Identification by Mathematical Morphology

• Published in: Automatic Diatom Identification Vol. 51; pp. 221 - 244
• Main Authors: WILKINSON, MICHAEL H. F., JALBA, ANDREI C., URBACH, ERIK R., ROERDINK, JOS B. T. M.
• Format: Book Chapter
• Language: English
• Published: Singapore World Scientific Publishing Company 2002; WORLD SCIENTIFIC
• Subjects: Artificial intelligence; Pattern recognition
• ISBN: 9789810248864; 9810248865; 9789812777867; 9789814489218; 9814489212; 9812777865
• DOI: 10.1142/9789812777867_0011

This chapter describes diatom identification methods based on algorithms from mathematical morphology. Two types of feature extraction method are described: a contour-based method which analyzes features of the diatom outline, and another method which computes features of the valve ornamentation, i.e. the striation pattern. The first method uses morphological curvature scale spaces, in combination with unsupervised cluster analysis, in order to select the best scale-space features of contours. The second method computes multi-scale ornamentation features on the basis of size and shape distributions. The sets of feature vectors were used as input for decision-tree classifiers with bagging. An identification rate of 83.7 ± 11.7% was obtained in the case of the Sellaphora pupula data set, whereas the mixed genera data set was identified with a rate of 89.6 ± 1.9%). The latter result increased to 97.5% when the rules were relaxed to allow the occurrence of the correct taxon within the top-five matches.