图像金字塔是图像处理和计算机视觉中常用到的概念,常常用于多尺度处理领域(multiscale processing),尤其早年的图像匹配、识别等算法中都用到了图像金字塔。
高斯金字塔(Gaussian pyramid)
下图为高斯金字塔的示意图,金字塔的底层为原始图像,每向上一层则是通过高斯滤波和1/2采样得到(去掉偶数行和列)。
我们可以使用如下Matlab代码来进行得到高斯金字塔:
function [ pyr ] = gaussian_pyramid( I,nlev )
%GAUSSIAN_PYRAMID Summary of this function goes here
% Detailed explanation goes here
pyr = cell(nlev,1);
pyr{1} = I;
filter = fspecial('gaussian');
for i=2:nlev
% gaussian filter
I = imfilter(I,filter,'symmetric');
% downsample
I = I(1:2:end,1:2:end);
pyr{i} = I;
end
end
下图就是生成的金字塔图像
高斯滤波器可以看做一个低通滤波器,那么每经过一次的高斯滤波,图像中仅能够保留某个频率值以下的频率部分,所以高斯金字塔也可以看做一个低通金字塔(每一级只保留某个频率以下的成分)。
拉普拉斯金字塔(Laplacian pyramid)
在进行高斯金字塔运算时,由于不断的进行高斯滤波和下采样,我们丢失了很多高频信号,而拉普拉斯金字塔的目的就是保存这些高频信号,保存这些高频信号所采用的方式就是保存差分图像。比如,拉普拉斯金字塔的第0层,就是原始图像和原始图像下采样(Reduce)后再次上采样(Expand)的图像的差值。
function pyr = laplacian_pyramid(I,nlev)
r = size(I,1);
c = size(I,2);
if ~exist('nlev')
% compute the highest possible pyramid
nlev = floor(log(min(r,c)) / log(2));
end
% recursively build pyramid
pyr = cell(nlev,1);
filter = pyramid_filter;
J = I;
for l = 1:nlev - 1
% apply low pass filter, and downsample
I = downsample(J,filter);
odd = 2*size(I) - size(J); % for each dimension, check if the upsampled version has to be odd
% in each level, store difference between image and upsampled low pass version
pyr{l} = J - upsample(I,odd,filter);
J = I; % continue with low pass image
end
pyr{nlev} = J; % the coarest level contains the residual low pass image
%下采样函数
function R = downsample(I, filter)
border_mode = 'symmetric';
% low pass, convolve with separable filter
R = imfilter(I,filter,border_mode); %horizontal
R = imfilter(R,filter',border_mode); %vertical
% decimate
r = size(I,1);
c = size(I,2);
R = R(1:2:r, 1:2:c, :);
%上采样函数
function R = upsample(I,odd,filter)
% increase resolution
I = padarray(I,[1 1 0],'replicate'); % pad the image with a 1-pixel border
r = 2*size(I,1);
c = 2*size(I,2);
k = size(I,3);
R = zeros(r,c,k);
R(1:2:r, 1:2:c, :) = 4*double(I); % increase size 2 times; the padding is now 2 pixels wide,注意这里要乘以4!
% interpolate, convolve with separable filter
R = imfilter(R,filter); %horizontal
R = imfilter(R,filter'); %vertical
% remove the border
R = R(3:r - 2 - odd(1), 3:c - 2 - odd(2), :);
%产生拉普拉斯滤波器
function f = pyramid_filter()
f = [.05, .25, .4, .25, .05]; % original [Burt and Adelson, 1983]
%f = [.0625, .25, .375, .25, .0625]; % binom-5
f = f'*f;
end
通过上面的代码,我们可以得到拉普拉斯金字塔如下所示。
由于拉普拉斯金字塔保留了高频信号,那么我们可以用它来重建原始图像。
function R = reconstruct_laplacian_pyramid(pyr)
r = size(pyr{1},1);
c = size(pyr{1},2);
nlev = length(pyr);
% start with low pass residual
R = pyr{nlev};
filter = pyramid_filter;
for l = nlev - 1 : -1 : 1
% upsample, and add to current level
odd = 2*size(R) - size(pyr{l});
R = pyr{l} + upsample(R,odd,filter);
%figure
%imshow(R,[]);
%imwrite(mat2gray(R),[num2str(l),'.jpg']);
end
需要注意的地方
为什么在处理高斯金字塔的时候需要采用滤波呢,直接下采样不可以吗?
如果把图像看做频率信号的话,直接进行下采样则会出现采样不足的情况,消除这种情况的方法是采用低通滤波器(高斯滤波器)对图像进行滤波,将采样不足的高频信号过滤掉,这样在进行下采样的时候就保证了不出现采样不足的情况。-
一般在图像处理中,将上面Matlab实现的下采样函数(包括高斯滤波和图像尺寸减半)这一部分叫做Reduce,将上面Matlab实现的上采样函数(包括高斯滤波和图像尺寸增加一倍)这部分叫做Expand,如果用数学方式表达的话,Expand函数如下:
注意前面需要乘以4。
拉普拉斯金字塔可以看做一个带通滤波器,在每一级都保留了图像某个频率值附近的成分。(这一点与高斯金字塔不同,高斯金字塔是低通金字塔)
两个低通滤波器的差值就构成了一个带通滤波器。
Python实现
下面是高斯金字塔和拉普拉斯金字塔的Opencv-Python实现
import cv2
import numpy as np
def gaussian_pyr(img,lev):
img = img.astype(np.float)
g_pyr = [img]
cur_g = img;
for index in range(lev):
cur_g = cv2.pyrDown(cur_g)
g_pyr.append(cur_g)
return g_pyr
def laplacian_pyr(img,lev):
img = img.astype(np.float)
g_pyr = gaussian_pyr(img,lev)
l_pyr = []
for index in range(lev):
cur_g = g_pyr[index]
next_g = cv2.pyrUp(g_pyr[index+1])
cur_l = cv2.subtract(cur_g,next_g)
l_pyr.append(cur_l)
l_pyr.append(g_pyr[-1])
return l_pyr
def lpyr_recons(l_pyr):
lev = len(l_pyr)
cur_l = l_pyr[-1]
for index in range(lev-2,-1,-1):
print(index)
cur_l = cv2.pyrUp(cur_l)
next_l = l_pyr[index]
cur_l = cur_l + next_l
return cur_l