LIA2 (Laboratorio de Informatica Aplicada)

back

CLASSIFICATION AND COUNTING OF POLLEN GRAINS

Image Processing

Palynological data are used in a wide range of applications such as paleoclimatic reconstruction, allergenic process, and food technology. Fossil pollen spectra are used for paleoenvironmental reconstruction in the Quaternary and in the geological past. The count of pollen in the air allows to suggest safety treatments to patients who are allergic to pollen. Pollinic analysis of honeybees is used to determine their geographical origins. The principal tasks in almost all of these applications are classification and counting, which are laborious and time consuming due to the fact that they are done manually by highly skilled experts. A high level of training is needed to obtain accurate identification results. The morphology of the grains is fundamental to identify the type of pollen. The key parameters are shape, polarity, symmetry, aperture, size, exine stratification, and sculpturing type of the grain.

We are interested in discriminating species of the same family Urticaceae, namely, Parietaria Judaica, Urtica Urens, and Urtica Membranacea which is not achieved in routine pollinic analysis. Our attempt is to classify and count the grains of pollen in a slice that represents the visual information reported by an optical microscope. The method consists of two steps: detection of pollen grains in the slice based on the Hough transform, and their classification. The detection step includes two phases: finding the location of the grains and extracting their silhouette. The classification step also includes the phase of computing the feature vector that represents each pollen grain which includes area, diameter, compactness, roundness, holes, thickness, elongatedness, eccentricity, several statistical measures and fourier descriptors. For classification itself we use minimum distance classifiers, multilayer perceptron, and support vector machine.

The final result obtained was 86% of correct pollen classification using the most convenient combinations of features and classification algorithms. The largest values of correctness has been reached for the Urtica Membranacea with 98%, but the value fell down to 82% for the Urtica Urens and to 76% for the Parietaria Judaica.


Financed Projects


Publications


People Involved




Partners

More about Laboratorio de Informática Aplicada
Laboratorio de Informática Aplicada