Cross-power spectrum based blind source separation method

The invention provides a cross-power spectrum based blind source separation method aiming at an existing blind source separation method and particularly at time-frequency variations under the circumstance of non-stable mixed signals. The method includes firstly, whitening input signals; secondly, computing power spectrums and cross-power spectrums of received whitened mixed signals or mixed images by the aid of a lag autocorrelation method; thirdly, organizing a power spectrum and a cross-power spectrum of each sampling point to construct a cross-power spectrum matrix, and separating a cross-power spectrum matrix of the mixed signals or mixed images by a joint diagonalization mode to obtain separated matrixes; and finally, utilizing the obtained separated matrixes for separating observed mixed image signals to achieve the purpose for blind source separation of the signals or images. The blind source separation method achieves satisfactory signal separation effect under the circumference of non-stable mixed signal mixing, and has important applications in radio communication systems and audio frequency and acoustic and medical signal processing in the industrial and military fields.