Abstract:

A method for extracting image semantics of local features is presented based on the Expectation Maximization algorithm(EM). The local image features are first extracted and the visual words in codebook are  used to describe every feature, and then the semantic model mapping from lowlevel image features to highlevel image semantics is achieved by using probabilistic latent semantic analysis. The latent semantic probability distribution is calculated for local features,and their spatial distribution in image is calculated using the ExpectationMaximization algorithm. Finally, this semantic probability distribution is used to image analysis and understanding. Compared to other semanticbased image understanding methods, the proposed method extract local latent semantics directly, which does not require manual annotation. It not only obtains the local semantic information, but also receives the distribution of semantic space. And thus it is better to model the scenes. The experimental results show that this method has satisfactory classification performances on a large set of 15category scenes.

Key words: local feature；image semantic；EM algorithm；probabilistic latent semantic analysis