引言：
LBP指局部二值模式，英文全称：Local Binary Pattern，是一种用来描述图像局部特征的算子，LBP特征具有灰度不变性和旋转不变性等显著优点。它是由T. Ojala, M.Pietikinen, 和 D. Harwood [1][2]在1994年提出，由于LBP特征计算简单、效果较好，因此LBP特征在计算机视觉的许多领域都得到了广泛的应用，LBP特征比较出名的应用是用在人脸识别和目标检测中，在计算机视觉开源库Opencv中有使用LBP特征进行人脸识别的接口，也有用LBP特征训练目标检测分类器的方法，Opencv实现了LBP特征的计算，但没有提供一个单独的计算LBP特征的接口。
LBP特征的原理：
LBP特征原理及代码实现 - SnailTyan - CSDN博客 https://blog.csdn.net/quincuntial/article/details/50541815
1.原始的LBP特征描述及计算方法
原始的LBP算子定义在像素3*3的邻域内，以邻域中心像素为阈值，相邻的8个像素的灰度值与邻域中心的像素值进行比较，若周围像素大于中心像素值，则该像素点的位置被标记为1，否则为0。这样，3*3邻域内的8个点经过比较可产生8位二进制数，将这8位二进制数依次排列形成一个二进制数字，这个二进制数字就是中心像素的LBP值，LBP值共有2^8种可能，因此LBP值有256种。中心像素的LBP值反映了该像素周围区域的纹理信息。
备注：计算LBP特征的图像必须是灰度图，如果是彩色图，需要先转换成灰度图。
上述过程用图像表示为：
代码实现：
#include "pch.h"
#include <iostream>
#include "opencv2/opencv.hpp"
#include <opencv2/xfeatures2d.hpp>
using namespace std;
using namespace cv;
using namespace cv::xfeatures2d;
int main()
{
//待检测图像
Mat src = imread("F:\\visual studio\\Image\\women1.jpg");
if (src.empty())
{
cout << "Can't load the image" << endl;
return -1;
}
imshow("src", src);
//转化为灰度图
Mat graysrc;
cvtColor(src, graysrc, COLOR_BGR2GRAY);
int width = graysrc.cols;
int height = graysrc.rows;
Mat lbpImage = Mat::zeros(height -2,width-2 , CV_8UC1);
for (int row = 1; row < height - 1; row++)
{
for (int col = 1; col < width - 1; col++)
{
uchar centerVal = graysrc.at<uchar>(row, col);//获取中心像素值
//计算LBP值
uchar code = 0;
code |= ((graysrc.at<uchar>(row - 1, col - 1) > centerVal) << 7);
code |= ((graysrc.at<uchar>(row - 1, col ) > centerVal) << 6);
code |= ((graysrc.at<uchar>(row - 1, col + 1) > centerVal) << 5);
code |= ((graysrc.at<uchar>(row , col + 1) > centerVal) << 4);
code |= ((graysrc.at<uchar>(row + 1, col + 1) > centerVal) << 3);
code |= ((graysrc.at<uchar>(row + 1, col) > centerVal) << 2);
code |= ((graysrc.at<uchar>(row + 1, col - 1) > centerVal) << 1);
code |= ((graysrc.at<uchar>(row , col - 1) > centerVal) << 0);
lbpImage.at<uchar>(row - 1, col - 1) = code;
}
}
imshow("LBP Image", lbpImage);
waitKey(0);
}效果展示：
2.LBP特征的改进版本
在原始的LBP特征提出以后，研究人员对LBP特征进行了很多的改进，因此产生了许多LBP的改进版本。
2.1圆形LBP特征(Circular LBP or Extended LBP)
由于原始LBP特征使用的是固定邻域内的灰度值，因此当图像的尺度发生变化时，LBP特征的编码将会发生错误，LBP特征将不能正确的反映像素点周围的纹理信息，因此研究人员对其进行了改进。基本的 LBP 算子的最大缺陷在于它只覆盖了一个固定半径范围内的小区域，这显然不能满足不同尺寸和频率纹理的需要。为了适应不同尺度的纹理特征，并达到灰度和旋转不变性的要求，Ojala 等对 LBP 算子进行了改进，将 3×3 邻域扩展到任意邻域，并用圆形邻域代替了正方形邻域，改进后的 LBP 算子允许在半径为 R 的圆形邻域内有任意多个像素点。从而得到了诸如半径为R的圆形区域内含有P个采样点的LBP算子：
这种LBP特征叫做Extended LBP，也叫Circular LBP。使用可变半径的圆对近邻像素进行编码，可以得到如下的近邻：
对于给定中心点(xc,yc)，其邻域像素位置为(xp,yp)，p∈P，其采样点(xp,yp)用如下公式计算：
R是采样半径，p是第p个采样点，P是采样数目。由于计算的值可能不是整数，即计算出来的点不在图像上，我们使用计算出来的点的插值点。目的的插值方法有很多，Opencv使用的是双线性插值，双线性插值的公式如下：
代码展示：
#include "pch.h"
#include <iostream>
#include "opencv2/opencv.hpp"
#include <opencv2/xfeatures2d.hpp>
using namespace std;
