Structural Decomposition

 

[pdf]                               80-citation article

[talk]                               slides about the categorization system

[link]                                       explaining parallel pop-out of visual search

 

A structure is decomposed with two principally different methods, the generation of a local/global space for each contour, and the generation of the symmetric axes for image regions. The decomposition can also explain most pop-out phenomena of human visual search.

 

Local/Global Space: For each contour, a window is iterated through the contour, which classifies whether a selected segment is a ‘bow’ or an ‘inflexion’, creating thereby signatures for a given window size. This is carried out for different window sizes to generate the local/global space, which contains a wealth of structural information. Here the one for a wiggly arc:

 

LG_arc_wig

 

 

Symmetric Axes: I use a wave-propagation process to generate the symmetric axes:

 

Fig_2

 

 

This shows the full decomposition output for one image at different scales:

 

s4_1_p4_o132002_cs

 

 

The decomposition returns many parameters. The challenge is now to create a useful multi-dimensional space, with which one can perform perfect categorization for arbitrary images. I have already worked toward that direction by classifying the images of the COREL and Caltech 101 collection:

 

            [link to image classification]          shows results from image classification (categorization), searches and sorting.

[link to COREL categorization]     basic-level categories sorted according to percentage correct

[link to COREL category labels]     how we categorized the COREL image classes