After over 10 years research, scientists at the Brain and Cognition Research Centre in Toulouse have developed a vision recognition technology based upon the human visual system.
The technology, dubbed SpikeNet, works like the human visual system in that it uses networks of millions of neurones linked together by synaptic connections. Learning can modify these connections, so that the network can recognise particular objects. The main difference between SpikeNet and conventional neural network based approaches is that the neurones send information in the form of pulses, or spikes.
SpikeNet is fast when compared to conventional artificial neural network approaches, which generally require a new activation level to be calculated for each neurone in the network, at each step in the computation.
However, the very large numbers of units required to implement anything approaching the complexity of the visual system has effectively ruled out the use of conventional artificial neural networks approaches in real-time image problems.
In contrast, the coding algorithms used in SpikeNet keep the computation required to a minimum. Where a network has hundreds of millions of units, computation only occurs when a unit actually fires a spike. Since only a few percent of the units actually emit a pulse, the speed of computation is increased dramatically.
The principal benefits of SpikeNet over more conventional recognition systems are high performance, small kernel ‘footprint’, a high level of robustness despite large variations in image contrast, luminance and the ability to cope with large amounts of noise. This has been demonstrated in applications where SpikeNet has been used to solve difficult problems, such as identification of hundreds of faces in natural surroundings, with performance being close to that achieved by human operators. SpikeNet can be run on a single desktop computer, effectively simulating networks with hundreds of thousands of neurones in real time. This provides sufficient processing power to enable recognition of small video images at up to 10 frames per second. A basic system could comprise a standard PC running Windows 98/NT, a camera, a video acquisition board, and the SpikeNet software.
Applications for the technology are varied. An early example concerned itself with car recognition.
A large automotive manufacturer suffered significant losses through the theft of cars. A number plate recognition system had already been installed to control the exit barriers in the mass car park but thieves used false number plates to defeat the system. Now, a webcam has been placed to view the cars at the barrier. An image of the front of the car is taken and SpikeNet matches the number plate to the type of car and its colour, thereby increasing security.
For object recognition, used in car identification, a learning session is undertaken to teach the various car models. Importantly, the number of types that may be taught is unlimited.
This capability makes the technology suitable for applications requiring the potential recognition of thousands of objects, such as car recognition on roads, car parks, airports, etc. Using a dynamic learning mode, if SpikeNet does not know an object, the learning process can be initiated automatically, like our own brain.
Whilst there are many potential vision recognition applications, such as identification, sorting and image analysis, the underlying technology has other applications, involving the processing of inputs from other sensors, such as audio signals.