using namespace cv;
using namespace cv::xfeatures2d;
int radius = 2;
int maxVal = 10;
void ELBP_demo(int pos, void* data);
Mat graysrc;
int main()
{
//待检测图像
Mat src = imread("F:\\visual studio\\Image\\women1.jpg");
if (src.empty())
{
cout << "Can't load the image" << endl;
return -1;
}
imshow("src", src);
//转化为灰度图
cvtColor(src, graysrc, COLOR_BGR2GRAY);
namedWindow("LBP Image", WINDOW_AUTOSIZE);
createTrackbar("radius", "LBP Image", &radius, maxVal, ELBP_demo);
ELBP_demo(0, 0);
waitKey(0);
}
void ELBP_demo(int pos, void* data)
{
int height = graysrc.rows;
int width = graysrc.cols;
int offset = 2 * radius;
Mat elbpImage = Mat::zeros(height - offset, width - offset, CV_8UC1);
int neighbours = 8; //理论上应根据半径变化，这里为了简便固定为8
//圆形LBP特征计算，这种方法适于理解，但在效率上存在问题
for (int row = radius; row < height - radius; row++)
{
for (int col = radius; col < width - radius; col++)
{
//获取中心像素值
uchar centerVal = graysrc.at<uchar>(row, col);
uchar code = 0;
for (int i = 0; i < neighbours; i++)
{
//获取邻点坐标 这个地方可以优化 不必每次都进行计算 radius*cos、radius*sin
float x = col + static_cast<float>(radius*cos(2.0*CV_PI*i / neighbours));
float y = row + static_cast<float>(radius*sin(2.0*CV_PI*i / neighbours));
//取整并根据取整结果处的值进行双线性插值,得到第k个采样点的灰度值
//1.分别对x,y进行上下取整
int x1 = static_cast<int>(floor(x));
int y1 = static_cast<int>(floor(y));
int x2 = static_cast<int>(ceil(x));
int y2 = static_cast<int>(ceil(y));
//2.计算四个点(x1,y1),(x1,y2),(x2,y1),(x2,y2)的权重
float u = x - x1;
float v = y - y1;
float w1 = (1 - u)*(1 - v);
float w2 = u*(1 - v);
float w3 = (1 - u)*v;
float w4 = u * v;
//进行双线性插值
float neighbour = graysrc.at<uchar>(y1, x1)*w1 + graysrc.at<uchar>(y2, x1)*w2 +
graysrc.at<uchar>(y1, x2)*w3 + graysrc.at<uchar>(y2, x2)*w4;
//通过与中心元素值进行比较 获得LBP值
code |= ((neighbour > centerVal) << (neighbours - i - 1));
}
elbpImage.at<uchar>(row - radius, col - radius) = code;
}
}
Mat roi = elbpImage(Range(300, 325), Range(200, 205));
cout << roi << endl;
imshow("LBP Image", elbpImage);
}效率改进版：
#include "pch.h"
#include <iostream>
#include "opencv2/opencv.hpp"
#include <opencv2/xfeatures2d.hpp>
using namespace std;
using namespace cv;
using namespace cv::xfeatures2d;
int radius = 2;
int maxVal = 10;
void ELBP_demo(int pos, void* data);
Mat graysrc;
int main()
{
//待检测图像
Mat src = imread("F:\\visual studio\\Image\\women1.jpg");
if (src.empty())
{
cout << "Can't load the image" << endl;
return -1;
}
imshow("src", src);
//转化为灰度图
cvtColor(src, graysrc, COLOR_BGR2GRAY);
namedWindow("LBP Image1", WINDOW_AUTOSIZE);
createTrackbar("radius", "LBP Image1", &radius, maxVal, ELBP_demo);
ELBP_demo(0, 0);
waitKey(0);
}
void ELBP_demo(int pos, void* data)
{
int height = graysrc.rows;
int width = graysrc.cols;
int offset = 2 * radius;
Mat elbpImage = Mat::zeros(height - offset, width - offset, CV_8UC1);
