Neural Network based Classification of Pollen Grains

Abstract

Palynological data are used in a wide range of applications. A new classification algorithm is proposed for pollen grains. With a view to extract features from pollen images, an classifier algorithm is developed which proposes two-dimensional discrete Walsh-Hadamard Transform domain coefficients in addition to image statistics and shape descriptor. The suitability of classifiers based on Multilayer Perceptron (MLP) Neural Network, Generalized Feedforward (GFF) Neural Network, Support Vector Machine (SVM), Radial Basis Functions (RBF) Neural Networks, Recurrent Neural Networks (RNN) and Modular Neural Network (MNN) is explored with the optimization of their respective parameters in view of reduction in time as well as space complexity. Performance of all six classifiers has been compared with respect to MSE, NMSE, and Classification accuracy. The Average Classification Accuracy of MNN comprising of two hidden layers and four parallel MLP neural networks organized in a typical topology is found to be superior (85 % on Cross Validation dataset) amongst all classifiers. Finally, optimal classifier algorithm has been developed on the basis of the best performance. The algorithm suggested could be easily modified to classify more than 10 species. The classifier algorithm will provide an effective alternative to traditional method of pollen image analysis for plant taxonomy and species identification.