結合時頻遮罩與壓縮感測之盲訊號源分離方法

Blind Source Separation Using Time-Frequency Masking and Compressive Sensing

張淳皓(Chun-hao Zhang)

指導教授 ： 王家慶

國立中央大學/資訊電機學院/資訊工程學系/碩士(2012年)

摘要

為了解決旋積盲訊號源分離這個問題，本論文提出了一個結合時頻遮罩分離與壓縮感測新方法。首先我們先定義兩個特徵參數包括了level-ratio以及phase-difference，然後利用KNN Graph方式，去除資料中的離群樣本，並用K-Means演算法對每個頻帶非離群資料群聚。運用DOA方位角估測方法可以將每個頻帶的群聚中心，找出每個群聚中心角度後可以找出其他頻帶有相近方位角群聚類別視為同一來源訊號。將先前利用KNN Graph方式設定為離群資料的資料點重新分群後。對於每個來源的時頻遮罩已經可以藉由群聚結果計算出來。我們利用壓縮感測去估算那些時頻域上未知的資料，藉此增加每個分離訊號分離效果。我們運由KSVD演算法將時頻能量矩陣訓練出重建字典。

關鍵字

時頻遮罩 ； 壓縮感測 ； 盲訊號源分離

並列摘要

To solve the convolutive blind source separation (BSS) problem, this thesis presents a new method which integrates time-frequency masking and compressive sensing (CS). We first define two features called level-ratio and phase-difference. Next, we eliminate outliers by KNN graph and use K-Means clustering to obtain the separated clusters in each frequency bin. A DOA detection method is then used to associate the cluster centroid with the corresponding source and this procedure is performed in all the frequency bands. The outliers eliminated by KNN graph are then reassigned to cluster centroids and time-frequency masking associated with each source can be designed. We use compress sensing (CS) to impute the unknown time-frequency points to enhance the quality of the separated sources. To build the atoms of the redundant dictionary for CS, frequency magnitude vectors obtained by short time Fourier Transform are trained by K-SVD algorithm to assure the sparseity of the dictionary.

並列關鍵字

Blind Source Separation ； Compressive Sensing ； Time-Frequency Masking

參考文獻

〔1〕 A. Hyvarinen, J. Karhunen, and E. Oja, Independent Component Analysis. New York: Wiley, 2001.

〔3〕 S.C. Douglas, H. Sawada, and S. Makino, “A spatio-temporalFastICA algorithm for separating convolutivemixtures,” Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing, Philadelphia, PA, vol. 5, pp. 165-168, Mar. 2005.

〔4〕 H. Saruwatari, T. Kawamura, T. Nishikawa, A. Lee, and K. Shikano, “Blind source separation based on a fast-converging algorithm combining ICA and beamforming,” IEEE Trans. Audio, Speech, Language Process., vol. 14, no. 2, pp. 666–678, Mar. 2006.

〔5〕 A. Belouchrani and M. Amin, “Blind source separation based on timefrequency signal representation,” IEEE Trans. Signal Processing, vol. 46, pp. 2888–2898, Nov. 1998.

〔6〕 J.-F. Cardoso, “Blind signal separation: Statistical principles,” Proc. IEEE, vol. 9, pp. 2009–2025, Oct. 1998.