Scene Retrieval in Videos by Query Image

Abstract

In this paper the concept and details of an elaborated video scene retrieval system is described. The videos can be segmented into scenes, and these semantic parts of video are the available set for the end user and in our retrieval system. The end user gives a query image, and would like to retrieve the most relevant scenes from a long video, furthermore the user may wonder the results in more details ‐ in less unit than scene, e.g. in sample image ‐ ordered as well. The key problem of such retrieval is the speed of the answer, and the solution for fast search is described in this paper. An index structure, as contribution of this paper is outlined for speed up; but for building this structure a clustering phase should has been applied before. After feature extraction from query image and comparison with index structure the candidate images for relevant hits are available; however many irrelevant images also can be found in the hit list, which causes difficulty. The paper shows a solution for filtering the hits, and finally the images will be in descending order based on relevance. At the end of the paper a solution has been shown for ranking of the scenes based on the rank of the images in the hit list.