Primitive attempt to turn images into percepts

Abstract

Images are about pictures. Percepts are about information. People need images. Machines do not. This paper suggests that it may be possible for machines to perceive things, not just register images in real time. This goal inspired our search for methods that might work in real time to perceive their world. In our brains, the image is processed separately in different parts of the brain, and the results of those parallel operations are somehow fused to form the percept. We have done extensive development of a spectral segmentation of images using “Artificial Color” based on the way animals use spectral information. That research showed that there is a massively parallel approach to recognize targets or their background using Fourier texture discrimination. In this paper, a Fourier-based system that uses nonlinear discrimination to recognize shape, size, pose, and location of a target in a scene is discussed. The sought-after conversion of that information into an Artificial Percept in which a “cartoon” of the situation is formed with labeled targets giving their appearance (pose and size) and location. This is a primitive percept. The machine no longer has an image. Instead it “knows” what targets of interest are in the field, where they are, and their range and pose.

Appendix: Margin Setting Algorithm

Sample feature vectors $A = \{ \vec{x}_{1} , \ldots ,\vec{x}_{n} \} \subset [0,\;1]^{q} \subset R^{q}$ taking from the target class and the sample vectors $B_{l} = \{ \vec{y}_{l,1} , \ldots ,\vec{y}_{{l,m_{l} }} \} \subset [0,\;1]^{q} \subset R^{q}$ ($l = 1, \ldots ,k$) taking from the other $k$ objects. Denote $B = \bigcup\limits_{l = 1}^{k} {B_{l} }$.

1.1 Notation:

Rand ($F$): random numbers taking from distribution $F$.

Unif ($D$): uniform distribution function on set $D$.

Card ($D$): the figure of merit of set $D$.

$[a,\;b]^{q} \;$: a $q$-dimensional cube.

$O(\vec{z},\,R)$: a $q$-dimensional ball centered at $\vec{z}$ with radius $R$.

$I$: designed number of generations.

$\varepsilon$: designed error tolerance.

$s$: designed number of samples taken for each generation.

$\delta$: designed size of perturbation.

$L$: designed number of mutations for each generation.

1.2 Algorithm:

• 1. Compute $N \equiv$ Card($A$), and set $i = 0$.

• 2. $i: = i + 1$. If $i > I$ or Card($A$)/$N$<$\varepsilon$, output $\{ (\vec{c}_{j} ,\,r_{j} ;\,N_{j} ):\;j = 1, \ldots ,i - 1\}$, and stop!

• 3. Take $\vec{c}_{i,1} , \ldots ,\vec{c}_{i,s}$ from Rand(Unif($[0,\;1]^{q}$)) and set $t_{i} = 0$. For $j = 1, \ldots ,s$, compute $r_{i,j} = \mathop {\hbox{min} }\limits_{{\vec{y} \in B}} |\vec{y} - \vec{c}_{i,j} |$. If $N_{i,j}$ $\equiv$ Card($A \cap O(\vec{c}_{i, j} ,\,r_{i, j} )$ ) = 0, discard ball $O(\vec{c}_{i,j} ,\,r_{i,j} )$; otherwise, $t_{i} : = t_{i} + 1$, and record $(\vec{c}_{i,j} ,r_{i,j} ;N_{i,j} )$.

• 4. If $t_{i} = 0$, record $(\vec{c}_{i} ,\,r_{i} ;\,N_{i} ) = (\vec{0},\,0;\,0)$, go to step 2; otherwise, for $\{ (\vec{c}_{i,j} ,\,r_{i,j} ;\,N_{i,j} ):\;j = 1,\; \ldots ,t_{i} \}$, compute

$$ w_{i,j} = \frac{{N_{i,j} }}{{\sum\nolimits_{j = 1}^{{t_{i} }} {N_{i,j} } }},\quad j = 1,\,2, \ldots ,t_{i} . $$

• 5. Take $v_{i}$ from Rand(Unif($[0,\;1]$)), if $v_{i} \in [\sum\nolimits_{j = 1}^{{\ell_{i,0} - 1}} {w_{i,j} ,\;\sum\nolimits_{j = 1}^{{\ell_{i,0} }} {w_{i,j} } } ]$ for some $\ell_{i,0} \in \{ 1,\,2, \cdots ,t_{i} \}$, Set $N_{i} = N_{{i,\ell_{i,0} }}$ and the corresponding ball as $O(\vec{c}_{i} ,\,r_{i} )$; and discard all the other balls. Set $l = 1$.

• 6. Take $\vec{s}_{l}$ from Rand(Unif($[ - \delta ,\;\delta ]^{q}$)), mutate $\vec{c}_{i}$ to $\vec{c}_{i}^{1} = \vec{c}_{i} + \vec{s}_{l}$, and compute $r_{i,j} = \mathop {\hbox{min} }\limits_{{\vec{y} \in B}} |\vec{y} - \vec{c}_{i,j} |$. If Card($A \cap O(\vec{c}_{i}^{1} ,\,r_{i}^{1} )$)$\le N_{i}$, record $(\vec{c}_{i} ,\,r_{i} ;N_{i} )$ and set $A: = A\backslash O(c^{i} ,R^{i} )$, go to step 2; otherwise, $N_{i} : =$Card($A \cap O(\vec{c}_{i}^{1} ,\,r_{i}^{1} )$), $\vec{c}_{i} : = \vec{c}_{i}^{1}$, $r_{i} : = r_{i}^{1}$ and $l: = l + 1$. If $l > L$, record $(\vec{c}_{i} ,\,r_{i} ;N_{i} )$ and set $A: = A\backslash O(c^{i} ,R^{i} )$, go to step 2; otherwise, go to step 6.