These experiments were designed to assess the feasibility of two different approaches to the recognition of pollen grains. Whilst the model based approach requires considerable further work, the Paradise neural network clearly demonstrates the potential for a fully automated pollen analysis system.
Whilst the model based approach has a very good recognition rate for the pollen grains it also misclassifies large amounts of non-pollen as pollen. In order to be a useful classifier it needs refining and another feature of the objects must be included in the discrimination. The more refined model would almost certainly be specific to a particular type of pollen, even the general model used in this paper will only recognise some types of pollen.
In contrast, the Paradise network automatically classifies similar objects together. Without any a priori knowledge (apart from choosing the feature detectors) the network learns to separate the pollen from the non-pollen with a high degree of accuracy. If other types of pollen need to be recognised, the network can be trained with examples of this pollen and will learn to recognise them.
Mr I France