Applications of Emergent Phenomena - Systematic Design for Emergence in Cellular Nonlinear Networks - page 132

one in Fig. 8.1, the upper left corner) loaded as an initial state, and after several
iterations an output image is generated where rectangular boxes are defined indi-
cating the pixels to be cropped for each isolated character.
A simple feature extractor is then proposed and applied to the segmented image
to generate a sequence of seven normalized feature vectors for each character.
A classifier can then learn to recognize the features providing a list of codes at the
output of the proposed smart-sensor. A brief description of the sensor architecture
and functionality follows, for more details the reader is directed to [88].
8.2.2 Architecture and Functionality of the CA-Based Sensor
The architecture is centered around a 2s5 Cellular Automata (CA) array with sim-
ple semitotalistic cells (Fig. 8.2). Since only two genes are used a simple circuit is
associated to each cell. A controller selects one of the two possible cell functions
and runs the CA until it contains only empty rectangular boxes, each framing an
isolated character from the visual field. Then using a simple “reading” algorithm
described next, the controller identifies and stores a list with positions and sizes
for each box (as seen in Fig. 8.1).
Fig. 8.2. The CA-based smart-sensor architecture: Each cell is semitotalistic and linearly
separable, i.e. its output at the next clock cycle depends only on the simple sum of the
neighboring cells (2-4) plus a weighted value of the central cell (or its negate). If the sum is
greater than 0 the new state is 1, otherwise is 0. A simple controller provides an efficient
way to extract features from each block and recognize the corresponding character provid-
ing at the output a list of codes abstracting the information from the visual field