对图像进行平移,只会改变频域空间的相位,不会改变幅值:
f
2
(
x
,
y
)
=
f
1
(
x
−
x
0
,
y
−
y
0
)
F
2
(
f
x
,
f
y
)
=
e
−
j
2
π
(
f
x
x
+
f
y
y
)
×
F
1
(
f
x
,
f
y
)
M
1
=
M
2
,
w
i
t
h
M
i
=
∣
F
i
∣
\begin{aligned} &f_2(x,y)=f_1(x-x_0,y-y_0)\\ &F_2(f_x,f_y)=e^{-j2\pi(f_xx+f_yy)}\times F_1(f_x,f_y)\\ &M_1=M_2,with\;M_i=|F_i| \end{aligned}
f2(x,y)=f1(x−x0,y−y0)F2(fx,fy)=e−j2π(fxx+fyy)×F1(fx,fy)M1=M2,withMi=∣Fi∣
对图像进行旋转,相当于频域空间极坐标下角度的平移:
f
2
(
x
,
y
)
=
f
1
(
x
c
o
s
θ
0
+
y
s
i
n
θ
0
,
−
x
s
i
n
θ
0
+
y
c
o
s
θ
0
)
M
2
(
f
x
,
f
y
)
=
M
1
(
f
x
c
o
s
θ
0
+
f
y
s
i
n
θ
0
,
−
f
x
s
i
n
θ
0
+
f
y
c
o
s
θ
0
)
M
1
(
ρ
,
θ
)
=
M
2
(
ρ
,
θ
−
θ
0
)
\begin{aligned} &f_2(x,y)=f_1(xcos\theta_0+ysin\theta_0,-xsin\theta_0+ycos\theta_0)\\ &M_2(f_x,f_y)=M_1(f_xcos\theta_0+f_ysin\theta_0,-f_xsin\theta_0+f_ycos\theta_0)\\ &M_1(\rho,\theta)=M_2(\rho,\theta-\theta_0) \end{aligned}
f2(x,y)=f1(xcosθ0+ysinθ0,−xsinθ0+ycosθ0)M2(fx,fy)=M1(fxcosθ0+fysinθ0,−fxsinθ0+fycosθ0)M1(ρ,θ)=M2(ρ,θ−θ0)
对图像的缩放,相当于频域空间极坐标下半径取对数的平移:
f
2
(
x
,
y
)
=
f
1
(
a
x
,
b
y
)
F
2
(
f
x
,
f
y
)
=
1
∣
a
b
∣
F
1
(
f
x
a
,
f
y
b
)
F
2
(
log
f
x
,
log
f
y
)
=
F
1
(
log
f
x
−
log
a
,
l
o
g
f
y
−
log
b
)
M
1
(
log
ρ
,
θ
)
=
M
2
(
log
ρ
−
log
a
,
θ
)
\begin{aligned} &f_2(x,y)=f_1(ax,by)\\ &F_2(f_x,f_y)=\frac{1}{|ab|}F_1(\frac{f_x}{a},\frac{f_y}{b})\\ &F_2(\log f_x,\log f_y)=F_1(\log f_x-\log a,logf_y-\log b)\\ &M_1(\log\rho,\theta)=M_2(\log\rho-\log a,\theta) \end{aligned}
f2(x,y)=f1(ax,by)F2(fx,fy)=∣ab∣1F1(afx,bfy)F2(logfx,logfy)=F1(logfx−loga,logfy−logb)M1(logρ,θ)=M2(logρ−loga,θ)
综上,若
f
2
f_2
f2相对于
f
1
f_1
f1平移任意尺度,缩放a倍,旋转角度
θ
0
\theta_0
θ0时,它们在频域的幅值有以下关系
M
1
(
ξ
,
θ
)
=
M
2
(
ξ
−
d
,
θ
−
θ
0
)
w
i
t
h
ξ
=
log
ρ
,
d
=
log
a
\begin{aligned} &M_1(\xi,\theta) = M_2(\xi-d,\theta-\theta_0)\\ &with\;\;\xi=\log\rho,d=\log a \end{aligned}
M1(ξ,θ)=M2(ξ−d,θ−θ0)withξ=logρ,d=loga
在这个基础上,猜想若将
M
1
M_1
M1和
M
2
M_2
M2以对数极坐标表示,再对它们做傅里叶变换,
M
1
M_1
M1和
M
2
M_2
M2傅里叶变换幅值相等,即
∣
F
(
M
1
)
∣
=
∣
F
(
M
2
)
∣
|\mathcal{F}(M_1)|=|\mathcal{F}(M_2)|
∣F(M1)∣=∣F(M2)∣
记作
M
1
=
M
2
\mathcal{M}_1=\mathcal{M}_2
M1=M2
这相当于即使
f
2
f_2
f2相对于
f
1
f_1
f1做了一系列仿射变换,但是仍可通过一系列变换从两张图像中提取出一种相等的特征。