opencv之局部方向模式(LDP)

人脸图像的局部方向模式基本思想是图像每个相素点都有一个局部3×3 邻域像素灰度值，该相素处于邻域中心。将该局部与3×3 邻域的灰度值与8 个Kirsch 模板卷积得到相应方向的边缘梯度值(i=0,1,…,7)，将边缘梯度值的绝对值 进行排序，求出第k大的值 并将大于等于 的 对应的第i位二进制数设置为1，剩余8-i位置为0，得到一个八位的二进制编码，然后根据不同位置进行加权求和，所得的十进制数即为该像素点的 LDP 特征值。八个方向的 Kirsch 模板(M0~M7)如图 1 所示，LDP 特征提取流程如图 2 所示。源码如下：完整的Qt工程下载：http://download.csdn.net/detail/j_d_c/9328865main.cpp

#include"ldp.h"

#include <iostream>

using namespace cv;

using namespace std;

int main(int argc,char *argv[])

{

int R=1;

Mat grayImg,ldpImg;

Mat pImg = imread("F:/MyQt/Gaussian/1.bmp",1);

grayImg.create(pImg.rows,pImg.cols,CV_8UC1);

cvtColor(pImg,grayImg,CV_RGB2GRAY);//转换成灰度图像

imshow("oringe",pImg);

LDP ldp;

ldp.ldppattern(grayImg,ldpImg,R);

namedWindow("ldpImg",WINDOW_AUTOSIZE);

imshow("ldpImg",ldpImg);

waitKey(0);

}

ldp.cpp

#include "ldp.h"

#include <qDebug>

#include <iostream>

#include <vector>

#include <opencv2/core/core.hpp>

#include <opencv2/highgui/highgui.hpp>

using namespace cv;

using namespace std;

typedef struct

{

int position;

long point_value;

} PAIR;

/*按照降序排列*/

bool operator<(const PAIR &x, const PAIR &y)

{

return x.point_value > y.point_value;

}

LDP::LDP()

{

}

void  LDP::ldpEdirect(Mat &Ein,Mat &Eout)//Kirsch卷积模板

{

Mat dst=Mat::zeros(3,3,CV_32F);

//kirsch

Mat M0 = (Mat_<float>(3,3) << -3,-3,5,-3,0,5,-3,-3,5);

Mat M1 = (Mat_<float>(3,3) << -3,5,5,-3,0,5,-3,-3,-3);

Mat M2 = (Mat_<float>(3,3) << 5,5,5,-3,0,-3,-3,-3,-3);

Mat M3 = (Mat_<float>(3,3) << 5,5,-3,5,0,-3,-3,-3,-3);

Mat M4 = (Mat_<float>(3,3) << 5,-3,-3,5,0,-3,5,-3,-3);

Mat M5 = (Mat_<float>(3,3) << -3,-3,-3,5,0,-3,5,5,-3);

Mat M6 = (Mat_<float>(3,3) << -3,-3,-3,-3,0,-3,5,5,5);

Mat M7 = (Mat_<float>(3,3) << -3,-3,-3,-3,0,5,-3,5,5);

filter2D(Ein,dst,dst.depth(),M0,Point(-1,-1));

Eout.ptr<float>(1)[2] = dst.ptr<float>(1)[1];

filter2D(Ein,dst,dst.depth(),M1,Point(-1,-1));

Eout.ptr<float>(0)[2] = dst.ptr<float>(1)[1];

filter2D(Ein,dst,dst.depth(),M2,Point(-1,-1));

Eout.ptr<float>(0)[1] = dst.ptr<float>(1)[1];

filter2D(Ein,dst,dst.depth(),M3,Point(-1,-1));

Eout.ptr<float>(0)[0] = dst.ptr<float>(1)[1];

filter2D(Ein,dst,dst.depth(),M4,Point(-1,-1));

Eout.ptr<float>(1)[0] = dst.ptr<float>(1)[1];

filter2D(Ein,dst,dst.depth(),M5,Point(-1,-1));

Eout.ptr<float>(2)[0] = dst.ptr<float>(1)[1];

filter2D(Ein,dst,dst.depth(),M6,Point(-1,-1));

Eout.ptr<float>(2)[1] = dst.ptr<float>(1)[1];

filter2D(Ein,dst,dst.depth(),M7,Point(-1,-1));

Eout.ptr<float>(2)[2] = dst.ptr<float>(1)[1];

}

void LDP::ldppattern(Mat &srcImg,Mat &ldpImg,int r)

{

int L = 2*r+1;

ldpImg.create(srcImg.rows,srcImg.cols,CV_8UC1);

Mat borderImg;

copyMakeBorder(srcImg,borderImg,L/2,L/2,L/2,L/2,BORDER_CONSTANT,Scalar::all(0));//卷积前对原图像上下左右进行扩展

Mat A,m;

A.create(L,L,CV_8U);

m = Mat::zeros(3,3,CV_32F);

int t[8];

for(int y=L/2;y<=borderImg.rows-L;y++)//行数变化

{ if(y>(borderImg.rows-L)) break;

for(int x=L/2;x<=borderImg.cols-L;x++)//列变化

{ if(x>(borderImg.cols-L)) break;//该语句是为了防止超过边界值

A = borderImg(Rect(x-L/2,y-L/2,L,L));

ldpEdirect(A,m);

t[0] = m.ptr<float>(1)[2];

t[1] = m.ptr<float>(0)[2];

t[2] = m.ptr<float>(0)[1];

t[3] = m.ptr<float>(0)[0];

t[4] = m.ptr<float>(1)[0];

t[5] = m.ptr<float>(2)[0];

t[6] = m.ptr<float>(2)[1];

t[7] = m.ptr<float>(2)[2];

PAIR pair[8];

vector<PAIR> vec;

int codebit[8]={0};

for(int i=0;i<8;i++)

{

pair[i].position = i;

pair[i].point_value = t[i];

vec.push_back(pair[i]);

}

sort(vec.begin(), vec.end());

// cout << "排序的结果为:" << endl;

//将最大的前三个值对应位置置1

for(vector<PAIR>::iterator it = vec.begin(); it < vec.begin()+3; ++it) {

codebit[it->position] = 1;

}

ldpImg.ptr<uchar>(y-L/2)[x-L/2] = codebit[7]*pow(2.0,7.0)+codebit[6]*pow(2.0,6.0)+codebit[5]*pow(2.0,5.0)+codebit[4]*pow(2.0,4.0)

+ codebit[3]*pow(2.0,3.0)+codebit[2]*pow(2.0,2.0)+codebit[1]*pow(2.0,1.0)+codebit[0]*pow(2.0,0.0);

}

}

}

运行结果：(1)卷积后取绝对值(1)卷积后不取绝对值