Research On Video Similarity Estimation And Segmentation

In the age of Internet, the flood of video data has been formed, and its size is rising sharply. The huge amounts of video data facilitate and enrich the life for people, while they also bring lots of management problems. This paper focuses on two basic problems in the area of video content analysis and measurement. First, how to improve the similarity estimation under the limitations in both image understanding techniques and distance measure methods. The second problem is the technique of video scene segmentation, which is an essential pre-processing task for a wide range of video manipulation applications, such as video indexing. In order to solve two problems, this paper presents two new algorithms, i.e., score distribution based similarity estimation improving, and the saliency and object recognition based scene segmentation approaches.1. Propose a similarity estimation improving approach via score distributions. In the approach, we propose and prove a potential hypothesis that the distributions of similarity scores are similar for true-relevant images and dissimilar for non-relevant images when they query an independent database. The approach involves the score distribution information by representing each distribution with the area under the corresponding similarity scores, then designs a weighting function. These are combined together to update the original distance between images. Experiments on three public datasets with various feature representations show that the enhanced similarity estimation remarkably outperforms the original distance measure.2. Propose a video segmentation approach based on saliency maps and object recognition. This paper gives a new and simplified definition of video scene, which is all the relations among objects in the current picture. This definition is consistent with the video scene which is usually defined as a logical story unit, and it further refines the logical story unit. The proposed method takes advantage of saliency maps and object recognition to design the segmentation algorithm, in which it is divided into the crude segmentation and the refined segmentation. The experimental results show that the method has better video scene segmentations for many types of videos, the selected key scenes have strong connectivity in video content.