A vision system based upon the human visual system

SpikeNet, is a revolutionary image recognition and tracking technology using networks of asynchronous spiking neurones, based on 15 years of research by Simon Thorpe and his colleagues at the Brain and Cognition Research Centre in Toulouse.

SpikeNet uses processing algorithms that are directly inspired by the strategies used by the human visual system which outperforms even the most sophisticated machine vision systems. Indeed, the human visual system is able to analyse a complex scene in a fraction of a second.

Rank order coding

There are billions of neurons in the visual system, and each transmits information by generating electrical pulses or spikes with maximum firing rates that are usually under 100 spikes per second. The conventional view is that computation in the nervous system uses the rate at which neurons emit spikes to code information. Thorpe argued that since vision is so fast, each neuron probably only gets to fire at most one spike, ruling out conventional rate coding based on neuron discharge. Instead, he proposed that the brain could use the order in which neurons fire spikes to code information (rank order coding) and this is the basic idea behind SpikeNet.

Onset latencies varies with activation: The first spikes fired are the most significant and provide sufficient information to recognize an object. Indeed, a strong stimulus (yellow line) will generate a spike sooner than a weak stimulus (red line).

Just like the human visual system, SpikeNet sends information in the form of spikes and simulates the activity of very large networks of neurons in which these spikes are generated asynchronously. It is designed to perform high level tasks with only one spike per neuron.

Real time image processing with conventional CPU power

Although the CPU of a conventional desktop PC doesn't actually contain anything that looks even remotely like a real neuron, very large networks of spiking neurons can be simulated very efficiently with such hardware because the rank order coding algorithms keep the computation required to a minimum. A basic recognition system needs just a standard PC, a camera, a video acquisition board, and the Spikenet software. This is enough processing power to allow real-time processing of small video images at up to 10 frames per seconds.

SpikeNet is also suitable for implementation on parallel hardware in the case of more demanding problems with large images and high number of target forms.

Robust recognition

Order coding is contrast independent and SpikeNet Technology provides great tolerance towards noise and light conditions.

SpikeNet's recognition mechanism can also tolerate variations in rotation and zoom. By default, the system handles variations in rotation and size of a few percent, even with one single prototype for each target. To increase this level of tolerance, SpikeNet can automatically generate multiple models to cover a wide range of zoom factors and/or orientations.

A simple learning process

You can train the system on the individual targets you wish to locate. For each target you create models which are saved in a database. You may load (and learn) images of different sizes and add any number of models. SpikeNet easily handles large model databases.The computation time increases linearly with the number of models.

Versatile

SpikeNet Technology works for any kind of image: video images, finger prints, satellite image, x-rays, microscopic slides. It can be applied to single video processing as well as to multiple video stream processing.