使用下采样帧构成的列向量的PCA系数作为每帧特征（原则上，也可以使用其他类型的关键帧特征），多帧特征构成特征矢量，以高斯概率模型的后验概率衡量相似性、并建立k维树结构的索引进行搜索，文中对噪声、模糊、再压缩、加Logo，及其这些变换的两两组合进行了实验（丢帧很少时也可以），速度很快，适合于大规模因特网视频的搜索。-Use sampling frame column vector of the PCA coefficients as each frame features (in principle, can also use other types of key-frame features), multi-frame features constitute the feature vector, the posterior probability of the Gaussian probability model to measure the similarity,and the establishment of the index of the k-dimensional tree structure search, to noise, blur, compression, Logo, its combination of two of these transformation experiments (dropped frames rarely), very fast, suitable for largescale internet video search.