# College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

1 、预备知识

( 处理 ) 工具，是在克服傅立叶变换缺点的基础上发展而来的，

、滤波器等的基础知识。

College of Mathematics and Computer Science, Hebei University

+∞

−∞

2

f (t ) dt = 0

(1.1)

College of Mathematics and Computer Science, Hebei University

512x512 ，量化级是 256 ，即
y

0 ≤ f ( x, y ) ≤ 255

0 ≤ x, y ≤ 511

x

College of Mathematics and Computer Science, Hebei University
2 、 L2(R) 空间的正交分解和变换 [1]

R ， i=1,2,… 使得

f (t ) =

+∞

∑ c g (t )
i

(1.2)

i

i =1

ci =< f (t ), g i (t ) >=
< g k (t ), g l (t ) >=

+∞

−∞

+∞

−∞

f (t ) g i (t )dt

g k (t ) g l (t )dt = δ kl，k , l ∈ Z (1.3)

College of Mathematics and Computer Science, Hebei University

f(t) 在这组基下的表现呈现出我们需要的特性，但是如果某

(1) K-L 变换
(2) Walsh 变换
(3) 傅立叶变换
(4) 小波变换

(1)小波必须时振荡的；
(2)小波的振幅只能在一个很短的一段区间上非零，即是局

College of Mathematics and Computer Science, Hebei University

1 、 Daubechies 小波

College of Mathematics and Computer Science, Hebei University

2 、 Coiflets 小波

3 、 Symlets 小波

College of Mathematics and Computer Science, Hebei University

Morlet 小波

5 、 Mexican Hat 小波

6 、 Meyer 小波

SKIP

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University
RETURN

College of Mathematics and Computer Science, Hebei University
3 、傅立叶变换与时频分析 [4]

a0
f (t ) =
+
2

+∞

∑ (a

k

cos kω0t + bk sin kω0t )

(1.4)

i =1

College of Mathematics and Computer Science, Hebei University

2
ak =
T
2
bk =
T

T

0

T

0

f (t ) cos kω0tdt，k = 0,1,2

(1.5)

f (t ) sin kω0tdt，k = 0,1,2

(1.6)

，即

f (t ) ⇔ { a0 , (a1 , b1 ), (a2 , b2 ),}

(1.7)

College of Mathematics and Computer Science, Hebei University

lim

T

N →∞ 0

 a0
f (t ) −  +
2

2

( ak cos kω0t + bk sin kω0t )  dx = 0
k =1
 (1.8)
N

fˆ (ω) =

+∞

−∞

f (t )e

−iωt

dt

(1.9)

f (t ) =

+∞

−∞

ˆf (ω)e iωt dω

(1.10)

College of Mathematics and Computer Science, Hebei University

 傅立叶分析不能刻画时域信号的局部特性；
 傅立叶分析对非平稳信号的处理效果不好。

College of Mathematics and Computer Science, Hebei University

，在哪一时刻有低音，因此结果是所有的音符都挤在了一起，如图所示。

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

，同时也能在频域反映信号的局部性，这种数学工具就是“小波”。从

 任何复杂的信号 f(t) ，都能由一个母函数 (t) 经过伸缩和平移产生

 信号用新的基展开的系数要能反映出信号在时域上的局部化特性；
 新的基函数 (t) 及其伸缩平移要比三角基 sint 更好地匹配非平稳信

1

ψ(t ) = 
−1

 1
x ∈0, 
 2
1 
x ∈ ,1
2 

(1.11)

College of Mathematics and Computer Science, Hebei University

{ψ (2 t − k ) | j, k ∈ Z }
j

(1.12)

j
ψ j ,k (t ) = 2 2 ψ (2 j t − k ) ( j , k ∈ Z )
(1.13)
2

∈L2(R) 可以分解为

f (t ) =

∑∑ c

ψ j ,k (t )

(1.14)

j ,k

j∈Z k∈Z

+∞

−∞

f (t )ψ j ,k (t )dt

(1.15)

College of Mathematics and Computer Science, Hebei University

Cψ =

+∞

−∞

(t)∈L2(R) 称为基本小波，如果它满足以下的“

ψˆ (t )
dω < ∞
ω

(2.1)

+∞

ψˆ (0) = 0 ⇔ ∫ ψ (t )dt = 0
−∞

(2.2)

College of Mathematics and Computer Science, Hebei University
ψ(t) 又称为母小波，因为其伸缩、平移可构成 L2(R) 的一个标

