The identification of diatoms is a very tedious work, because microscope slides must be prepared, scanned, photographed, and the diatoms found must be linked to described taxa. The latter is normally done by comparing their images with those in atlasses, which is quite difficult even for experts because of the very subtle variations of the shapes and ornamentations.
The ADIAC project aims at developing algorithms for an automatic identification of diatoms using image information, i.e. both shape and ornamentation. For this purpose the consortium, existing of diatomists, taxonomists, ecologists and pattern recognition experts, will create image databases and software for feature extraction, identification, as well as automatic microscope slide scanning including autofocusing. The ADIAC project is very innovative and can, in a near future, introduce a revolution in diatom research in all application areas.
Because ADIAC is a pilot study (it is the first project to systematically study the identification taking the ornamentation into account), active collaborations with both diatomists and pattern recognition experts are sought. Much if not all information, image databases and software, will be made freely available by means of these webpages. Please keep track of ADIAC, and contact the coordinator or the partners if you think you could contribute!
After 3 years the public database contains approx 3400 images, whereas the non-public database contains approx 2500 more images.
The online diatom browser contains approx 2300 images of 500 taxa plus 250 synonyms, about 750 names in total.
On a very difficult set of Sellaphora pupula with 6 demes, different contour feature sets and different classifiers can achieve identification rates from 75% to 90%.
On a large set of 781 images with 37 taxa, a combination of all feature sets and 10-fold cross-validation with the C4.5 decision tree yielded an identification rate of 89%. For detailed information go to ADIAC and Results section.
YOU CAN EXPERIMENT WITH AN ONLINE DEMO IDENTIFICATION:
Partner RUG has prepared a web demo version that allows to select an image that will then be identified using part of the database. The user cannot yet identify his own images (this requires a continuation project ADIAC-2), but the demo shows the results we can obtain and how diatom identification in the 21st century might look like. Click: ADIAC ONLINE DEMO
Dept. of Electronics and Computer Science - UCEH,
University of Algarve,
Campus de Gambelas, 8000 Faro, Portugal
Tel (+351 89) 800900 extension 7761 (is new number!)
Fax (+351 89) 818560
This page has been visited
times since July 1998.
Last update: February 1999 (Hans du Buf)