Keyframe Generation Method via Improved Clustering and Silhouette Coefficient for Video Summarization

In order to solve the issue that the traditional k-means algorithm falls into the local optimal solution in video summarization due to unreasonable initial parameter setting, a video summarization generation algorithm by using improved clustering and silhouette coefficient was proposed. Firstly, color features and texture features are extracted and fused from the decomposed video frames. Secondly, the hierarchical clustering algorithm is used to obtain the initial clustering results. And then, the improved k-means algorithm with silhouette coefficient is introduced to optimize the initial clustering results. Finally, the nearest frame from the cluster center is selected as the keyframe, and all the final keyframes are arranged in the order of the time sequence in the original video to constitute video summarization. The proposed algorithm is evaluated on two video datasets and the results show that the proposed algorithm achieves an average 84% accuracy rate and only 24% error rate in YouTube dataset. At the same time, the algorithm is validated on the benchmark Open Video Database dataset with an average 71% precision, 84% recall rate, and 76% F-score, which is higher than state-of-the-art video summarization methods. Moreover, it generates video keyframes that are closer to user summaries, and it improves effectively the overall quality of the generated summary.