The Vertebrate Eye

The vertebrate eye is built like a camera - or rather we should say a camera is modeled after the vertebrate eye. It contains a light-tight chamber with a lense system in front that focuses an image of the visual field on a light-sensitive surface (the retina) in back.

The Retina

The retina is the sensory tissue responsible for the transduction of light energy into nervous impulses. The retina is actually made up of three sets of neurons (nerve cells) in series with each other: photoreceptors (rods and cones), intermediate neurons, and ganglionic neurons whose axons form the optic nerve. When a quantum of light strikes a rod it causes a chemical process which triggers the discharge of a nerve impulse in the receptor cell. The optic nerve leads the nervous impulses from the retina to the brain.

The Neuron

The neuron is a cell body with all its processes. From the nucleated cell body extends an axon, which carries impulses away from the cell. Typically several branching dentrites extend from the cell body. These carry impulses towards the cell body.

Synapses are the junction points between neurons. The membranes of synapses are separated by a narrow gap (synaptic cleft). As the impulses aproaches the presynaptic membrane, it causes the release of a neurochemical transmitter which rapidly diffuses across the confines of the synaptic cleft and approaches the postsynaptic membrane, where it causes a change in its ionic permeability. The resulting flow of ions creates a brief electric potential difference which can be recorded by a microelectrode.

The ISE-Ultramicroelectrode

The most vulnerable components of any electrochemical system are the electrodes - the interface between the real world and the electronic circuitry which controls the system. The Ion-Selective Ultramicroelectrode senses electronic potential differences at the neural synapses and transmitts the signal to the data processor. The Ultramicroelectrode has an architecture similar to a microchip. It consists of a thin layer of iridium film that was deposited onto a polymer layer by evaporating of the electrode material under vacuum.

The film is sandwiched between thin polymer supports by gluing a second sheet of polymer over the exposed surface of the electrode material.At one end of the electrode the polymer is cut with a laser into an array of singular electrode units which act as the measurement ends, while at the other end of the electrode an electrical contact is exposed. In order to assure the contact between neuronal synapse and the electrode unit the surface of the measurement end of the electrode is coated with a film containing proteinmolecules of the neural membrane.

The Neural Accelerator Chip (NAC)

Neural networks are computer systems that emulate the interconnections and behaviour of neurons in the brain. They are trained from examples, reinforcing the connectivity between neurons. Applications include cognitive modelling, control systems, knowledge-based systems, associative memory, system optimization and pattern recognition. The hardware is based on the Neural Accelerator Chip (NAC).

The NAC is a modular multi-processing architecture consisting of sixteen processing elements and implemented in a 1.5u double metal, single polysilicon CMOS Process. The device contains approximately 200,000 transistors in an area of 74 sq mm, and calculates pipe-lined sums of products at the rate of 500 million operations per second operating with a 16 Mhz clocking frequency.

PIGVISION uses a configuration which contains an array of 32 NAC's (4x8) which perform sixteen 8x8 convolution masks on the electrode data stream.

High Speed Image Classification

The natural unit for visual information processing, in machine as in animal is the "movie" - not the individual picture. In fact, a human observer artificially constrained to look at an isolated static image, nonetheless, rapidly decomposes the picture into a time sequential montage of subimages - thus generating a movie.

Image computation connotes complex operations which modify individual pixel values of one or more images. PIGVISION uses the following image computations to process the data stream from the ISE-Ultramicroelectrode for the output device (Computer Screen):

Intensity transformation

serial pixel modification

logical filtering

edge detection

domain transformation (Fourier, Walsh-Hadamard, Karhounen-Loeve)

statistical operations ( contigency tables, co-occurence matrices, auto-correlation, cross-correlation)