汕头大学学报（自然科学版）

针对现有视频摘要提取方法存在的动作断裂、语义缺失等问题，研究提出了一种基于嵌入位置编码增强注意力机制的人工智能（artificial intelligence,AI）动画关键帧摘要提取方法.该方法通过GoogLeNet构建多模态特征提取体系，利用嵌入位置编码增强注意力机制提取长视频特征，结合多尺度Anchor机制和核时态分割实现镜头语义划分.通过非极大抑制与0/1背包算法优化关键镜头筛选.结果显示，在TVSum-Anime Extension和AnimeKeyFrames数据集上，该模型的精确率分别达97.84%和95.12%,F1分数为95.68%和90.12%，明显优于其余对比模型，表明研究提出的AI动画关键帧摘要提取方法有效，提升了关键帧提取的准确性与叙事连贯性，推动了动画产业向智能化生产方向发展.

Aiming at the problems of action breakage and semantic loss in existing video summarization methods, a method for extracting keyframe summaries of artificial intelligence(AI)animations based on embedded position encoding enhanced attention mechanism is proposed. This method constructs a multimodal feature extraction system using Goog Le Net, utilizes embedded position encoding to enhance attention mechanism to extract long video features, and combines multi-scale Anchor mechanism and kernel temporal segmentation to achieve shot semantic segmentation. Optimize key shot screening through nonmaximum suppression and 0/1 knapsack algorithm. The results showed that on the TVSum-Anime Extension and AnimeKeyFrames datasets, the accuracy of the model reached97.84% and 95.12%, respectively, with F1 scores of 95.68% and 90.12%, significantly better than other compared models. This indicates that the AI animation keyframe extraction method proposed in the study effectively improves the accuracy and narrative coherence of keyframe extraction, and promotes the development of the animation industry towards intelligent production.