Neural network ensembles for image identification using Pareto-optimal features

In this paper, an ensemble classifier is constructed for invariant image identification, where the inputs to the ensemble members are a set of Pareto-optimal image features extracted by an evolutionary multi-objective Trace transform algorithm. The Pareto-optimal feature set, called Triple features, gains various degrees of trade-off between sensitivity and invariance. Multilayer perceptron neural networks are adopted as ensemble members due to their simplicity and capability for pattern classification. The diversity of the ensemble is mainly achieved by the Pareto-optimal features extracted by the multi-objective evolutionary Trace transform. Empirical results show that the general performance of proposed ensemble classifiers is more robust to geometric deformations and noise in images compared to single neural network classifiers using one image feature.