1
2

t −b
+
ψ a ,b (t ) = a ψ 
，a ∈ R , b ∈ R
 a 

(2.3)

(Wψ f )(a, b) =< f (t ),ψ

−j

−j

a ,b (t )

a = 2 ，b = 2 k，j , k ∈ Z

>

(2.4)

(2.5)

College of Mathematics and Computer Science, Hebei University

( DWψ f )( j , k ) =< f (t ),ψ j ,k (t ) >

ψ j ,k (t ) =

j
22

ψ (2 t − k )，j , k ∈ Z
j

(2.6)
(2.7)

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

 多尺度 / 多分辨的特点，可以由粗及细地处理信号

 可以看成用基本频率特性为 (ω) 的带通滤波器

 适当地选择小波，使 ψ(t) 在时域上为有限支撑 ,Ψ
(ω) 在频域上也比较集中，就可以使 WT 在时、频域

College of Mathematics and Computer Science, Hebei University

 尺度伸缩

f (t ) = sin(t ); a = 1
f (t ) = sin( 2t ); a = 1

2

f (t ) = sin( 4t ); a = 1

4

College of Mathematics and Computer Science, Hebei University

f (t ) = ψ (t ); a = 1
f (t ) = ψ (2t ); a = 1
f (t ) = ψ (4t ); a = 1

2

4

College of Mathematics and Computer Science, Hebei University

 时间平移

College of Mathematics and Computer Science, Hebei University

(1) 选择一个小波函数，并将这个小波与要分析的信

(2) 计算在这一时刻要分析的信号与小波函数的逼近

College of Mathematics and Computer Science, Hebei University

(3) 将小波函数沿时间轴向右移动一个单位时间，然后

College of Mathematics and Computer Science, Hebei University

(4) 将所选择的小波函数尺度伸缩一个单位，然后重复

(5) 对所有的尺度伸缩重复步骤
(1) 、 (2) 、 (3) 、 (4) 。

College of Mathematics and Computer Science, Hebei University

 尺度与频率的关系

 小尺度 a → 压缩的小波 快速变换的细节 高频部

 大尺度 a → 拉伸的小波 缓慢变换的粗部 低频部

College of Mathematics and Computer Science, Hebei University

1 、多分 辨分析 (MRA) 的概念 [5]

j

ψ j ,k (t ) = 2 2 ψ (2 j t − k )，j , k ∈ Z，t ∈ R

(3.1)

2
Z ，这个闭子空间序列充满了整个 L (R) 空间。
 在 V0 子空间找一个函数 g(t) ，其平移 {g(t-k)}k ∈Z 构成 V0 子空

 对函数 g(t) 进行正交化，得到函数称为正交尺度函数 (t) 。
 由 (t) 计算出小波函数 (t) 。

College of Mathematics and Computer Science, Hebei University
Riesz 基

{

}

1) span g j (t ) | j ∈ Z = H ,即∀f ∈ H , ∀ε > 0,

{ }

n

∑c g
2) 存在常数0 < A ≤ B < ∞, 使得∀{c }

2

l
, 使得 f (t ) −
j∈Z

j

+∞

A

∑c

j = −∞

2
j

+∞

∑c g
j

j = −∞

j

≤B

(3.2)

j =− n

j

2

j (t )

+∞

∑c

2
j

2

l
,有
j∈Z

(3.3)

j = −∞

A 和 B 分别称为 Riesz 基的上下界， Riesz 基又称为稳定基。

College of Mathematics and Computer Science, Hebei University

{V j } j∈Z

1)单调性 :  ⊂ V−1 ⊂ V0 ⊃ V1 ⊂ ;
2)逼近性 :

V = { 0} ,

2

V j = L ( R );

j

j∈Z

j∈Z

3)伸缩性 : f (t ) ∈ V j ⇔ f (2t ) ∈ V j +1 ;
4)平移不变性 : f (t ) ∈ V j ⇒ f (t − k ) ∈V j , ∀k ∈ Z ;

5)存在函数g (t ) ∈V0 , 使得{ g (t − k )} k∈Z 构成V0的Riesz基。

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

j

ϕ j ,k (t ) = 2 2 ϕ (2 j t − k )，j , k ∈ Z，t ∈ R

(3.4)

∀f (t ) ∈ L2 ( R)，则f (t )在每个V j空间的正交投影是
fV j =

∑ < f (t ),ϕ
j∈Z

j , k (t )

> ϕ j ,k (t )

(3.5)

