論文ID: e25.46
Advancements in determined blind source separation (BSS) have been achieved through two approaches: design of better source models and derivation of better optimization algorithms. This paper proposes novel BSS algorithms based on alternating direction method of multipliers (ADMM) to easily incorporate additional constraints and regularization terms into a source model, which provides more flexibility for the source model design. In addition, the structure of spatially whitened signals is effectively utilized to simplify the computation and speed up the ADMM algorithm. In experiments, we applied the proposed ADMM algorithm to the source model of independent vector analysis and compared it with the majorization-minimization (MM) algorithms. Experimental results showed that the proposed ADMM algorithm with the speeding up technique can achieve performance and convergence speed comparable to the state-of-the-art MM algorithm. Our MATLAB codes are available at https://github.com/WATARAI-Hiroko/ADMM-IVA.
