图像二值化算法-优快云博客

本文链接：https://blog.youkuaiyun.com/szfhy/article/details/49512269

一：大津阈值及其变种

该方法的核心就是计算背景和前景的最大类间方差，把类间方差最大的像素值作为全局二值化的阈值，计算费时。全局阈值，要根据适用场景选择性的应用。

变种：在[0 T] 内继续寻找ostu值，在[T 255]内继续寻找ostu值。

[plain] view plain copy

clear;
close all;
clc;
G = imread('input.JPG');
I = rgb2gray(G);
[m,n] = size(I);
Hist = zeros(255);
dHist = zeros(255);
variance = zeros(255);
PXD = 0;
for i = 1:m
for j = 1:n
Hist(uint8(I(i,j))) = Hist(uint8(I(i,j))) + 1;
end
end
for i = 1:255
dHist(i) = Hist(i)/(m*n);
end
for PXD = 1:255
w0 = 0;
w1 = 0;
g0 = 0;
g1 = 0;
for i = 1:PXD
g0 = g0 + i*dHist(i);
w0 = w0 + dHist(i);
end
for i = PXD+1 : 255
g1 = g1 + i*dHist(i);
w1 = w1 + dHist(i);
end
variance(PXD) = w0*w1*(g0 - g1)*(g0 - g1);
end
PXD = 1;
for i = 1:255
if variance(PXD) < variance(i)
PXD = i;
end
end
for i = 1:m
for j = 1:n
if I(i,j) > PXD
I(i,j) = 255;
else
I(i,j) = 0;
end
end
end
imagBW = I

二：Bernsen算法实现图像二值化，该方法属于局部方法

[html] view plain copy

clc;
clear;
close all;
I = imread('input.bmp');
w = 1;
T = 0;
max = 0;
min = 0;
[m,n] = size(I);
T = zeros(m - 2*w,n - 2*w);
for i = (w + 1):(m - w)
for j = (w + 1):(n - w)
max = uint8(I(i,j));
min = uint8(I(i,j));
for k = -w:w
for l = -w:w
if max < uint8(I(i + k,j + l))
max = uint8(I(i + k,j + l));
end
if min > uint8(I(i + k,j + l))
min = uint8(I(i + k,j + l));
end
end
end
T(i,j) = 0.5*(max + min);
end
end
for i = (w + 1):(m - w)
for j = (w + 1):(n - w)
if I(i,j) > T(i,j)
I(i,j) = uint8(255);
else
I(i,j) = uint8(0);
end
end
end
imshow(I);

三：Niblack 二值化

该方法仍然是一种二值化方法，阈值的计算公式是T = local_avg + k*v，其中local_avg表示以该像素点为中心的区域的平均灰度值，v是该区域的标准差，k是一个可以调节的系数

[html] view plain copy

clc;
clear;
close all;
I = imread('input.bmp');
I = rgb2gray(I);
w = 2;
max = 0;
min = 0;
[m,n] = size(I);
T = zeros(m ,n );
for i = (w + 1):(m - w)
for j = (w + 1):(n - w)
sum = 0;
for k = -w:w
for l = -w:w
sum = sum + uint32(I(i + k,j + l));
end
end
average = double(sum) /((2*w+1)*(2*w+1));
s = 0;
for k = -w:w
for l = -w:w
s = s + (uint32(I(i + k,j + l)) - average)*(uint32(I(i + k,j + l)) - average);
end
end
s= sqrt(double(s)/((2*w+1)*(2*w+1)));
T(i,j) = average + 0.2*s;
end
end
for i = 1:m
for j = 1:n
if I(i,j) > T(i,j)
I(i,j) = uint8(255);
else
I(i,j) = uint8(0);
end
end
end
imshow(I);

四：局部block分析二值化方法

Block Analysis大体思想是基于Niblack算法的，只不过把图像进行分块，然后对每一块计算阈值，T = local_avg + k*v，其中local_avg表示以该像素点为中心的区域的平均灰度值，v是该区域的标准差，k是一个可以调节的系数

[html] view plain copy

clc;
clear;
close all;
I = imread('100_1243.JPG');
I = rgb2gray(I);
[m,n] = size(I);
block = 10;
ver = floor(m/block);
hor = floor(n/block);
T = zeros(m,n);
for b_ver = 1:block
for b_hor = 1: block
t = 0;
for i = (ver * (b_ver - 1)+1) : (ver * b_ver)
for j = (hor * (b_hor - 1) + 1):(hor * b_hor)
t = t + uint32(I(i,j));
end
end
t = double(t)/(ver * hor);
std_deviation = 0;
for i = (ver * (b_ver - 1)+1) : (ver * b_ver)
for j = (hor * (b_hor - 1) + 1):(hor * b_hor)
std_deviation = std_deviation + (uint32(I(i,j)) - t)*(uint32(I(i,j)) - t);
end
end
std_deviation = sqrt(double(std_deviation)/(ver*hor));
thr = t + 0.2*std_deviation;
for i = (ver * (b_ver - 1)+1) : (ver * b_ver)
for j = (hor * (b_hor - 1) + 1):(hor * b_hor)
if I(i,j) > uint8(floor(thr))
T(i,j) = 255;
else
T(i,j) = 0;
end
end
end
end
end
imshow(T);

五：sauvola二值化算法

将图像分块，每块内的均值为local_avg. 阈值T=local_avg*(1+k*(s/R-1)) 其中k为可以调节的阈值，常取0.5，s为局部像素和均值的标准差，s=(pix_val - local_avg)^2/block_pixel_count. R = 128. 该方法在保留图像细节方面有很强的效果。