CID(Constrained Iterative Deconvolution)算法
CID算法是Richards, Mark在1988年提出的,大大降低了解卷积对信噪比的要求。对于原始的解卷积问题可以表示为
Σ
(
ω
)
=
Y
(
ω
)
−
N
(
ω
)
H
(
ω
)
\Sigma \left( \omega \right)=\frac{Y\left( \omega \right)-N\left( \omega \right)}{H\left( \omega \right)}
Σ(ω)=H(ω)Y(ω)−N(ω)
将其重写,可以得到
Σ
(
ω
)
=
Y
(
ω
)
H
(
ω
)
=
a
×
H
∗
(
ω
)
×
Y
(
ω
)
a
×
∣
H
(
ω
)
2
∣
=
a
×
H
∗
(
ω
)
×
Y
(
ω
)
{
1
−
(
1
−
a
×
∣
H
(
ω
)
2
∣
)
}
=
a
{
∑
n
=
0
∞
(
1
−
a
×
∑
∣
H
(
ω
)
2
∣
)
n
}
H
∗
(
ω
)
Y
(
ω
)
\begin{aligned} & \Sigma \left( \omega \right)=\frac{Y\left( \omega \right)}{H\left( \omega \right)}=\frac{a\times {{H}^{*}}\left( \omega \right)\times Y\left( \omega \right)}{a\times \left| H{{\left( \omega \right)}^{2}} \right|}=\frac{a\times {{H}^{*}}\left( \omega \right)\times Y\left( \omega \right)}{\{1-(1-a\times \left| H{{\left( \omega \right)}^{2}} \right|)\}} \\ & =a\left\{ {{\sum\limits_{n=0}^{\infty }{\left( 1-a\times \sum\limits_{{}}^{{}}{\left| H{{\left( \omega \right)}^{2}} \right|} \right)}}^{n}} \right\}{{H}^{*}}\left( \omega \right)Y\left( \omega \right) \\ & \end{aligned}
Σ(ω)=H(ω)Y(ω)=a×∣∣∣H(ω)2∣∣∣a×H∗(ω)×Y(ω)={1−(1−a×∣∣∣H(ω)2∣∣∣)}a×H∗(ω)×Y(ω)=a{n=0∑∞(1−a×∑∣∣∣H(ω)2∣∣∣)n}H∗(ω)Y(ω)
式中:
H
∗
(
ω
)
{{H}^{*}}\left( \omega \right)
H∗(ω)表示
H
(
ω
)
H\left( \omega \right)
H(ω)的共轭。当
∥
1
-a
×
∣
H
i
(
ω
)
∣
2
∥
<
1
,
a
{
∑
n
=
0
∞
(
1
−
a
×
∑
∣
H
i
(
ω
)
2
∣
)
n
}
\left\| 1\text{-a}\times {{\left| {{H}_{i}}\left( \omega \right) \right|}^{2}} \right\|<1,a\left\{ {{\sum\limits_{n=0}^{\infty }{\left( 1-a\times \sum\limits_{{}}^{{}}{\left| {{H}_{i}}{{\left( \omega \right)}^{2}} \right|} \right)}}^{n}} \right\}
∥∥∥1-a×∣Hi(ω)∣2∥∥∥<1,a{n=0∑∞(1−a×∑∣∣∣Hi(ω)2∣∣∣)n}
,上式收敛于
1
/
∣
H
i
(
ω
)
∣
2
{1}/{{{\left| {{H}_{i}}\left( \omega \right) \right|}^{2}}}
1/∣Hi(ω)∣2,因此需要选择合理的
a
a
a来满足上式。
令
H
′
(
ω
)
=
H
(
ω
)
×
H
∗
(
ω
)
,
Y
′
(
ω
)
=
Y
(
ω
)
×
H
∗
(
ω
)
{H}'\left( \omega \right)=H\left( \omega \right)\times {{H}^{*}}\left( \omega \right),{Y}'\left( \omega \right)=Y\left( \omega \right)\times {{H}^{*}}\left( \omega \right)
H′(ω)=H(ω)×H∗(ω),Y′(ω)=Y(ω)×H∗(ω)
利用迭代算法重新计算上式,可以得到
Σ
k
+
1
(
ω
)
=
a
Y
′
(
ω
)
+
(
1
