.png){kind=icon}
.png){kind=icon}
Return
Deep Source Separation With Prior
.png){kind=icon}
PRE
AI
DOI：10.1109/ACCESS.2026.3675935.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Blind Source Separation (BSS) methods are used in digital signal processing to recover latent sources from a set of observed mixed signals, with little or no prior knowledge of the sources or the mixing process. Sparse Component Analysis (SCA) methods are particularly effective for addressing underdetermined cases, where the number of sources exceeds the number of observed mixtures, assuming that the mixtures are instantaneous linear combinations of the sources. These methods generally follow a two-stage approach: first, estimate the mixing matrix; second, reconstruct the original sources. The latter stage is particularly challenging, especially in the presence of noise, because it can introduce undesirable distortions and artifacts in the recovered signals. In this work, we investigate the use of a Deep Image Prior (DIP) convolutional neural network, adapted to handle one-dimensional signals, for the reconstruction stage of latent sources in the underdetermined case with instantaneous mixtures. Experimental results show that the proposed approach is competitive to traditional SCA methods based on & ell;(1) norm minimization, particularly in scenarios involving non-overlapping sources and mixtures heavily contaminated with additive noise.
Keywords:
Electronics packaging
Inverse problems
Time-frequency analysis
Estimation
Blind source separation
Vectors
Training
Proposals
Convolutional neural networks
Time-domain analysis
convolutional neural network
deep image prior
disjoint signal
sparse component analysis
sparse signals
Journal
IF:
3.6
Papers:
9.8W
Citations:
29.4W