Research On Key Frame Extraction Based On Clustering Algorithm And Multi-Feature Fusion

Key frame extraction is one of the core technologies of content-based video retrieval(CBVR). However, with regard to the characters of videos such as animation, advertisingfilms, action films, the performances of existing key frame extraction methods have limit.On one hand, the key frames extracted cannot effectively represent the original video. Onthe other hand, there exist redundant key frames. It will greatly and negatively affect thespeed and accuracy of video retrieval if using these key frames.To solve these problems, it is necessary to design a new key frame extraction methodfor such a genre of video sequence. It is also necessary to accelerate the offlinekey-frame-computing speed to make it more practical.This paper proposes a new key frame extraction method based on clustering algorithmand multi-feature fusion with regard to the type of videos of complex content, a largerange of scenarios and rich action. Firstly, similarity measure using multi-feature fusioncan describe the original video content more comprehensively. Secondly, directlyclustering the video sequence into different scenes eliminates difficulty and complexity ofshot detection and split. Thirdly, extracting the key frames based on minimum value ofmotion can represent the original video content more precisely.In addition, this paper has improved the performance of affinity propagationclustering algorithm when handling large-scale data.Finally, this paper discusses the design and implementation of thekey-frame-extraction system. Analysis and comparison of the results confirm thefeasibility of the method this paper proposed. The results show that the proposed systemhas high performance and accuracy. This research has theoretical and application value ontechnology of extraction of key frames and content-based video retrieval.