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. |
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 |