College of Mathematics and Computer Science, Hebei University

(t) 并不是 L2(R

) 空间的小波函数，而是与其

Vj 是 Vj+1 的严格子空间，设 Wj

Vj⊕

( 子空间 ) ，即
Wj
j +1 = Vj+1

(

W j，且V j ⊥ W j

)

(

)

= V j − 2 ⊕ W j − 2 ⊕ W j −1 ⊕ W j = 
j

l = −∞

(3.6)

College of Mathematics and Computer Science, Hebei University
+∞

j → +∞

l = −∞

(3.7)

{ }

j∈Z

College of Mathematics and Computer Science, Hebei University

，假设有一幅 256 级量化的图像，不妨将它看成量化空间 Vj
V j = V j −1 ⊕ W j −1

(t)∈W0 ，使得 {ψ(t-k)}k∈Z

j

ψ j ,k (t ) = 2 ψ (2 t − k )
2

j

{

(3.8)

}

j , k∈Z

SKIP

College of Mathematics and Computer Science, Hebei University
Vj

Wj-1

RETURN

Vj-1

College of Mathematics and Computer Science, Hebei University
MRA 非常抽象，但是它给出了构造小波的一般框架。在实践中

2 、双尺 度方程

ϕ (t ) ∈V0 ⊂ V1
ψ (t ) ∈W0 ⊂ W1

ϕ (t) = ∑ h kϕ (2t − k)

(3.9)

ψ (t) = ∑ g kϕ (2t − k)

(3.10)

k

k

College of Mathematics and Computer Science, Hebei University

(t) 的正交性可得

hk = ϕ (t ),ϕ (2t − k ) ，k ∈ Z

(3.11)

g k = ψ (t ),ϕ (2t − k ) ，k ∈ Z

(3.12)

ω  ω 
(3.13)
ϕˆ (ω ) = h ϕˆ  

2 2
ω  ω 

ψˆ (ω ) = g  ϕˆ  
2 2

(3.14)

College of Mathematics and Computer Science, Hebei University

1
h(ω ) =
2

∑h e

1
g (ω ) =
2

k

− ikω

(3.15)

k

g k e −ikω

(3.16)

k

(t) 对应的高同滤波器， {h ， g} 既可以表示为时域上的离散序

(ω) ， g(ω)} 。两者本质上是一样的。

College of Mathematics and Computer Science, Hebei University

supp=[0 ， N] ， (3.9) 式变为：
N

ϕ (t ) = ∑ hk ϕ (2t − k )

(3.17)

k =0

College of Mathematics and Computer Science, Hebei University

ω  ω  ω  ω  ω 
ϕˆ (ω ) = h ϕˆ   = h h ϕˆ  
2 2 2 4 4
ω  ω  ω  ω 
= h h h ϕˆ   =  =
2 4 8 8

ω
h j
j =1  2
n

 ω 
ϕˆ  n 
 2 

(3.18)

ω 
h j 是收敛的，即
j =1  2 
n

+∞

ω
ϕˆ (ω ) =
h j
j =1  2

ϕˆ (0)

(3.19)

College of Mathematics and Computer Science, Hebei University

(3.9) 的序列 {hk}k∈Z ，通过公

ϕˆ (ω )

，经过傅立叶反变换，最终可得尺度函

College of Mathematics and Computer Science, Hebei University

(t) 与滤波器系数 {hk}k∈Z 之间的关系问题：
 如果有一个 L2(R) 空间的尺度函数 (t) ，一定能构造出双尺

 反过来，如果有一组滤波器 {hk}k∈Z 满足某个双尺度方程，由

2
2
(3.20)

h(0) = 1

(3.21)

(3.21) 称为构造正交小波的必要条件 。

College of Mathematics and Computer Science, Hebei University

(3.21) 外，还应满足其他条件。 S. Mallat[4] ， W. Lawton[6]

2

2

h(ω ) + h(ω + π ) = 1

h ( 0) = 1
+∞

ω

h j
j =1  2

aij =

N

∑h h
k

k =0

j − 2 i + k，

，由{ hk } k∈Z 构造矩阵A( 2 N −1)( 2 N −1)

N + 1 ≤ i, j ≤ N − 1

College of Mathematics and Computer Science, Hebei University

0,1,…,N

2
2

h(
ω
)

h(ω ) + h(ω + π ) = 1和h(0) = 1

+∞

ω 
ϕˆ (ω ) =
h j 
j =1  2 

(3.10) 计算。

College of Mathematics and Computer Science, Hebei University
3 、小波 与共轭 镜像 滤波器 [4]