−
a
H
′
(
ω
)
)
Σ
k
(
ω
)
k
=
0
,
1
,
2
⋯
{{\Sigma }_{k+1}}\left( \omega \right)=a{Y}'\left( \omega \right)+\left( 1-a{H}'\left( \omega \right) \right){{\Sigma }_{k}}\left( \omega \right)k=0,1,2\cdots
Σk+1(ω)=aY′(ω)+(1−aH′(ω))Σk(ω)k=0,1,2⋯
初始化
Σ
0
=
a
Y
′
(
ω
)
{{\Sigma }_{0}}=a{Y}'\left( \omega \right)
Σ0=aY′(ω),但是上式的求解容易受到噪声的影响,并且由于迭代过程负值没有及时被剔除而存在严重的“振铃”效应。观测数据的接收不可能存在负值,因此,在每次迭代的过程中可以先将负值剔除然后再进行迭代,这样就可以缓解最终结果出现的“振铃”效应,即
Σ
k
+
1
(
ω
)
=
F
{
P
{
F
−
1
[
a
Y
′
(
ω
)
+
(
1
−
a
H
′
(
ω
)
)
Σ
k
(
ω
)
]
}
}
k=0,1,2
⋯
{{\Sigma }_{k+1}}\left( \omega \right)=F\{P\{{{F}^{-1}}[a{Y}'\left( \omega \right)+\left( 1-a{H}'\left( \omega \right) \right){{\Sigma }_{k}}\left( \omega \right)\text{ }\!\!]\!\!\text{ }\!\!\}\!\!\text{ }\!\!\}\!\!\text{ }\text{k=0,1,2}\cdots
Σk+1(ω)=F{P{F−1[aY′(ω)+(1−aH′(ω))Σk(ω) ] } } k=0,1,2⋯
式中:
F
(
∙
)
F\left( \bullet \right)
F(∙)表示傅里叶变换,
F
−
1
(
∙
)
{{F}^{-1}}\left( \bullet \right)
F−1(∙)表示傅里叶反变换,
P
(
∙
)
P\left( \bullet \right)
P(∙)表示去除数据中出现的负值。上述操作可以在较低信噪比的情况下分辨出同一距离向的不同目标,但是计算量过于繁重,下面将介绍CID的快速算法FCID。
FCID(Fast Constrained Iterative Deconvolution)算法
将
Σ
(
ω
)
\Sigma \left( \omega \right)
Σ(ω)重写可以得到
Σ
(
ω
)
=
a
{
∑
n
=
0
∞
(
1
−
a
H
′
(
ω
)
)
n
}
Y
′
(
ω
)
=
a
{
1
+
(
1
−
a
H
′
(
ω
)
)
+
(
1
−
a
H
′
(
ω
)
)
2
+
⋯
}
\Sigma \left( \omega \right)=a\left\{ {{\sum\limits_{n=0}^{\infty }{\left( 1-a{H}'\left( \omega \right) \right)}}^{n}} \right\}{Y}'\left( \omega \right)\text{=}a\left\{ 1+\left( 1-a{H}'\left( \omega \right) \right)+{{\left( 1-a{H}'\left( \omega \right) \right)}^{2}}+\cdots \right\}
Σ(ω)=a{n=0∑∞(1−aH′(ω))n}Y′(ω)=a{1+(1−aH′(ω))+(1−aH′(ω))2+⋯}
从上式中可以看出,迭代求解中有公因子
1
−
a
H
′
(
ω
)
1-a{H}'\left( \omega \right)
1−aH′(ω),令
T
(
ω
)
=
1
−
a
H
′
(
ω
)
T\left( \omega \right)=1-a{H}'\left( \omega \right)
T(ω)=1−aH′(ω),则上式可以表示为
Σ
(
ω
)
=
a
{
1
+
T
(
ω
)
+
T
(
ω
)
2
+
T
(
ω
)
3
+
⋯
}
Y
′
(
ω
)
=
a
{
(
1
+
T
(
ω
)
)
+
T
(
ω
)
2
(
1
+
T
(
ω
)
)
+
T
(
ω
)
4
(
1
+
T
(
ω
)
)
+
⋯
}
Y
′
(
ω
)
=
a
{
(
1
+
T
(
ω
)
)
(
1
+
T
(
ω
)
2
+
T
(
ω
)
4
+
T
(
ω
)
6
+
⋯
)
}
Y
′
(
ω
)
=
a
Y
′
(
ω
)
∏
n
=
1
∞
(
1
+
T
(
ω
)
2
n
)
