Abstract

Background: Scene image classification is a fundamental problem in the field of computer vision, as described in various patents. But, so far, it is still a challenging task to solve the semantic gap of the scene image between low level feature and high level topic.

Method: In this paper, we propose a new scene image classification method based on super pixel segmentation and correlated topic model. The method is composed of the following steps: Firstly, considering super pixel providing the spatial support for computing region, we divide image into sub-regions through super pixel segmentation model. Then, each sub-region is described by lots of local scale invariant feature transform key points. In order to preserve the mode information of key points, we use Median- shift method to build word of bag to represent image. Lastly, in order to reflect the relation of the low level features and the high topics of images, we use a correlated topic model based on word of bag to classify scene image.

Result: We evaluated the proposed method on the classical Caltech 10 database. The experiment results show that the presented method have average precision rate with 72.6% for scene image classification.

Conclusion: From the experimental results we can draw the conclusion that the super pixel segmentation method can preserve more spatial support to scene image, and the correlated topic model can mine the high-level semantic information scene categories from low-level feature, which make the presented method highly completive than other approaches.

Keywords: Scene image, super pixel segmentation, scale invariant feature transform, correlated topic model, median-shift method.