A New Method to Generate Semantic Templates Based on Multilayer Perceptron

摘要

Content-based image retrieval pays more attention to reducing semantic gap. Semantic template is a promising method for reducing semantic gap, and consists of mapping between high-level and low-level visual features. The work presented here proposes a semantic template method via multilayer perceptron, which has three layers: an input layer, a hidden layer, and an output layer. In the proposed method, the pixel features of an interesting region are selected as input features, the features weights are originally designed randomly with random seeds, and softmax is selected as the activation function. Experiments show the proposed method has high accuracy for image retrieval, and the accuracy can be improved by adding samples to train the MLP (Multilayer perceptron) classifier until a relative stable state is achieved.