基于时频单源点的欠定混合矩阵估计

摘要

This paper focuses on the mixing matrix estimation in Underdetermined Blind Source Separation(UBSS). A method of single source points detection in the Time-Frequency(TF)domain is proposed, which detects the points by comparing the normalized real and imaginary parts of the mixtures. The proposed method is simple and effective, which also relaxes the condition on detecting the single source points compared with other existing method. Then K-means clustering method is adopted to estimate the mixing matrix. The performance of the mixing matrix estimation is further improved by removing those points which are far away from the mean direction of each cluster. It is experimentally shown that the proposed mixing matrix estimation algorithm estimates the mixing matrix with high accuracy compared with other algorithms.%主要研究了欠定盲源分离中的混合矩阵估计问题。提出了一种检测时频单源点的新方法，通过比较归一化的观测信号时频点的实部和虚部向量来检测时频单源点。与其他时频单源点检测方法相比，该方法简单而有效，同时降低了对检测条件的要求。采用K-means方法估计混合矩阵，通过去除聚类后每一类数据中偏离中心方向较远的数据点，进一步提高了混合矩阵的估计精度。仿真实验表明，与已有欠定混合矩阵估计算法相比，提出的方法有更高的估计精度。