opencv学习+踩坑
环境
- ubuntu 19.04
- vscode 1.37.0
- opencv 3.4.7
- cmake 3.13.4
拜一下julao的数字图像处理提纲
https://bitlecture.github.io/notes/数字图像处理/
然后开始跟着毛星云的blog跑demo来学opencv
实际上如果论实用性的话,以下的系列blog可能还会更好一些?
https://blog.csdn.net/morewindows/article/category/1291764
https://www.cnblogs.com/long5683/p/10094122.html
实际上学一会就会发现RM里面使用到的视觉(仅仅看这篇开源的话)并不困难
https://blog.csdn.net/u010750137/article/details/91344986
https://blog.csdn.net/qq_31669419/article/details/53053321
反而是去年的神符里面涉及到了一些类似机器学习一样的东西,更有研究的空间在
那么让我们开始视觉学习之路
文件读取和输出
https://blog.csdn.net/poem_qianmo/article/details/20537737
定义图像
Mat image = imread("Filename");
namedWindow("Windowname");
imshow("Windowname",image);
需要注意的是图片要放到build的文件夹里面,如果没能成功imread的话,会报错——
error: (-215:Assertion failed) size.width>0 && size.height>0 in function 'imshow'
视频读取
VideoCapture cap;
cap.open("Filename");
打开摄像头
cap.open(0);
检测是否读取到的方法:
//方法1
if(!image.data){printf("未能读取")};
//方法2
if(image.empty()){printf("未能读取")};
划定特定区域(ROI)
https://blog.csdn.net/poem_qianmo/article/details/20911629
Mat imageROI;
//方法一
imageROI= srcImage4(Rect(200,250,logoImage.cols,logoImage.rows));
//方法二
imageROI= srcImage4(Range(250,250+logoImage.rows),Range(200,200+logoImage.cols));
图像变换应该也挺重要的
https://blog.csdn.net/xiaowei_cqu/article/details/7616044
图像线性混合
使用addWeighted可以直接混合两张图片,
int main()
{
double alphavalue = 0.5;
double betavalue;
Mat satori = imread("satori.jpg");
Mat name = imread("name.png");
if(satori.empty()){cout << "未能成功读取图片satori" << endl;exit;};
if(name.empty()){cout << "未能成功读取图片satori2" << endl;exit;};
betavalue = 1 - alphavalue;
//在satori上划出ROI
Mat ROI = satori(Rect(0,0,name.cols,name.rows));
//将划出了ROI的satori和name做合并
addWeighted(ROI,alphavalue,name,betavalue,0.,ROI);
namedWindow("混合效果");
imshow("混合效果",satori);
waitKey();
return 0;
}
分离/合并颜色通道
split()/merge()
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int main()
{
Mat satori = imread("satori.jpg");
Mat name = imread("name.png",0);
vector<Mat>channels;
Mat blue_channel;
if(satori.empty()){cout << "未能成功读取图片satori" << endl;exit;};
if(name.empty()){cout << "未能成功读取图片satori2" << endl;exit;};
//分割成几个颜色通道
split(satori,channels);
blue_channel = channels.at(0);
addWeighted(blue_channel(Rect(0,0,name.cols,name.rows)),1.0,name,0.5,0,blue_channel(Rect(0,0,name.cols,name.rows)));
//混合通道
merge(channels,satori);
namedWindow("混合效果");
imshow("混合效果",satori);
waitKey();
return 0;
}
从颜色通道的角度来说,可以扒掉另外两个通道,只留一个通道做合成来形成单色图片
opencv里面可以设置图片类型,比如CV_8UC1,就是unsigned int8+channel_1,所以这里的操作还是挺简单的,就是用black来取代掉另外两个通道(black意味着灰度值为0),把它给另外两个通道即可
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int main()
{
Mat satori = imread("satori.jpg");
vector<Mat> channels(satori.channels());
vector<Mat> channels_mix(satori.channels());
Mat mixed;
if(satori.empty()){cout << "未能成功读取图片satori" << endl;exit;}
int w = satori.cols;
int h = satori.rows;
split(satori,channels);
Mat black;
black.create(h,w,CV_8UC1);
black = Scalar(0);
channels_mix[0] = channels[0];
channels_mix[1] = black;
channels_mix[2] = black;
merge(channels_mix,mixed);
imshow("mixed",mixed);
waitKey();
return 0;
}
颜色通道和ROI,以及线性混合的内容再补充一个画矩形?
https://blog.csdn.net/wc781708249/article/details/78518447
会用rectangle就行了
这个是边缘查找,感觉也是个有意思的demo
https://www.cnblogs.com/skyfsm/p/6890863.html
还是继续跑demo,拿小圆当看板是有点东西的,tracebar的话,应该相当于提供了类似嵌入式开发中的在线debug一样的功能?