int neighbours = 8; //理论上应根据半径变化，这里为了简便固定为8
//圆形LBP特征计算，效率优化版本
for (int i = 0; i < neighbours; i++)
{
//计算采样点对于中心点坐标的偏移量rx,ry;
float rx = static_cast<float>(radius*cos(2.0 * CV_PI*i / neighbours));
float ry = -static_cast<float>(radius*sin(2.0 * CV_PI*i / neighbours));
//为双线性插值做准备
//1.对采样点偏移量分别进行上下取整
int x1 = static_cast<int>(floor(rx));
int y1 = static_cast<int>(floor(ry));
int x2 = static_cast<int>(ceil(rx));
int y2 = static_cast<int>(ceil(ry));
//计算权重
float u = rx - x1;
float v = ry - y1;
//if (u < std::numeric_limits<float>::epsilon())
// u = 0;
//if (v < std::numeric_limits<float>::epsilon())
// v = 0;
float w1 = (1 - u)*(1 - v);
float w2 = u * (1 - v);
float w3 = (1 - u) * v;
float w4 = u * v;
//循环处理每个像素
for (int row = radius; row < height - radius; row++)
{
for (int col = radius; col < width - radius; col++)
{
uchar centerval = graysrc.at<uchar>(row, col);
float neighbour = graysrc.at<uchar>(row + y1, col + x1) * w1 +
graysrc.at<uchar>(row + y2, col + x1) * w2 +
graysrc.at<uchar>(row + y1, col + x2) * w3 +
graysrc.at<uchar>(row + y2, col + x2) * w4;
elbpImage.at<uchar>(row - radius, col - radius) |= ((neighbour > centerval) << (neighbours - i - 1));
}
}
}
Mat roi = elbpImage(Range(300, 325), Range(200, 205));
cout << roi << endl;
imshow("lbp image1", elbpImage);
}效果展示：
使用上述同一种方法的不同实现，得到的2个LBP特征图像不同。根据多次调试发现，代码逻辑本身没有问题，只是涉及到精度、有效数字之类的问题，下面进行修改，主要修改了采样点与中心点处值比较时采用的阈值，由于计算精度的影响，对于同一组中心点和采样点，2种方法获得的数值有及其细微的差别，但是比较结果有时却不一致。因此我们调大比较时的阈值，可以忽略其细微差别。
代码修改：
#include "pch.h"
#include <iostream>
#include "opencv2/opencv.hpp"
#include <opencv2/xfeatures2d.hpp>
#include <bitset>
using namespace std;
using namespace cv;
using namespace cv::xfeatures2d;
int radius = 2;
int maxVal = 10;
void ELBP_demo(int pos, void* data);
Mat graysrc;
int main()
{
//待检测图像
Mat src = imread("F:\\visual studio\\Image\\women1.jpg");
if (src.empty())
{
cout << "Can't load the image" << endl;
return -1;
}
imshow("src", src);
//转化为灰度图
cvtColor(src, graysrc, COLOR_BGR2GRAY);
namedWindow("LBP Image", WINDOW_AUTOSIZE);
createTrackbar("radius", "LBP Image", &radius, maxVal, ELBP_demo);
ELBP_demo(0, 0);
waitKey(0);
}
void ELBP_demo(int pos, void* data)
{
int height = graysrc.rows;
int width = graysrc.cols;
int offset = 2 * radius;
Mat elbpImage = Mat::zeros(height - offset, width - offset, CV_8UC1);
int neighbours = 8; //理论上应根据半径变化，这里为了简便固定为8
//圆形LBP特征计算，这种方法适于理解，但在效率上存在问题，声明时默认neighbors=8
for (int row = radius; row < height - radius; row++)
{
for (int col = radius; col < width - radius; col++)
{
//获取中心像素值
uchar centerVal = graysrc.at<uchar>(row, col);
uchar code = 0;
for (int i = 0; i < neighbours; i++)
{
//获取邻点坐标 这个地方可以优化 不必每次都进行计算 radius*cos、radius*sin
float x = col + static_cast<float>(radius*cos(2.0*CV_PI*i / neighbours));
