关键帧提取是基于内容视频检索领域中一个重要的研究课题。提出了一种基于视觉注意模型的自适应视频关键帧提取方法。该方法分别提取视频中的运动和空间显著度,并用一种运动优先非线性混合模式将显著度合成为视觉注意度。在此基础上提出一种基于视觉注意度的局部和整体两级关键帧提取策略,先采用局部策略,选择镜头内注意度最大的帧作为关键帧候选;再根据视觉注意度的变化,为各个镜头自适应分配关键帧数目作为整体关键帧分配策略。实验证明,该方法提取的关键帧较为符合人类的视觉系统特性,而且该方法具有根据内容变化自适应提取关键帧等特点。
In this paper, we propose a novel video key-frame extraction method based on visual attention model. Firstly, the spatiotemporal saliency levels are generated and fused in a motion priority fashion to produce the overall attention degree. Then, a new adaptive key-frame extraction algorithm using attention and the variation of attention is put forward. For the shot level, the frames with higher attention value are selected as the candidates of the key-frames. For the clip level, the key-frame number is generated by the attention variation in a shot. Experimental results indicate the proposed method performs well in key-frame extraction with high efficiency.