Key-words:
Cellular Neural Network, Smart Sensors, Genetic Algorithm, Texture Analysis, Deconvolution
Abstract:
We have developed a new single-chip texture
classifier smart-sensor system. Its main part is a programmable
Cellular Neural Network (CNN) VLSI chip with optical input and
an execution time of a few microseconds. This chip contains a
dynamic and locally interconnected 2-D cell array. It executes
a theoretically new method for texture classification, compared
to the other convolution-based feature extraction methods, since
here we have feedback convolution as well. Depending on the kernel-parameters,
this array can execute filtering, moving, linear and nonlinear
effects at the same time. The parameters of the feedback and feed-forward
convolutions are optimized through a genetic learning using the
22*20 CNN chip itself. This chip has a simplified architecture
with binary input/output, but it gives good recognition results:
this CNN chip with a simple 3*3 kernel can reliably classify 5
natural Brodatz textures. Using more templates for decision-making, more
textures can be separated and a classification-error of less than
1% has been achieved.