Neural networks for object recognition
Nicola Cross, Roland Wilson · Warwick Research Archive Portal (University of Warwick) · 1995
The overall aim is to produce a vision system using artificial neural networks that is capable of working with whole images. To this end some major problems in Computer Vision are considered. These are questions of control, binding, multiple pattern recognition, invariant processing and the modelling of complex visual behaviours. The majority of the practical work presented here is concerned with the development of higher order neural networks. In the higher order neural network, preprocessing creates a nonlinear decision surface for the network. A single layer will then often be sufficient for the network to train in a small number of passes. In the more traditional multilayered neural network, the hidden layers must extract the nonlinear quantities in the data for themselves. The number of passes through the data that must be made by the multilayer net for nonlinear problems is thus typically much greater. In addition, claims are made by some workers for superior robustne...