The Variant of Latent Dirichlet Allocation for Natural Scene Classification

Abstract

The paper proposes a novel model based on classic LDA (latent Dirichlet allocation), which is used to learn and recognize natural scene category. Unlike previous work, the model performs variational Bayesian inference (VB) two times in order to get more precise prior Dirichlet parameters for each scene category. Although the scenes is represented in common topic simplex, the model has retained the diversities of each scene category based on the same topic simplex. Furthermore, two discriminations have been done to get good performance. We investigated the classification performance with classic 13 scenes image database and the experiments had demonstrated that our method can get better performance with less training time.