Blind Signal Separation by Independent Subspace Analysis

Most of the proposed algorithms for solving the Blind Signal Separation(BSS) problem rely on statistical independence (independent component analysis, ICA) or at least uncorrelation assumption of source signals. However, the independence property of sources may not hold in some real-world situations, especially in biomedical signal processing and image processing, therefore the standard ICA cannot give the expected results. Independent Subspace Analysis (ISA) as an extended ICA method for BSS has more application than ICA. In this paper, we briefly present a new perspective of ISA for BSS. The general and detailed definition of the ISA model is given, the relationships between ICA and ISA methods is discuss simultaneously. Moreover, because a fundamental difficulty in the ISA problem is that-it is non unique without extra constraints, the separateness and uniqueness of the ISA model have been discussed and reviewed too. At last, the state-of-art ISA algorithms are overviewed from different theory foundations, some ISA algorithms based on the original relative gradient(natural gradient) ICA, FastICA and JADE are constructed for the BSS problem in detail.