Research On Dynamic Video Summarization Technology Via Attention Mechanism

With the rapid development of computer technology and the widespread application of digital video technology,video data is growing explosively.How to effectively manage and store video information has become a difficult and hot spot in current information processing research.Video summarization for the convenience of video browsing and management has gained increasing attention.It aims to shorten the length of the video while preserving the main content,so as to still convey the entire storyline of the original video.Moreover,it can reduce the huge time spent searching for videos and save a lot of storage space.However,due to the variety of video types and the complexity of video content,higher requirements have been placed on the design of video summarization algorithms.Aiming at the dynamic video summarization problem,this paper conducts research on improving algorithm performance and reducing model complexity.The main research content includes the following three parts.(1)Based on the introduction of the latest Independently Recurrent Neural Network(Ind RNN)and the combination with attention mechanism,a novel video summarization algorithm is proposed.This method solves the problem of gradient disappearance and explosion in the traditional RNN when modeling long-term dependencies for videos with thousands of frames.At the same time,the model leverages the attention mechanism to selectively focus on the coding sequence and guide the current decoding process,which effectively improves the prediction accuracy of the model.(2)We further explore the possibility of fusing Ind RNN and attention mechanism and propose a video summarization algorithm based on Ind RNN and self-attention mechanism,which reduces the model complexity brought by the use of bidirectional RNN in the first part.The self-attention mechanism can quickly explore the internal correlation of video frames,while the Ind RNN captures the long-term temporal dependencies.The model combines the advantages of the self-attention mechanism and the Ind RNN to generate more accurate video summaries.(3)In order to further improve the running speed and parallelize the operation,we propose to implement a position-aware self-attention mechanism video summarization algorithm.To build an attention mechanism with the ability of position awareness,we apply a bidirectional mask to model the video context-aware information.Finally,the experimental results demonstrate the effectiveness of the methods.