Mixed-method identifications
Michel A. Westenberg and Jos B.T.M. Roerdink. Mixed-method identifications. In: Automatic Diatom Identification, Du~Buf, J. M. H. and Bayer, M. M. (eds.), Series in Machine Perception and Artificial Intelligence vol. 51, World Scientific Publishing Co., Singapore, 2002, chapter 12, pp. 245-257.
This chapter reports the identification performance obtained by making combinations of the feature sets which have been described in previous chapters. We develop an application framework that integrates the contributions of the project partners to make these mixed-method identifications possible. This chapter also describes a web-based application, called \adiacweb, that allows users to interact with the automatic identification system. Currently, it can identify 37 different taxa. Identification performance of various combinations of feature sets is measured by bootstrap aggregating (bagging) C4.5 decision trees. Combinations of contour-based features show that over 90% of the diatoms can be identified correctly, and a similar result is obtained using ornamentation features. If all features are combined, identification performance increases to almost 97%. From the analysis of a collection of 25 decision trees, a set of 17 robust features is derived by selecting those features which were used by at least 12 trees. The performance of this feature set shows an identification performance of almost 96%.