https://blog.csdn.net/poem_qianmo/article/details/21479533
关于向量这个数据类型
https://www.cnblogs.com/mr-wid/archive/2013/01/22/2871105.html
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
Mat satori;
int threval = 160;
static void trace_bar(int,void*)
{
Mat image = threval > 128? (satori < threval) : (satori > threval);
vector<vector<Point>> contours;
vector<Vec4i> hierarchy;
findContours(image,contours,hierarchy,CV_RETR_CCOMP,CV_CHAIN_APPROX_SIMPLE);
Mat dst = Mat::zeros(satori.size(),CV_8UC3);
if(!contours.empty() && !hierarchy.empty())
{
for (int i = 0; i >=0; i=hierarchy[i][0])
{
Scalar color((rand()&255),(rand()&255),(rand()&255));
drawContours(dst,contours,i,color,CV_FILLED,8,hierarchy);
}
}
imshow("satori",dst);
}
int main()
{
satori = imread("satori.jpg",0);
if(satori.empty()){cout << "未能成功读取图片satori" << endl;exit;}
namedWindow("satori");
createTrackbar("treashould","satori",&threval,255,trace_bar);
trace_bar(threval,0);
waitKey();
return 0;
}
所以实际上主要就是找轮廓+填色,有点意思
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int Contrast,Bright;
Mat srcImage,dstImage;
static void trace_bar(int,void *)
{
for (int i = 0; i < srcImage.cols; i++)
{
for (int j = 0; j < srcImage.rows; j++)
{
for (int k = 0; k < 3; k++)
{
dstImage.at<Vec3b>(j,i)[k] = saturate_cast<uchar>((Contrast*0.01)*srcImage.at<Vec3b>(j,i)[k] + Bright);
}
}
}
imshow("satori",dstImage);
};
int main()
{
srcImage = imread("satori.jpg");
if(srcImage.empty()){cout << "未能成功读取图片satori" << endl;return -1;}
dstImage = Mat::zeros(srcImage.size(),srcImage.type());
Contrast = 80;
Bright = 80;
namedWindow("satori");
createTrackbar("contrast","satori",&Contrast,255,trace_bar);
createTrackbar("bright","satori",&Bright,255,trace_bar);
trace_bar(Contrast,0);
trace_bar(Bright,0);
waitKey();
return 0;
}
这个demo主要是试了一下针对像素调bright和contrast,我没想到居然就是这么简单的线性运算关系,另外就是对单独的像素操作
其实我们已经看出来了,图片的一种表现方式就是每个Image.at
到滤波了
https://blog.csdn.net/poem_qianmo/article/details/22745559
https://blog.csdn.net/xiaowei_cqu/article/details/7785365
方框滤波——boxblur函数
均值滤波(邻域平均滤波)——blur函数
高斯滤波——GaussianBlur函数
中值滤波——medianBlur函数
双边滤波——bilateralFilter函数
https://wenku.baidu.com/view/f55e1bc6f90f76c661371ac5.html
二维卷积挺有用的,包括之后做边沿检测用的sobel算子等
https://blog.csdn.net/dang_boy/article/details/76150067
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int main()
{
Mat srcImage = imread("satori.jpg");
if(srcImage.empty()){cout << "未能成功读取图片satori" << endl;return -1;}
Mat dstImage1,dstImage2,dstImage3,dstImage4,dstImage5;
dstImage1 = srcImage.clone();
dstImage2 = srcImage.clone();
dstImage3 = srcImage.clone();
dstImage4 = srcImage.clone();
dstImage5 = srcImage.clone();
imshow("原图",srcImage);
boxFilter(srcImage,dstImage1,-1,Size(5,5));
imshow("方框滤波",dstImage1);
blur(srcImage,dstImage2,Size(5,5));
imshow("均值滤波",dstImage2);
GaussianBlur(srcImage,dstImage3,Size(3,3),0,0);
imshow("高斯滤波",dstImage3);
medianBlur(srcImage,dstImage4,5);
imshow("中值滤波",dstImage4);
bilateralFilter(srcImage,dstImage5,25,25*2,25/2);
imshow("双边滤波",dstImage5);
waitKey();
destroyAllWindows();
return 0;
}
简单的滤波跑了一下而已,Size(w,h)规定了卷积核的大小,卷积核的大小会影响模糊的效果
然后是非线性滤波
中值滤波和双线性滤波
双线性滤波的效果非常神奇,把原图上一些类似于陈旧的纹理一样的效果给修没了,非常6p(磨皮?)
