced, one with a linear scale and the other one with a logarithmic scale. Any one can be used for the others steps of the procedures.

2) Choice of the automatic or interactive mode of segmentation
    - Manual (interactive) mode:





.......

After the number of classes (N) has beeen selected, the user has to draw N regions of interest on the scatterplot. Each time a ROI is drawn, pressing "Enter" allows to store the ROI and select a grey level value corresponding to that ROI (204 in the exemple above).
Whan all the ROIs have been drawn, a composite figure (with the different classes of pixels) is built in the feature space (scatterplot) and in the image space (see below). Black pixels are pixels that were not included in any ROI.

                                       
 Partitionned scatterplot                                  Partitionned (segmented) image space


  - Automatic mode:

                                 

Note that the number of classes is irrelevant            First phase: look at the variation of the number of modes (= classes)
in the automatic mode                                             as a function of the smoothing parameter (sigma)




The curve plots the number of modes as a function of the smoothing parameter. A plateau is generally expected for the "true" number of classes. Here, we can see a plateau for 4 classes for sigma between 7 and 15. Note that several plateaus can sometimes be obtained, reflecting the intrinsic hierarchical nature of clustering.

This first phase can be repeated several times, when necessary.


  
             
Second phase: fix a value of sigma                                                                        Estimated pdf
 for the final pdf estimation (here, sigma=8)                                                                                                                                                                    

            
         Automatically partitionned pdf          Automatically partitionned (segmented) image space

Version 2:
In the manual or the automatic mode, several coefficients of colocalisation are computed (for the whole image, for the computed scatterplot, and for the selected or automatically computed regions.