Comparison of support vector machines and multilayer perceptron networks in building mine classification models

The augmentation of a currently employed baseline feature set for mine classifier design by "transform" or "moment" derived features, e.g. such as Discrete Cosine Transform and Pseudo-Zernike Moments, results in an aggregate feature set which is large in size. A "traditional" approach to this problem in the context of using multilayer perceptron(MLP) neural networks for classification consists first in the use of feature selection techniques, followed by some cross-validation based training algorithm. In this paper we contrast results obtained using the described "traditional" approach, with those obtained from using the Support Vector Machine(SVM) based framework for classifier design. The SVM approach is regarded as more attractive for large feature sets due to the optimization of a criterion in training, which is closely related to theoretical bounds on classifier generalization ability.