A Graph Representation Based Fuzzy C Means Approach to Video Summarization

The Video summarization normally includes shot boundary detection, key frame extraction from each shot and generation of summary. The proposed video summarization approach eliminates shot boundary detection by employing segmentation based key frame extraction scheme and the segmentation is achieved by fuzzy C Means (FCM) clustering. The video summarization algorithms basically perform structural analysis of the video. Recently algorithms utilizing graph theoretic approaches for structure analysis are found in literature. The proposed method models a video segment formed by FCM clustering as undirected weighted graph with frames in cluster as nodes of graph and Euclidean feature distance as edge weight. The eccentricity of each graph node is used to determine its connectivity to other nodes of the graph. If the eccentricity of node is greater than the mean eccentricity of graph that node is one of the important node and the frame represented by that node is selected as key frame. This is done over all the clusters and the summary is created using key frames and merging them on the basis of their timeline. This method ensures that video summary represents the most unique frames of the input video and gives equal attention to preserving continuity of the summarized video. The performance of the algorithm is evaluated for compactness, fidelity and informativeness measures.