IM_DBR

Image Database Retrieval GUI

     [RANK,TARG,OUTL] = IM_DBR(DBASE,FSETS,CLASSF,COMB)

Input
DBASE	Dataset or datafile with N object images
FSETS	Cell array with maximum 4 feature sets
CLASSF	Cell array with untrained classifiers (Default: KNNC([],1))
COMB	Combining classifier (Default: MEANC)

Output
RANK	Index array ranking the N object images
TARG	Index array pointing to user defined target images
OUTL	Index array pointing to user defined outlier images

Description

This command generates a Graphical User Interface (GUI) enabling the user  to label a database of images in 'target' and 'outlier' images in an  interactive and iterative way. Up to four feature sets can be given and  corresponding classifiers that assist the user by predict an object ranking  based on classification confidences for the 'target' class.

The GUI shows the top-10 of the ranking and the user should classify  them as targets or outliers (original object labels in DBASE are  neglected). There are buttons for browsing through the ranked database  or through the selected targets and outliers. Classifiers can be trained  according to two different strategies using the top right buttons
Classify - uses all stored target and outlier objects (shown in the top  left windows) for building a training set as well as the  hand labeled images in the present screen.

Label	uses just the hand labeled images in the present screen  and neglects the stored targets and outliers. This enables  a more flexible, but still controlled browsing throug the  database.
Reset	Resets the entire procedure by deleting all selected targets  and outliers.
Quit	Deletes the GUI and returns the ranking and selected targets  and outliers to the user.
A	few additional buttons and sliders for controlling the system behavior:
-	Delete and move buttons for the selected targets and outliers
-	Weights for the feature sets. For each feature set a different  classifier is computed generating target confidences for all images.  This influences the operation of the combiniong classifier.  The weights can be changed by a slider for every feature set.  By default weights are 1.
-	Two buttons for setting all labels as target ('All target') or outlier  ('All outlier').
-	Labels for the individual images can be changed by a mouse-click in the  image or on the image check-box.
-	For all images a target confidence is computed. Depending on the 'all'  and 'unlabeled' radio buttons at the bottom the ranking of all images  or of the yet unlabeled images are shown.
Note:	It is not an error, but for most classifiers useless or
counterproductive	to label an object as target as well as outlier.

Example(s)

 % This example assumes that the Kimia images are available as datafile
 % and that the DipImage image processing package is available.
 prwaitbar on
 a = kimia_images;
 x = im_moments(a,'hu');
 x = setname(x,'Hu moments');
 y = im_measure(a,a,{'size','perimeter','ccbendingenergy'});
 y = setname(y,'Shape features');
 [R,T,L] = im_dbr(a,{x,y});  % do your own search
 delfigs
 figure(1); show(a(R,:)); % show ranking
 figure(2); show(a(T,:)); % show targets
 figure(3); show(a(L,:)); % show outliers
 showfigs

See also

datasets, datafiles, mappings, knnc, meanc,