{h(ω),g(ω)}ω∈R 从频域上反映出来，甚至离散小波变换都可以借

3.1 正交尺 度函数 产生 共轭镜 像滤 波器

h(ω) 满足以下条件：
2

2

h(ω ) + h(ω + π ) = 1

h ( 0) = 1
h(π ) = 0

College of Mathematics and Computer Science, Hebei University

{hk}k∈Z 作用后，其低频成分能被保留下来，而高频成分 (ω=π)

2
2

g (ω ) + g (ω + π ) = 1
(3.22)

ϕ (t − k ),ψ (t ) = 0

(3.23)

RETURN

College of Mathematics and Computer Science, Hebei University
S. Mallat[4] 同时给出了这样的结论：若高通滤波器 g(ω) 满足公式
(3 .2 2) 和(3 .2 3) ，则由公式

ω  ω 

ψˆ (ω ) = g  ϕˆ  
2 2

(3.23) 的 g(ω) ，就一定能找到对应的小波 (t) ，但是这样的解

g (ω ) = e −iω h(ω + π )

(3.24)

g k = (−1)1− k h1− k

(3.25)

College of Mathematics and Computer Science, Hebei University

R ，产生一组共轭镜像滤波器 {h,g} ，满足：
2

2

h(ω ) + h(ω + π ) = 1
2

2

g (ω ) + g (ω + π ) = 1

(3.26)

h(ω ) g (ω ) + h(ω + π ) g (ω + π ) = 0

College of Mathematics and Computer Science, Hebei University

 在频域上 (3.26) 式成立；
 在时域上以下公式成立：

∑h h
j

j∈Z

∑g
j∈Z

j g j +2k

∑h g
j

j∈Z

j + 2k

j +2k

= 2δ k ,0
= 2δ k ,0

k ∈Z

(3.27)

=0

 定义调制矩阵

 h(ω ) h(ω + π ) 

m(ω ) = 
 g (ω ) g (ω + π ) 

(3.28)

College of Mathematics and Computer Science, Hebei University

m(ω )mT (ω ) = 1，∀ω ∈ R

(3.29)

3.2 利用共 轭镜像 滤波 器实现 快速 正交 小波变换 [4]
L2(R) 空间的一个 MRA 产生了两个子空间：尺度空间 {Vj}j∈Z

 fV = < f (t ),ϕ j ,k (t ) > ϕ j ,k (t )
 j k∈Z

 fW j = < f (t ),ψ j ,k (t ) > ψ j ,k (t )
k∈Z

(3.30)

College of Mathematics and Computer Science, Hebei University

f (t ) =

∑f
j∈Z

Wj

=

∑∑ < f (t ),ψ

j , k (t )

> ψ j ,k (t )

(3.31)

j∈Z k∈Z

= VJ − 2 ⊕ WJ − 2 ⊕ WJ −1

= V j ⊕ W j ⊕ W j +1 ⊕  ⊕ WJ −1

( j < J)

VJ = VJ −1 ⊕ WJ −1

f (t ) =

∑ < f (t ),ϕ
k∈Z

j , k (t )

> ϕ j ,k (t ) +

∑ ∑ < f (t ),ψ

j ≤ j ′< J k∈Z

j ′, k (t )

> ψ j ′,k (t )

College of Mathematics and Computer Science, Hebei University

 c j ,k =< f ,ϕ j ,k >

d j ,k =< f ,ψ j ,k >

{

ϕ j ,k = ∑ < ϕ j ,k ,ϕ j +1,n > ϕ j +1,n
n

}

j , n∈Z

(3.32)

College of Mathematics and Computer Science, Hebei University

< ϕ j ,k ,ϕ j +1,n >=
=

1
2

+∞

−∞

2j

2∫

+∞

−∞

ϕ (2 t − k )ϕ (2
j

t − n)dt

j

ϕ (t )ϕ (2t + 2k − n)dt (令t = 2 t − k )

hk =< ϕ (t ),ϕ (2t − k ) >=

j +1

< ϕ j ,k (t ),ϕ j +1,n (t ) >=

1
2

+∞

−∞

ϕ (t )ϕ (2t − k )dt

hn − 2 k

College of Mathematics and Computer Science, Hebei University

ϕ j ,k =

1

h

2

n−2k

1

h ϕ

2

ϕ j +1,n =

n′

n

n′

j +1, n′+ 2 k (令n′

= n − 2k )

c j ,k =< f (t ),ϕ j ,k >=

1

h

2

n

< f (t ),ϕ j +1,n + 2 k >

n

=

1

hc

2

n j +1, n + 2 k

n

c j ,k =

1

hc

2
n∈Z

n j +1, n + 2 k

(3.33)

