Underdetermined Blind Source Separation Algorithm Based On Deep Learning

Blind source separation(BSS)is to recover the source signals from the mixtures signals of the sources without about a prior information of the mixing system and the sources.BSS technology can be widely used in various fields such as image processing,speech signal processing and biomedical signal processing,it is a research hotspot in signal and information processing.In recent years,with the development of deep learning,which has been widely used in the field of image and speech signal processing.In this paper,the problem of underdetermined blind source separation(UBSS)based on deep clustering algorithm is studied by using deep learning method,i.e.,the number of sources is more than the number of observed signals.The main contributions of this dissertation are summarized below:1.Using the deep auto-encoder,the mixtures of some speaker's speeches are mapped from a low-dimensional space to a high-dimensional space to obtain the embedding vectors of the mixtures in the time-frequency domain.The nearest neighbor clustering algorithm is used to estimate the clustering centers of the embedding vectors and the cardinality of the clustering centers,and the number of the sources is estimated by the clustering centers with larger cardinality.The source assignment algorithm is used to estimate the sources.2.Using the deep auto-encoder,the deep nearest neighbor clustering algorithm is proposed to realize the BSS of the mixtures in the underdetermined case.This algorithm adopts the nearest neighbor clustering algorithm to estimate the clustering center vectors of the high-dimensional embedding vectors for the output of the encoder.These clustering center vectors are used as the input of the decoder to train the deep auto-encoder network,and the clustering centers of the embedding vectors and the cardinality of the clustering centers are obtained.The number of the sources is estimated by the clustering centers with larger cardinality.Finally,the embedding vectors and their corresponding mixtures can be assigned to these clusters by a hard or probability assignment algorithm to reconstruct the sources.