CUDA框架下的视频关键帧互信息熵多级提取算法

Mutual Information Entropy Multi-level Extraction Algorithm of the Video Key Frame with CUDA

  • 摘要: 在传统视频关键帧提取过程中,需要对每一帧视频图像进行特征提取、图像匹配、重复检测等大量计算,导致算法运行时间过长。对此,该文提出了CUDA框架下的关键帧互信息熵多级提取算法。在CPU调度及GPU划分线程基础上,依据帧间三通道互信息熵,将视频序列初次划分为静态片段类和动态片段类;运用相邻帧间互信息量极小值法,将动态片段划分成多个关键子类,在关键子类中选取预备关键帧;并运用SUSAN算子分块计算,快速完成帧间的边缘匹配,从预备关键帧中滤除冗余,得到最终的关键帧序列。实验结果表明,与其他算法相比,该算法的查全率和查准率均为91%以上,提取关键帧的数量平均减少约42.82%,降低了视频数据量的存储,与其他CPU串行方法相比,其关键帧提取时间减少约50%,提高了算法运算效率。

     

    Abstract: The video key frame extraction involves feature extraction and matching, it easily leads to high computation complexity. The paper proposes mutual information entropy multi-level extraction algorithm with compute unified device architecture (CUDA). Under CPU scheduling and GPU partition thread, three-channel mutual information entropy among the frames is designed to divide the video clips into the static and the dynamic fragment coarsely. By minimum value method for inter-frame mutual information, the dynamic fragments are categorized into multiple subclasses further, from which pre-key frames are selected. Furthermore, in order to filter out the redundancy of the pre-key frames, the SUSAN operator based on block computing is used to complete the edge matching among the inter-frames, and the final key frame sequence can be obtained by the threshold setting. The experiment results show that, compared with the other algorithms, the precision and the recall ratio of the algorithm in the paper are at least 91%, and the amount of the key frames extracted is reduced by 42.82%. It greatly cut down the video data and saves storage space. Besides it, compared with the CPU serial method, the extraction time by CUDA is shorted by about 50% and it improves the efficiency.

     

/

返回文章
返回