An approach is proposed to obtain the dictionary capacity of bag of words(BoW) model efficiently

which is combined with The Latent Dirichlet Allocation (LDA) model to analyze the performance of scene category. Based on the feature matrix of scene image data sets constructed by SIFT feature

the affinity propagation method is firstly employed to obtain the clustering numbers

and to take the minimal clustering number as the visual dictionary capacity before generating a visual dictionary. Secondly

the scene training and testing sets are described by these visual words. Finally

the LDA model is employed to classify the testing data set. The experiments show that the proposed approach in this paper maintains higher accuracy of scene classification and can improve efficiency greatly.