Combining Spatial and Semantic Information to Represent Image

Recently,bag-of-words(BoW) model has been approved by researchers due to their good performance.There are mainly two categories of bag-of-words models.One is to add spatial information into the image feature representation and the other is to add semantic information.This paper proposes an image feature representation method which combines the spatial information between feature points with the semantic information,and makes the feature show better performance.This paper extracts similar visual words by computing distribution divergence and forms visual phrase,which can present the meaning of image.Image classification experiments based on this method are conducted on UIUC-Sports8 dataset and Scene-15 dataset,and the results show that the visual phrase method has better classification accuracy compared with the conventional bag-of-words model and other models.