Video summary generation based on density peaks clustering with temporal characteristics

Abstract

Video summary is a new content-based video compression technology, which can effectively find important information from the video and eliminate redundant data in video. The density peaks clustering (DPC) can quickly find the density peaks of datasets of arbitrary shapes and efficiently allocate data. In order to apply it to video summary generation, we consider the temporal characteristics of the video, and introduce it into the DPC algorithm, and propose an improved DPC algorithm with temporal characteristics (called T-DPC), which is applied for the Hue histogram clustering of video frames, and the video shot is segmented based on the clustering results. In the keyframe selection stage, calculate the similarity between each frame and its cluster center, and the entropy of each frame, then select the frame with the largest linear combination of entropy and similarity in each category as the keyframe. At the same time, the histogram intersection method is employed to remove similar frames in the keyframes to generate video summary. The proposed method in this paper is evaluated with 50 videos in the open video library. The experimental results show that the accuracy of the video summary generated by our method is higher than that of OV, STIMO, and VSUMM1, but not as good as DT and VSUMM2. The recall rate is higher than the OV, DT, and VSUMM2, the same as the STIMO, and slightly lower than the VSUMM1. The F values are all higher than the comparison algorithms OV, DT, STIMO, VSUMM1 and VSUMM2.