1 Institute of Chemical Engineering, Biotechnology and Environmental Technology, Faculty of Engineering, SDU2 The Maersk Mc-Kinney Moller Institute, Faculty of Engineering, SDU3 Institute of Chemical Engineering, Biotechnology and Environmental Technology, Faculty of Engineering, SDU4 The Maersk Mc-Kinney Moller Institute, Faculty of Engineering, SDU
One branch of general image processing research deals with 2D object classification where classes are categorized by different features of the objects such as area, perimeter, elongation, color and texture. When dealing with plant specie classification some of the widely used and well known object features are less useful because the task is to categorize soft objects in outdoor scenes. A general feature set for robust description of soft objects such as plants in an early growth stage is, to our knowledge non existing. We propose a novel way of parametrizing a distance transformation of an object silhouette that may prove to posses value in object classification. The method approximate the distance distribution of an object with a high degree Legendre polynomial where the polynomial coefficients constitutes a feature set. This feature set will be referred to as Legendre Polynomial Feature Set (LPFS). The method have been tested through a discrimination task where two similar plant species were to be divided into their respective classes. Since the LPFS feature set is meant to be used with a classification algorithm, the performance assessment is the classification accuracy of 4 different classifiers (kNN, Naive-Bayes, Linear SVM, Non-linear SVM). A set of well known features is used for comparison. This feature set will be referred to as Standard Feature Set (SFS). The used dataset consisted of 139 samples of Corn Flower (Centaura cyanus L.) and 63 samples of Night Shade (Solanum nigrum L.). The highest achieved discrimination accuracy with the LPFS feature set was 98.75 % and contained 10 numerical features. The SFS feature set achieved an accuracy of 87.1 % using 22 features. The results show the LPFS feature set can compete with the SFS feature set. Further testing is needed to reveal the true value of the LPFS feature set.