An Efficient Model for Dynamic Video Summarization based-on Deep Visual Features

Digital video consists of frames that are presented at variant frame rates such as 15, 24, 30, 60 fps. There is a saying “A picture is worth a thousand words.” Pertaining to Digital Video the saying is “a video represents a million of those words strung together”. A digital video’s core components are frames. Video summarization is the procedure of distilling an original video to build a short raw video that keeps the important stories/content without sacrificing too much information. In this paper, a proposed model is offered for enhancement of dynamic video summarization based on deep learning producing a video skimming summary by selecting keyframes that have importance score met the specification. The task of assigning importance scores to certain frames in a video is critical for summarization, but it is also tough. Proposed model assigns a probability to each video frame, indicating how important it is to be selected, and then selects frames based on the expected values, creating video summaries. Using InceptionV3, ResNet50 pretrained models, it early fuses two feature sets for object-based and place-based in order to predict importance scores. The performance of proposed model has been compared to state-of-the-art models on TVSum dataset. Experimental evaluation show that proposed model performs better in terms of F-score to assess the effectiveness of the proposed model.