Using Neural Network to combine measures of word semantic similarity for image annotation

Abstract

This paper proposes a Feed-forward Neural Network (FNN) based method to combine word-to-word semantic similarity metrics for improving the accuracy of image annotation. The network fuses various estimates of word similarity to output a hybrid score which is used in the random walker with restarts method of image annotation refinement. A particle swarm optimization algorithm is designed to train the network to achieve the optimal annotation accuracy. Each particle represents a FNN configuration, the fitness value of which is the accuracy evaluation of image annotation based on the corresponding FNN. We conducted the experiments of image annotation on the Corel-5K dataset. The experimental comparisons between single measures and our combined measure show that the proposed method is effective and promising.