Scopul nostru este sprijinirea şi promovarea cercetării ştiinţifice şi facilitarea comunicării între cercetătorii români din întreaga lume.
Domenii publicaţii > Ştiinţe informatice + Tipuri publicaţii > Capitol de carte
Autori: Adrian Ciobanu and Hans du Buf
Editorial: Hans du Buf, Micha M. Bayer, World Scientific, Automatic Diatom Identification, p.167-185, 2002.
Rezumat:
In this chapter we present an improved method for contour profiling which is based on dynamic ellipse fitting. We also apply Legendre polynomials for characterizing upper and lower contour parts, assuming that a contour with a pennate form is in horizontal position. Three classification methods are used: decision trees, backpropagation neural networks, and a hand-optimized syntactical classifier. Using only contour features of the Sellaphora pupula data set with distinct training and test sets, the characteristic profile features outperformed Legendre polynomials and correct identification rates of 91.7 resp. 93.3% were obtained by a neural network and decision tree. The hand-optimized syntactical classifier could reach 75% on the test set, but a training on all 120 available samples resulted in 99.2% when also width and striation features were used. In the case of the mixed genera data set with 48 taxa and a total of 1009 images, neural networks in combination with Legendre polynomials performed best, yielding 82%. After excluding taxa with either similar shapes, incorrect contours or an insufficient number of images, the remaining 17 taxa with a total of 359 images could be identified with an ID rate of 91.1%.
Cuvinte cheie: procesare computerizata, diatomee, clasificarea formelor // computer processing, diatoms, shape classification