High-Speed Clustering of Regional Photos Using Representative Photos of Different Regions

In recent years, a huge number of photographs have been posted on SNS by many users, and users view photos posted by other users. When browsing photos, even if you find a photo of the scenery you want to see, it is difficult to go to that place if you were taken at a remote location such as overseas. Then, there are demands to search for areas that look like the photo in nearby places. To this end, there is a method of extracting representative photos for each area and clustering a large number of photos based on the representative ones. The k-medoids clustering method extracts representative objects called medoids and clusters them, so it coincides with this purpose, but it takes a large amount of computation time. In this paper, we aim to propose two methods of speeding up for k-medoids clustering utilizing representative photos in other areas which have been already extracted. In a method using representative photos of a single area, the clustering quality varies depending on the area to be used. It is difficult to know in advance the area that increases the clustering quality. In a method of selecting from representative photos in multiple regions, it is expected that highly accurate clustering results can be obtained because the representative photographs that minimize the objective function of the k-medoids method are selected across regions. In our experimental evaluation using large real datasets, we confirm that our proposed method works much faster than existing methods, greedy methods equipped with the lazy evaluation and the pivot pruning techniques, and obtains high quality.