A Model Of Visual Recognition Implemented Using Neural Networks

The ability to recognise and classify objects in the environment is an important property of biological vision. It is highly desirable that artificial vision systems also have this ability. This thesis documents research into the use of artificial neural networks to implement a prototype model of visual object recognition. The prototype model, describing a computtional architecture, is derived from relevant physiological and psychological data, and attempts to resolve the use of structural decomposition and invariant feature detection. To validate the research a partial implementation of the model has been constructed using multiple neural networks. A linear feed-forward network performs pre-procesing after being trained to approximate a conventional statistical data compression algorithm. The output of this pre-processing forms a feature vector that is categorised using an Adaptive Resonance Theory network capable of recognising arbitrary analog patterns. The implementation has been applied to the task of recognising static images of human faces. Experimental results show that the implementation is able to achieve a 100% successful recognition rate with performance that degrades gracefully. The implentation is robust against facial changes minor occlusions and it is flexible enough to categorise data from any domain.

Grantor dc:publisher

Edith Cowan University, Research Online, Perth, Western Australia

Year dc:date

1994

Author dc:creator

• Phillips, Vincent C.

• Pattern recognition systems
• Neural networks
• Computer science
• Artificial Intelligence and Robotics
• Computer Sciences

Repository record dc:identifier

https://ro.ecu.edu.au/theses/1472

OAI identifier oai:identifier

oai:ro.ecu.edu.au:theses-2473