An Effective Method for Scene Image Management

This paper proposes an effective method to manage the huge number of scene images captured by camera users. Here, "manage" mainly refers to cluster these images into semantically meaningful categories and further try to find some most representative images of each scene to characterize the scene. In this work, our contributions mainly focus on two aspects: (i) during the image clustering process, we propose a voting clustering method based on dense patches extracted from the image; (ii) during the representative image selection process, we propose an adjacency matrix-based method to find the best candidates as the scene's preview or peer-view. Keypoint-based method is used to determine the similarity between the images in each scene category. Finally, to verify the proposed method's performance, a dataset is constructed from the real world image collection. The experimental results show that the proposed method is very effective in managing the huge number of scene images, and also outperforms the conventional method.