Research On Method And Application Of Source Separation Based On Multichannel Compressed Sensing

Compressive sensing breaks through the limitations of the Nyquist samplingtheory and has attracted more and more attentions in the field of signal processing.The compression efficiency can be further increased when the CS theory is appliedin multi-sensor case utilizing the correlation between the signals. And it can reducethe the pressure on hardware acquisition, transmission and storage.In the microphone array, when the sound sources are close, the microphonesmay collecte the mixed signal of sound sources; in the field of remote sensingimaging, when the spatial resolution of the sensor is not high, a pixel may containdifferent surface materials; in the EEG signal detection process, the signls thatmedical equipment collect may include signals from other parts of the body. Inthese cases, what the sensors collected are mixed signals. However, the sourcesignals are more valuable rather than the mixed signal. Therefore, it is necessary toresearch source separation methods under the framework of the multi-channelcompressed sensing. The main research contents and results are as follows:1. Summarized the general source separation algorithms based on multichannelcompressed sensing.Summarized three mixture models and three compressedsampling models of multichannel signals. Four joint sparse models in multichannelcompressed sensing area are introduced and the joint sparse models of mixed signalsare analyzed. Describe the major steps of the general algorithm, including thereconstruction of mixed signals and source separation. When there is no noise andnoisy, design the experiments of separating the voice mixtures, music mixtures andthe EEG mixtures by general algorithms. Experimental results demonstrate the goodperformance of the algorithms.2. Proposed the source separation method based on alternate estimation methodunder the framework of the multichannel compressed sensing. Describe the type ofsource which applies to a given situation. Introduce the principle and steps of thealgorithm. Namely, update the source signal and maixing matrix by iterations. Whenthe mixing matrix is fixed, reconstruct the source signals by compressed sensingalgorithms. When source signals are fixed, update the mixing matrix by maximizingthe posterior probability of the observed signals. When there is no noise and noisy,design experiments to verify the algorithm performance. The results show our algorithm have good performance. Design the comparative experiments between thealternating iterative algorithm and the general algorithm. When the compression ratiois less than0.6, the alternating iterative algorithm outperforms common algorithm.3. Applied the source separation methods under the framework of multichannelcompressed sensing to unmixing of high spectral image. Apply the alternateestimation algorithm to the separation of spectral reflectance matrix and theabundance matrix. When the priori information of endmenber is known, apply theclass algorithm to separation. When the priori information of endmenber is unknown,apply the alternative estimation method to separation. Experimental results show thatthe proposed method has good separation effect.