基于最优化分类的视频镜头谱聚类算法

Video shot spectral clustering algorithm by optimized automatic cluster model selection

谱聚类是目前最有效的视频镜头聚类算法之一,但是如何自动选择最优化的分类个数仍是谱聚类算法中的难题。该文提出一种基于最优化分类的视频镜头谱聚类算法,对每个镜头采用分区域的Gauss混合模型(DGMM)进行特征建模,并提取模型参数特征作为镜头谱聚类的特征向量,通过构造DGMM和谱聚类的联合评价函数来自动选择最优化的分类个数和特征空间维数。实验结果表明,该文提出的算法比原有谱聚类算法分类结果更加准确和有效。

Spectral clustering is one of the most efficient video shot clustering algorithms. The automatic cluster model selection is still an open issue for the spectral clustering algorithm. This paper presents a video shot spectral clustering algorithm that incorporates optimized automatic cluster model selection. A distributed gauss mixture model （DGMM） is used to represent the spatial-temporal features of each shot with the model parameters used as the feature vectors for the spectral clustering. Both the DGMM and the spectral clustering measurements are used to in a globally optimized method to automatically select the number of clusters and the feature-space dimension. Tests show that the method gives better cluster model selections and clustering results.