While investigating the snake results it became obvious that in many cases when pollen was being analysed that the snake was picking up a faint double edge (perhaps due to poor focussing) and was not then able to find the ``real'' double edge. Thus a lower percentage of double edge was being reported than was actually present.
In this work the Paradise network has been used in its general training form where no pre-selection of ``good'' examples of each class is performed. However, it has been shown in [ 11 ] that the classification ability of the Paradise network can be improved and the number of classes generated decreased if good examples of the objects to be recognised are presented to the network first.
There are many thousands of types of pollen, so a practical system would need to be trainable on-line by a palynologist.
Mr I France