enhance the robustness of recove rab ility, such as FSCA
(Feedback SCA) [10] and ML using the Markov method
[15,16]. Howev er, when the number of samples increases,
the complexity of clustering also increases significantly, and
the algorithms suffer from high computation complexity . The
inherent reason is the redundancy of cluster points in the
mixing matrix estimation stage. Some papers [5,6,8] propose
the SSPs (single source points) method to solve this problem
for one-dimensional signals. Those methods have achieved
high mixing matrix estimation performance from multiple
speech sources under different mixing conditions. Unfortu-
nately , the above algorithms cannot be used in BIS directly.
In this paper, we focus on BIS based on the SCA and
exploit the detailed structure of sparsity. We aim to
decrease the complexity of the mixing matrix estimation
and enhance the accuracy, if possible. The mixing matrix
estimation is performed by applying clustering algorithms
to a subset of observations in the Haar wavelet domain
where only one source is active. These subset points are
referred to as SSPs; the aim of this paper is to find these
SSPs and to use them to estimate the mixing vectors. Both
theory and experiment show that the proposed algorithm
achieves sparser representation, faster estimation of the
mixing matrix and higher accuracy compared to other
algorithms.
The paper is organised as follows: Section 2 introduces
the basic principles of our approach. In Section 3, SSPs in the
Haar wav elet domain are discussed. The proposed algorithm
is then presented in Section 4.Finally,Section 5 provides the
experimental results of our algorithm, comparisons with
[9,10,16,17], and a byproduct for LVA (Latent Variable
Analysis).
2. Basic principles of our approach
In this section, we review the linear instantaneous
model and describe the detection of SSPs from the wavelet
representation of the mixtures. Throughout this paper, we
will denote vectors using bold lower case, matrices using
bold upper case and scalars using upper or lower case.
2.1. Linear instantaneous mixing model
The problem of BSS based on the SCA can be stated as
follows. Consider the linear model [18],
X ¼ AS þN; X A ℝ
mT
; A A ℝ
mn
; S A ℝ
nT
; ð1Þ
where X is the observed signal matrix (mixtures), A is the
mixing matrix (mixing character of channels), S is the
source signals matrix, and N is the additive noise matrix. m
denotes the number of mixtures, n denotes the number of
sources, and T denotes the number of samples.
The simplified model that ignores noise is as follows:
X ¼ AS: ð2Þ
The SCA approach uses the sparsity of sources to solve
the BSS problem. The sparsity of source signals implies
that in each column of S, there are just a few significant
values (active sources) and most of the elements are
almost zero (inactive sources). The goal of SCA is then to
estimate A and S as accurately as possible, based only on
the X and the sparsity assumption.
In the state-of-the-art approach, the SCA problem is
usually solved in two steps. The first step is the estimation
of the mixing matrix A, and the second step is the recovery
of the source signals S using the mixing matrix. Therefore,
the key challenge for the whole algorithm (two steps
method) is to estimate the mixing matrix quickly and
accurately in the determined case.
2.2. Single source points—SSPs
2.2.1. Definitions and notations
Definition 1. If there is only one active source in the
sample point of all observed signals, this point is called
as a SSP.
In the wavelet domain, if all of the wavelet coefficients
of the observed signal points respond with only one
wavelet source coefficient, this point can be considered a
SSP in the wavelet domain.
Definition 2. Suppose that T
1
ðU Þ and T
2
ðU Þ denote two
types of linear transformations or two direction matrices
of a linear transformation. Let T
1
ðXÞ be a linear transfor-
mation of X, then T
1
ðXÞ¼AT
1
ðSÞ. Let X
T
1
and S
T
1
be the
corresponding transformed coefficients matrices of X and
S, then X
T
1
¼ AS
T
1
.
2.2.2. Basic principles of SSPs in Haar wavelet domain
Let X
T
1
and X
T
2
be the two directions of Haar wavelet
transformed coefficient matrices of X, then the SSPs have
the same absolute directions (column vectors) in the Haar
wavelet domain.
The linear instantaneous mixing system X ¼ AS can be
written as follows using matrix elements.
x
11
x
12
⋯ x
1T
x
21
x
22
⋯ x
2T
⋮⋮⋱⋮
x
m1
x
m2
⋯ x
mT
2
6
6
6
6
4
3
7
7
7
7
5
¼
a
11
a
12
⋯ a
1n
a
21
a
22
⋯ a
2n
⋮⋮⋱⋮
a
m1
a
m2
⋯ a
mn
2
6
6
6
6
4
3
7
7
7
7
5
s
11
s
12
⋯ s
1T
s
21
s
22
⋯ s
2T
⋮⋮⋱⋮
s
n1
s
n2
⋯ s
nT
2
6
6
6
6
4
3
7
7
7
7
5
ð3Þ
Then X
T
1
¼ AS
T
1
can be described as follows:
x
T
1
11
x
T
1
12
⋯ x
T
1
1k
x
T
1
21
x
T
1
22
⋯ x
T
1
2k
⋮⋮⋱⋮
x
T
1
m1
x
T
1
m2
⋯ x
T
1
mk
2
6
6
6
6
6
4
3
7
7
7
7
7
5
¼
a
11
a
12
⋯ a
1n
a
21
a
22
⋯ a
2n
⋮⋮⋱⋮
a
m1
a
m2
⋯ a
mn
2
6
6
6
6
4
3
7
7
7
7
5
s
T
1
11
s
T
1
12
⋯ s
T
1
1k
s
T
1
21
s
T
1
22
⋯ s
T
1
2k
⋮⋮⋱⋮
s
T
1
n1
s
T
1
n2
⋯ s
T
1
nk
2
6
6
6
6
6
4
3
7
7
7
7
7
5
:
ð4Þ
The mixture matrix and source matrix can be written in
row vector form. Then, (3) and (4) can be described as (5)
and (6), respectively
x
1:
x
2:
⋮
x
m:
2
6
6
6
4
3
7
7
7
5
¼
a
11
a
12
⋯ a
1n
a
21
a
22
⋯ a
2n
⋮⋮⋱⋮
a
m1
a
m2
⋯ a
mn
2
6
6
6
6
4
3
7
7
7
7
5
s
1:
s
2:
⋮
s
n:
2
6
6
6
4
3
7
7
7
5
; ð5Þ
J. Xu et al. / Signal Processing 95 (2014) 58–66 59