Mean multiclass Type I and II errors for training multilayer perceptron with particle swarm in image segmentation

Abstract

Image segmentation can be posed as a multiclass classification problem. In doing so, segmentation evaluation can be made through multiclass classification errors. Instead of being used for evaluation, in this work the mean multiclass type I and II errors are proposed for multilayer perceptron training via particle swarm optimization. Moreover, some relations involving mean multiclass errors and conditional errors are exposed. Applied to image segmentation, mean multiclass errors were compared to mean squared error as objective functions. The approach was effective and able to provide accuracy and precision gains, resulting in a lower number of function evaluations in a cross-validated experiment. © 2012 Springer-Verlag.