College of Mathematics and Computer Science, Hebei University

ψ j ,k = ∑ < ψ j ,k ,ϕ j +1,n > ϕ j +1,n
n

< ψ j ,k ,ϕ j +1,n >=
=

1
2

+∞

−∞

2j
2

+∞

−∞

ψ (2 j t − k )ϕ (2 j +1 t − n)dt

ψ (t ′)ϕ (2t ′ + 2k − n)dt (令t ′ = 2 j t − k )

(3.34)

College of Mathematics and Computer Science, Hebei University

g k =< ψ (t ),ϕ (2t − k ) >=

1
2

+∞

ψ (t )ϕ (2t − k )dt

−∞

g n−2k

ψ j ,k =

1

g

2

n−2k

ϕ j +1, n

n

=

1

g

2
n′

ϕ j +1,n′+ 2 k (令n′ = n − 2k )

n′

College of Mathematics and Computer Science, Hebei University

d j , k =< f (t ),ψ j ,k >=

1

g

2

n

< f (t ),ϕ j +1,n + 2 k >

n

=

1

g c

2

n j +1, n + 2 k

n

d j ,k =

1

g c

2
n∈Z

n j +1, n + 2 k

(3.35)

College of Mathematics and Computer Science, Hebei University

{cj,k,dj,k} ，这就是著名的 Mallat 算法：

College of Mathematics and Computer Science, Hebei University

ϕ j +1,n = ∑ < ϕ j +1,n ,ϕ j ,k > ϕ j ,k + ∑ < ϕ j +1,n ,ψ j ,k > ψ j ,k
k∈Z

k∈Z

< ϕ j +1, n ,ϕ j ,k >=

1

< ϕ j +1, n ,ψ j , k >=

ϕ j +1,n =

1

h

2
k∈Z

n−2k

ϕ j ,k +

2
1

hn − 2 k
g n−2k

2

1

g

2
k∈Z

ψ j ,k

n−2k

College of Mathematics and Computer Science, Hebei University

c j +1,n =< f (t ),ϕ j +1,n >
=

1

h

2
k∈Z

=

1

n−2k

h

2
k∈Z

< f (t ),ϕ j ,k > +

n − 2 k c j ,k

1

g

2
k∈Z

+

1

g

2
k∈Z

n− 2k d j ,k

n−2k

< f (t ),ψ j ,k

College of Mathematics and Computer Science, Hebei University

1 ）对原始信号作小波变换，将信号由空域变换到频

2 ）对小波系数做相应处理；
3 ）对处理后的小波系数做小波逆变换，还原原信号

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

 图像信号的小波分解：选择合适的小波及恰当的分

 对分解后的高频系数进行阈值量化：对于分解的每

 重构图像：根据小波分解后的第 N 层近似的低频系

College of Mathematics and Computer Science, Hebei University

College of Mathematics and Computer Science, Hebei University

[1] 唐远炎 , 王玲 . 小波分析与文本文字识别 , 科学出版社 ,2004
[2] 李弼程 , 彭天强 , 彭波 . 智能图像处理技术 , 电子工业出版社 ,
2004
[3] I. Daubechies, Ten Lectures on Wavelets. Philadelphia: SIAM,
1992.
[4] S. Mallat. A wavelet tour of signal processing. Academic Press,
USA, 1998
[5] S. Mallat, “A theory for multiresolution signal decomposition: The
wavelet representation,” IEEE Trans. Pattern Anal. Machine Intell.,
vol.11, pp. 674–693, 1989.
[6] W. Lawton. Tight frames of compactly supported wavelets, J.
Math. Phys., 31: 1898~1901, 1990.

College of Mathematics and Computer Science, Hebei University

[7] I. Daubechies, J. C. Lagarias. Two-scale difference operations I:
existence and glogal regularity of solutions, SIAM J. Math. Anal.,
22: 1388~1410,1991.
[8] D. Donoho, “De-noising by soft-thresholding,” IEEE Trans.
Inform.Theory, vol. 41, pp. 613–627, 1995.
[9] B.JAWERTH,etc. “an overview of wavelet based
multiresolution analyses,” SIAM REVIEW,vol. 36,No.33,pp.377412,September,1994.

College of Mathematics and Computer Science, Hebei University