This paper gives the results of using both the methods to differentiate between pollen and non-pollen. In this experiment the pollen species has the characteristic ``double edge'' mentioned in section 2.1 . Two sets of objects were collected and in order to fully test both methods the number of objects in each set was large, approximately 900 in each case, and the production of the object sub images was performed automatically to match the type of object that would be encountered in a fully automated identification system. Two sets of 100 images were taken together with a ``blank'' reference image. After subtraction of the reference image thresholding and connected component labelling was used to identify the separate objects. Each object was then saved as an individual image with the object roughly centred. A fixed size border was included around the object and this meant that the object images were of different sizes as can be seen from the example objects shown in figure 2 . The two sets were labeled A and B, Set A being used to train the Paradise network and both methods were used to classify set B. For each of the sets, all of the objects which were larger than a threshold size were removed since they were judged to be too large to be pollen. This reduced the size of set A to 753 objects and set B to 818.
Figure 2:
Examples of objects found on the microscope slides. Objects 1, 3, 5, 6,
9, 10 and 11 (counting clockwise from top-left) are pollen, the rest are
non-pollen.
Mr I France