float y = row - static_cast<float>(radius*sin(2.0*CV_PI*i / neighbours));
//取整并根据取整结果处的值进行双线性插值,得到第k个采样点的灰度值
//1.分别对x,y进行上下取整
int x1 = static_cast<int>(floor(x));
int y1 = static_cast<int>(floor(y));
int x2 = static_cast<int>(ceil(x));
int y2 = static_cast<int>(ceil(y));
//2.计算四个点(x1,y1),(x1,y2),(x2,y1),(x2,y2)的权重
float u = x - x1;
float v = y - y1;
float w1 = (1 - u)*(1 - v);
float w2 = u*(1 - v);
float w3 = (1 - u)*v;
float w4 = u * v;
//进行双线性插值
float neighbour = static_cast<float>(graysrc.at<uchar>(y1, x1)*w1 + graysrc.at<uchar>(y2, x1)*w2 +
graysrc.at<uchar>(y1, x2)*w3 + graysrc.at<uchar>(y2, x2)*w4);
//通过与中心元素值进行比较 获得LBP值
code |= (((neighbour > centerVal)&& abs(neighbour - centerVal) > /*std::numeric_limits<float>::epsilon()*/0.0001) << (neighbours - i - 1));
//if (col == (368 + radius) && row == (140 + radius))
//{
// cout << "1 = " << (int)graysrc.at<uchar>(y1, x1) << " w1 = " << w1 << endl;
// cout << "2 = " << (int)graysrc.at<uchar>(y2, x1) << " w2 = " << w2 << endl;
// cout << "3 = " << (int)graysrc.at<uchar>(y1, x2) << " w3 = " << w3 << endl;
// cout << "4 = " << (int)graysrc.at<uchar>(y2, x2) << " w4 = " << w4 << endl;
// cout << "W = " << w1 + w2 + w3 + w4 << endl;
// cout << "(y1,x1) = " << y1 << "," << x1 << endl;
// cout << "(y2,x2) = " << y2 << "," << x2 << endl;
// cout << "u = " << u << " v= " << v << endl;
// cout << "neighbour" << i << " = " << neighbour << endl;
// cout << "center = " << (int )centerVal << endl;
// cout << bitset<8>(code) << endl;
//}
}
elbpImage.at<uchar>(row - radius, col - radius) = code;
}
}
//Mat roi = elbpImage(Range(300, 325), Range(200, 205));
//cout << roi << endl;
imshow("LBP Image", elbpImage);
}
效率优化版本：
#include "pch.h"
#include <iostream>
#include "opencv2/opencv.hpp"
#include <opencv2/xfeatures2d.hpp>
#include <bitset>
using namespace std;
using namespace cv;
using namespace cv::xfeatures2d;
int radius = 2;
int maxVal = 10;
void ELBP_demo(int pos, void* data);
void mouseCallback(int event, int x, int y, int flags, void* data);
Mat graysrc;
Mat elbpImage;
int main()
{
//待检测图像
Mat src = imread("F:\\visual studio\\Image\\women1.jpg");
if (src.empty())
{
cout << "Can't load the image" << endl;
return -1;
}
imshow("src", src);
//转化为灰度图
cvtColor(src, graysrc, COLOR_BGR2GRAY);
namedWindow("LBP Image1", WINDOW_AUTOSIZE);
createTrackbar("radius", "LBP Image1", &radius, maxVal, ELBP_demo);
setMouseCallback("LBP Image1", mouseCallback, 0);
ELBP_demo(0, 0);
waitKey(0);
}
void ELBP_demo(int pos, void* data)
{
int height = graysrc.rows;
int width = graysrc.cols;
int offset = 2 * radius;
elbpImage = Mat::zeros(height - offset, width - offset, CV_8UC1);
int neighbours = 8; //理论上应根据半径变化，这里为了简便固定为8
//圆形LBP特征计算，效率优化版本
for (int i = 0; i < neighbours; i++)
