Gaussian Pyramid vs Laplacian Pyramid
A Laplacian Pyramid is a linear invertible image representation consisting of a set of band-pass images, spaced an octave apart, plus a low-frequency residual. Formally, let
d
(
⋅
)
d(\cdot)
d(⋅) be a downsampling operation which blurs and decimates a
j
×
j
j \times j
j×j image
I
I
I, so that
d
(
I
)
d(I)
d(I) is a new image of size
j
2
×
j
2
\frac{j}{2} \times \frac{j}{2}
2j×2j. Also, let
u
(
⋅
)
u(\cdot)
u(⋅) be an upsampling operator which smooths and expand
I
I
I to be twice the size, so
u
(
I
)
u(I)
u(I) is a new image of size
2
j
×
2
j
2j \times 2j
2j×2j. We first build a Gaussian pyramid
G
(
I
)
=
[
I
0
,
I
1
,
.
.
.
I
K
]
,
G(I)=[I_0,I_1,...I_K],
G(I)=[I0,I1,...IK], where
I
0
=
I
I_0=I
I0=I and
I
k
I_k
Ik is
k
k
k repeated applications of
d
(
⋅
)
d(\cdot)
d(⋅) to
I
I
I.
K
K
K is the number of levels in the pyramid, selected so that the final level has very small spatial extent (
≤
8
×
8
\le 8 \times 8
≤8×8 pixels).
The coefficients h k h_k hk at each level k k k of the Laplacian pyramid L ( I ) L(I) L(I) are constructed by taking the difference between adjacent levels in the Gaussian pyramid, upsampling the smaller one with u ( ⋅ ) u(\cdot) u(⋅), so that the sizes are compatible:
h k = L k ( I ) = G k ( I ) − u ( G k + 1 ( I ) ) = I k − u ( I k + 1 ) h_k = L_k(I)=G_k(I)-u(G_{k+1}(I))=I_k - u( I_{k+1}) hk=Lk(I)=Gk(I)−u(Gk+1(I))=Ik−u(Ik+1)
Intuitively, each level captures images structure present at a particular scale. The final level of the Laplacian pyramid h K h_K hK is not a difference image, but a low-frequency residual equal to the final Gaussian pyramid level, i.e. h K = I K h_K=IK hK=IK. Reconstruction from a Laplacian pyramid coefficients [ h 1 , . . . h K ] [h_1,...h_K] [h1,...hK] is performed using the backward recurrence:
I k = u ( I k + 1 ) + h k I_k=u(I_{k+1})+h_k Ik=u(Ik+1)+hk
which is started with I K = h K I_K=h_K IK=hK and the reconstructed image being I = I o I=I_o I=Io. In other words, starting at the coarsest level, we repeatedly upsample and add the difference image h at the next finer level until we get back to the full resolution image.
补充
[OpenCV / The Laplacian pyramid] 浅析拉普拉斯金字塔 / 利用拉普拉斯完成采样的恢复 学习笔记