曲阜师范大学学报（自然科学版）2025，Vol.51Issue(2)：73-79,7.DOI:10.3969/j.issn.1001-5337.2025.2.073
嵌入位置编码增强注意力机制的视频摘要改进算法
A video summarization improvement algorithm with embedded position encoding enhanced attention mechanism
摘要
Abstract
In order to solve the problems that the attention mechanism in the video summarization algo-rithm lacks the temporal order information of video frames and dynamic visual contextual content informa-tion,a video summarization algorithm that embeds position encoding to enhance the attention mechanism is proposed.Firstly,absolute position encoding is introduced in the multi-head attention mechanism to model the frames in temporal order and learn the correlation between frames.Then the image features and tempo-ral features of the video are fused and several shot segments are generated using the multi-scale Anchor mechanism processed by the video segmentation module.Finally,the generated shots are subjected to posi-tion regression and importance prediction,and key shots are selected to generate a video summarization.The proposed method has been experimentally validated to achieve an F1-score index of 52.5％on the SumMe normative dataset and 62.3％on the TVSum normative dataset.The algorithm effectively im-proves the accuracy of the video summarization task,and the algorithm research provides theoretical sup-port for the technical implementation of video summarization algorithms.关键词
视频摘要/多头注意力机制/位置编码/anchor机制Key words
video summarization/multi-head attention mechanism/position encoding/anchor mechanism分类
信息技术与安全科学引用本文复制引用
张雅,王玉德,樊令冲,杜婉童,林元元..嵌入位置编码增强注意力机制的视频摘要改进算法[J].曲阜师范大学学报（自然科学版）,2025,51(2):73-79,7.基金项目
国家自然科学基金(52102005) （52102005）
山东省研究生导师指导能力提升计划(SDYY18119) （SDYY18119）
山东省研究生教学案例库建设(SDYAL21090). （SDYAL21090）
