一种有效的视频镜头检索方法研究

Effective Approach for Video Shot Retrieval

镜头检索是基于内容的视频检索的重要内容。本文首先尝试将组合相似性用于镜头检索。与现有方法相比,本文提出的方法强调把镜头看作一个整体,全面客观地度量两个镜头的相似度。把两个镜头的相似度度量建立在组合相似性上:镜头看成由帧序列作为样本的组合,通过核方法,在高维空间假设特征向量表示的帧序列服从高斯分布,利用概率距离公式计算出两分布之间的距离,以此作为两个镜头的相似度。考虑到检索速度问题,给出了Chernoff距离和KL散度两种概率距离的改进算法。实验对比结果证实了本文所提方法在镜头检索中的优异表现。

Shot retrieval plays a critical role in content-based retrieval. Motivated by the theory of ensemble similarity, the paper proposed a novel approach based on probabilistic distance algorithm for shot retrieval. In contrast to existing algorithms, the p An ensemble similarity is used to approach emphasizes the integral of shot for effective similarity measure. calibrate the similarity between two shots： a shot can be treated as an ensemble that consists of a sequence of multiple frames. By kernel method, in a high dimension space the feature vector represented frames can be assumed to follow a Gaussian distribution model. Then the two Gaussian distributions is computed as the similarity ilistic distance between value between two shots. To improve the retrieval speed effectively, improved algorithms of Chernoff distance and KL divergence are also Experimental results indicate that the proposed approach achieves superior performance than some existing methods.