{
//计算采样点对于中心点坐标的偏移量rx,ry;
float rx = static_cast<float>(radius*cos(2.0 * CV_PI*i / neighbours));
float ry = -static_cast<float>(radius*sin(2.0 * CV_PI*i / neighbours));
if (abs(ry) < std::numeric_limits<float>::epsilon())
ry = 0;
if (abs(rx) < std::numeric_limits<float>::epsilon())
rx = 0;
//为双线性插值做准备
//1.对采样点偏移量分别进行上下取整
int x1 = static_cast<int>(floor(rx));
int y1 = static_cast<int>(floor(ry));
int x2 = static_cast<int>(ceil(rx));
int y2 = static_cast<int>(ceil(ry));
//计算权重
float u = rx - x1;
double v = ry - y1;
if (u < std::numeric_limits<float>::epsilon())
u = 0;
if (v < std::numeric_limits<float>::epsilon())
v = 0;
//if (i == 6)
//{
// cout << "rx = " << rx << "ry = " << ry << endl;
// cout << "x1 = " << x1 << " y1 = " << y1 << endl;
// cout << "x2 = " << x2 << " y2 = " << y2<< endl;
// cout << "u = " << u<< " v = " << v << endl;
//}
float w1 = (1 - u)*(1 - v);
float w2 = u * (1 - v);
float w3 = (1 - u) * v;
float w4 = u * v;
//循环处理每个像素
for (int row = radius; row < height - radius; row++)
{
for (int col = radius; col < width - radius; col++)
{
uchar centerval = graysrc.at<uchar>(row, col);
float neighbour = graysrc.at<uchar>(row + y1, col + x1) * w1 +
graysrc.at<uchar>(row + y2, col + x1) * w2 +
graysrc.at<uchar>(row + y1, col + x2) * w3 +
graysrc.at<uchar>(row + y2, col + x2) * w4;
elbpImage.at<uchar>(row - radius, col - radius) |= ((neighbour > centerval && abs(neighbour-centerval) >/*std::numeric_limits<float>::epsilon()*/0.0001) << (neighbours - i - 1));
// 取出一个点的值进行调试
//if (col == (368 + radius) && row == (140 + radius))
//{
// cout << "1 = " << (int)graysrc.at<uchar>(row + y1, col + x1) << " w1 = " << w1 << endl;
// cout << "2 = " << (int)graysrc.at<uchar>(row + y2, col + x1) << " w2 = " << w2 << endl;
// cout << "3 = " << (int)graysrc.at<uchar>(row + y1, col + x2) << " w3 = " << w3 << endl;
// cout << "4 = " << (int)graysrc.at<uchar>(row + y2, col + x2) << " w4 = " << w4 << endl;
// cout << "(y1,x1) = " << row + y1 << "," << col + x1 << endl;
// cout << "(y2,x2) = " << row + y2 << "," << col + x2 << endl;
// cout << "u = " << u << " v = " << v << endl;
// cout << "W = " << w1 + w2 + w3 + w4 << endl;
// cout << "neighbour" << i << " = " << neighbour << endl;
// cout << "center = " << (int)centerval << endl;
// cout << bitset<8>(elbpImage.at<uchar>(368, 140)) << endl;
//}
}
}
}
//Mat roi = elbpImage(Range(300, 325), Range(200, 205));
//cout << roi << endl;
imshow("LBP Image1", elbpImage);
}
void mouseCallback(int event, int x, int y, int flags, void* data)
{
if (flags == EVENT_FLAG_LBUTTON)
{
cout << "X = " << x << endl << " Y = " << y << endl;
cout << (int )elbpImage.at<uchar>(x, y) << endl;
}
}效果展示：
此次获得的结果一致性更高:
5543
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
