[Abstracts] [A] [B] [C] [D] [E] [F] [G] [H] [I] [J] [K] [L] [M] [N] [O] [P] [Q] [R] [S] [T] [U] [V] [W] [X] [Y] [Z]

Hans R. Thierstein, Jörg Bollmann, Miguel Vela:
Automated coccolith classification
(Talk)

We have developed a fully automated, SEM-based coccolith recognition system which can identify, measure and count common Holocene coccolith taxa. The current version of COGNIS (COmputer Guided Nannofossil Identification System) is now implemented in our fully automated Philips XL30 Scanning Electron Microscope. Classification is performed with a convolutional neural network (Brechner, 2000). Various network structures, using different taxonomic units, are being tested. Currently, the best performing net is for 14 Holocene taxa and averages 82% correct classifications, with a minimum of 42% for R. clavigera (side view) and a maximum of 96% for E. huxleyi. We used a training set of 979 preprocessed images (48x48 pixels) and a test set of 715 images. Preprocessing of objects in SEM images is necessary to achieve homologous positioning and includes translation, rotation, size-equalisation and grey-level normalisation. Preprocessing seems to be critical for the recognition rate of all tested networks. Training times range from hours to a few days on a NT-based PC (300MHz) and classification of individual images takes less than a second. Further progress will be demonstrated.

Reference

Brechner, S. 2000. Automated coccolith classification and extraction of morphological features in SEM images. Selected Readings in Vision and Graphics, 5. Hartung-Gorre Verlag Konstanz: 205pp.

[INA8] 

[Registration and Accomodation Form]
[First Circular and Pre-Registration]
[Second Circular]

 [Division of Micropalaeontology] [Department of Geosciences] [Bremen University]

 [INA Europe]  

Copyright © 2000, most recent revision July 26, 2000