基于距离相关图的音频相似性度量方法

相似性度量是基于内容的音频分析中的关键环节之一,目前往往沿用传统的统计分析思想实现,缺乏有效的呈现手段和对语义层信息的准确表达.提出了一种基于图像分割技术的相似性度量方法,通过特征提取构造特征空间,进而绘制特征向量间的距离相关图,通过对该图的可视化分析判断最大相似方向,并分别计算出局部相似性和全局相似性.给出了实验验证过程,该方法适用于数字广播等音频流中的片段查找等应用.

Audio similarity measure based on distance correlation image

Similarity measure is one of the keys of content-based audio analysis. Traditionally,statistic methods are used for accomplishing the aim.But it is hard to represent the results in a visual way and it shows very weak relationship with semantic layer.An image segmentation-based method was introduced to partly solve the problem.Features were extracted respectively to form a so-called feature space.The distance correlation images were calculated by comparing the feature vectors.A maximal similar direction was estimated by segmentation on the image to calculate local similarity and global similarity.Experiment was also provided.The method was proved to be suitable for applications like clip searching in digital broadcasting stream.