S-VSUM: Static Video Content SUMmarization using CNN
Ashvini A. Tonge, Sudeep D. Thepade · 2022
Due to complex structure of video data, it is very challenging to find the important part of the video and to understand the content present in it. There are different types of videos like sports, scenery, movies available online and offline. There is no prior knowledge for understanding video's important content, so there is need to do video summarization with precise and abstract way. The research work presented here proposes new enhanced system of static video summarization for the video data that creates static storyboards. A static video content summarization using modified Convolution Neural Network (CNN) is proposed here. First, video segmentation is done to separate video frames from a video input signal. The features are selected using proposed modified CNN architecture considering frame level importance of each frame available in a particular scene and combines the features from the three pre-trained models. Here along with significant frames near duplicate frames are also extracted. Then these redundant near duplicate frames are eliminated using thresholding. The whole implementation is performed using the standard benchmark public video dataset Open Video Project (OVP) and SumMe. The performance evaluation of the proposed static video content summarization (S-VSUM) is done using accuracy. The result shows that the proposed static video content summarization system outperforms traditional feature-based approaches like DT, VSUMM, k-means.