\begin{aligned} & \Sigma \left( \omega \right)=a\left\{ 1+T\left( \omega \right)+T{{\left( \omega \right)}^{2}}+T{{\left( \omega \right)}^{3}}+\cdots \right\}{Y}'\left( \omega \right) \\ & =a\left\{ \left( 1+T\left( \omega \right) \right)+T{{\left( \omega \right)}^{2}}\left( 1+T\left( \omega \right) \right)+T{{\left( \omega \right)}^{4}}\left( 1+T\left( \omega \right) \right)+\cdots \right\}{Y}'\left( \omega \right) \\ & =a\left\{ \left( 1+T\left( \omega \right) \right)\left( 1+T{{\left( \omega \right)}^{2}}+T{{\left( \omega \right)}^{4}}\text{+}T{{\left( \omega \right)}^{6}}+\cdots \right) \right\}{Y}'\left( \omega \right) \\ & =a{Y}'\left( \omega \right)\prod\limits_{n=1}^{\infty }{\left( 1+T{{\left( \omega \right)}^{2n}} \right)} \end{aligned}
Σ(ω)=a{1+T(ω)+T(ω)2+T(ω)3+⋯}Y′(ω)=a{(1+T(ω))+T(ω)2(1+T(ω))+T(ω)4(1+T(ω))+⋯}Y′(ω)=a{(1+T(ω))(1+T(ω)2+T(ω)4+T(ω)6+⋯)}Y′(ω)=aY′(ω)n=1∏∞(1+T(ω)2n)
则可以用上式写出FCID的迭代过程
{
T
k
+
1
=
T
k
(
ω
)
2
Σ
k
+
1
(
ω
)
=
(
1
+
T
k
(
ω
)
)
Σ
k
(
ω
)
k
=
0
,
1.2
⋯
\left\{ \begin{aligned} & {{T}_{k+1}}={{T}_{k}}{{\left( \omega \right)}^{2}} \\ & {{\Sigma }_{k+1}}\left( \omega \right)=\left( 1+{{T}_{k}}\left( \omega \right) \right){{\Sigma }_{k}}\left( \omega \right) \\ \end{aligned} \right.k=0,1.2\cdots
{Tk+1=Tk(ω)2Σk+1(ω)=(1+Tk(ω))Σk(ω)k=0,1.2⋯
式中:初始化
T
0
=
1
−
a
H
′
(
ω
)
Σ
0
(
ω
)
=
a
Y
′
(
ω
)
{{T}_{0}}=1-a{H}'\left( \omega \right){{\Sigma }_{0}}\left( \omega \right)=a{Y}'\left( \omega \right)
T0=1−aH′(ω)Σ0(ω)=aY′(ω)。
从推导过程可以看出FCID算法的运算速度比CID算法的运算速度要快很多,即CID算法运算 N N N次,则FCID算法只需要运算 log 2 N {{\log }_{2}}N log2N次就可以达到相同的效果,但是同样需要注意的是由于每次迭代的过程中可能会产生负值,造成“振铃”效应,所以在迭代的过程中需要将负值剔除。再者就是因为FCID算法可以大大提升运算的速度,这样就会造成去除负值的次数减少,从而造成噪声的累积,会对最终的成像效果造成一定的影响。实际应用过程中通常采用两者的结合:先用FCID算法将原始目标区分出来,再用CID算法减小噪声的影响,提高最终的成像效果。
参考文献
[1] R. W. Schafer, R. M. Mersereau and M. A. Richards, “Constrained iterative restoration algorithms,” in Proceedings of the IEEE, vol. 69, no. 4, pp. 432-450, April 1981, doi: 10.1109/PROC.1981.11987.
[2] M. A. Richards, “Iterative noncoherent angular superresolution (radar),” Proceedings of the 1988 IEEE National Radar Conference, 1988, pp. 100-105, doi: 10.1109/NRC.1988.10940.