/[gzz]/manuscripts/Paper/paper.tex
ViewVC logotype

Diff of /manuscripts/Paper/paper.tex

Parent Directory Parent Directory | Revision Log Revision Log | View Patch Patch

revision 1.113 by tjl, Mon Mar 31 16:39:15 2003 UTC revision 1.114 by tjl, Mon Mar 31 16:46:18 2003 UTC
# Line 638  the pixels. Line 638  the pixels.
638    
639  From the  From the
640  model  model
641  we can formulate the following points:  we can see that
642  To be distinguishable, \label{secfeaturevec}  to be distinguishable, \label{secfeaturevec}
643  \begin{itemize}  a feature vector for a given texture should always be the
644  \item A feature vector for a given texture should always be the  same.
     same.  
645      Fragments of      Fragments of
646      a non-repeating texture will be slightly different,      a non-repeating texture will be slightly different,
647      resulting in slightly different vectors even      resulting in slightly different vectors even
648      if the local structure is the same.      if the local structure is the same.
649      A repeating texture should thus be easier to recognize.      A repeating texture should thus be easier to recognize.
650      Our anecdotal observations confirm this.      Our anecdotal observations confirm this.
651  % XXX: Why wouldn't it always be the same?  
652  % - seeing different parts of the texture?  Additionally,
653  % - ambiguous perception?  the entropy of the feature vectors
654  \item The entropy of the feature vectors  over the distribution of textures should be maximized.
655      over the distribution of textures should be maximized.  \iftechreport
656  \end{itemize}      This means essentially
657        that if all square-like shapes were green, we would be
658        wasting recognitive power.
659        There should also be as many possible features in the distribution
660        as possible. For example, if there were no yellow textures,
661        or if there were no curved lines, we would be wasting
662        recognition potential by leaving some elements
663        of the feature vector always zero.
664    
665        Generally, the entropy is maximized when the features are distributed
666        independently from each other.
667    
668    \else
669        The distribution should contain occurrences
670        as many different features
671        as possible, and
672        the features should be distributed
673        independently from each other.
674    \fi
675    However, the results cited above\cite{olson02vstm}  also
676    indicate that in any {\em single} texture, only a limited range
677    of features should be used.
678    
679  % which facilitates recognition and memorization of images.  % which facilitates recognition and memorization of images.
680  % The structure of the features is assumed to be irrelevant.  % The structure of the features is assumed to be irrelevant.
# Line 672  To be distinguishable, \label{secfeature Line 692  To be distinguishable, \label{secfeature
692  % To achieve this, the model should maximize the entropy of the feature vector.  % To achieve this, the model should maximize the entropy of the feature vector.
693  %We call this the principle of saving bits.  %We call this the principle of saving bits.
694    
 \iftechreport  
 The last part means essentially  
 that if all square-like shapes were green, we would be \fi  
 wasting recognitive power.  
 There should also be as many possible features in the distribution  
     as possible. \iftechreport For example, if there were no yellow textures,  
     or if there were no curved lines, we would be wasting  
     recognition potential by leaving some elements  
     of the feature vector always zero. \fi  
     However, the results cited above\cite{olson02vstm}  also  
     indicate that in any {\em single} texture, only a limited range  
     of features should be used.  
   
695    
696  Generally, the entropy is maximized when the features are distributed  In a sense, the model of perception should be {\em inverted}
697  independently from each other:  in order to produce a unique background from
698  features orthogonal to human perception  a random vector.
699  (e.g.,~color, direction of fastest luminance change)  Features that are orthogonal for human perception
700    (e.g.,~color and  direction of fastest luminance change)
701  should be independently random, and features not orthogonal  should be independently random, and features not orthogonal
702  (e.g. colors of neighbouring pixels)  (e.g. colors of neighbouring pixels)
703  should be correlated so as to maximize the entropy.  should be correlated so as to maximize the entropy.
704  For example, pixels on a small area should correlate enough to  %For example, pixels on a small area should correlate enough to
705  facilitate perception of contours.  %facilitate perception of contours.
706    
707  In a sense, the model of perception should be {\em inverted}  A type of inversion approach has been used
 in order to produce a unique background from  
 a random vector.  
 This type of approach has been used  
708  by Ware and Knight\cite{ware95texture}, for inverting  by Ware and Knight\cite{ware95texture}, for inverting
709  the earliest stage of the visual system,  the earliest stage of the visual system,
710  the spatial frequency detectors, in order to place  the spatial frequency detectors, in order to place

Legend:
Removed from v.1.113  
changed lines
  Added in v.1.114

savannah-hackers-public@gnu.org
ViewVC Help
Powered by ViewVC 1.1.26