膨胀腐蚀
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int Elem_Size = 3;
int value1;
Mat srcImage,dstImage;
static void tracebar(int,void*)
{
Mat element = getStructuringElement(MORPH_RECT,Size(2*Elem_Size+1,2*Elem_Size+1),Point(Elem_Size,Elem_Size));
if(value1 == 0)
{
erode(srcImage,dstImage,element);
}
else
{
dilate(srcImage,dstImage,element);
}
imshow("satori",dstImage);
}
int main()
{
srcImage = imread("satori.jpg");
if(srcImage.empty()){cout << "未能成功读取图片satori" << endl;return -1;}
Mat element = getStructuringElement(MORPH_RECT,Size(2*Elem_Size+1,2*Elem_Size+1),Point(Elem_Size,Elem_Size));
erode(srcImage,dstImage,element);
imshow("satori",dstImage);
createTrackbar("腐蚀/膨胀","satori",&value1,1,tracebar);
createTrackbar("内核尺寸","satori",&Elem_Size,21,tracebar);
tracebar(value1,0);
tracebar(Elem_Size,0);
while(char(waitKey(1)) != 'q');
return 0;
}
腐蚀是将暗的像素扩大,膨胀是将亮的像素扩大
在这个基础上还有开运算,闭运算,黑帽运算......
开运算其实就是分开细微链接的像素,闭运算是填平小的裂痕
https://blog.csdn.net/hanshanbuleng/article/details/80657148
Mat element = getStructuringElement(MORPH_RECT,Size(2*Elem_Size+1,2*Elem_Size+1),Point(Elem_Size,Elem_Size));
morphologyEx(srcImage,dstImage,MORPH_OPEN,element);
更改第三个参数即可
终于到快乐的算子环节了
在具体介绍之前,先来一起看看边缘检测的一般步骤吧。
-
1)滤波:边缘检测的算法主要是基于图像强度的一阶和二阶导数,但导数通常对噪声很敏感,因此必须采用滤波器来改善与噪声有关的边缘检测器的性能。常见的滤波方法主要有高斯滤波,即采用离散化的高斯函数产生一组归一化的高斯核(具体见“高斯滤波原理及其编程离散化实现方法”一文),然后基于高斯核函数对图像灰度矩阵的每一点进行加权求和(具体程序实现见下文)。
-
2)增强:增强边缘的基础是确定图像各点邻域强度的变化值。增强算法可以将图像灰度点邻域强度值有显著变化的点凸显出来。在具体编程实现时,可通过计算梯度幅值来确定。
-
3)检测:经过增强的图像,往往邻域中有很多点的梯度值比较大,而在特定的应用中,这些点并不是我们要找的边缘点,所以应该采用某种方法来对这些点进行取舍。实际工程中,常用的方法是通过阈值化方法来检测。
边缘检测应该是RM里面非常常用的算法了,识别装甲板应该主要就用了这个,识别到边缘之后solvepnp
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int Elem_Size = 1;
int value1;
Mat srcImage,dstImage;
int main()
{
srcImage = imread("satori.jpg");
if(srcImage.empty()){cout << "未能成功读取图片satori" << endl;return -1;}
Canny(srcImage,dstImage,300,100);
imshow("satori",dstImage);
while (char(waitKey(1)) != 'q');
return 0;
}
用canny很容易就可以看到效果,调节一下两个阈值则可以起到抑制噪声的作用
sobel算子可以计算x方向和y方向各自的梯度方向,相比canny而言,可以在一些相对比较特定(特征在x/y方向)的场景起到作用
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int Elem_Size = 1;
int value1;
Mat srcImage,dstImage,dstImage2,dstImage3;
int main()
{
Mat satori;
satori = imread("satori.jpg");
if(satori.empty()){cout << "未能成功读取图片satori" << endl;return -1;}
imshow("image",satori);
bilateralFilter(satori,srcImage,25,25*2,25/2);
cvtColor(srcImage,srcImage,CV_RGB2GRAY);
Sobel(srcImage,dstImage,srcImage.depth(),1,0,3,1,0,BORDER_DEFAULT);
Sobel(srcImage,dstImage2,srcImage.depth(),0,1,3,1,0,BORDER_DEFAULT);
imshow("satori",dstImage);
imshow("satori2",dstImage2);
addWeighted(dstImage,1,dstImage2,1,1,dstImage3);
imshow("satori3",dstImage3);
while (char(waitKey(1)) != 'q');
return 0;
}
结合了双边滤波后在x,y方向做sobel检测,然后合成,效果还行
结果试了一下双边滤波后做laplace检测,效果更好,啧啧
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int Elem_Size = 1;
int value1;
Mat srcImage,dstImage,dstImage2,dstImage3;
int main()
{
Mat satori;
satori = imread("satori.jpg");
if(satori.empty()){cout << "未能成功读取图片satori" << endl;return -1;}
imshow("image",satori);
bilateralFilter(satori,srcImage,25,25*2,25/2);
cvtColor(srcImage,srcImage,CV_RGB2GRAY);
Laplacian(srcImage,dstImage,srcImage.depth());
imshow("satori",dstImage);
while (char(waitKey(1)) != 'q');
return 0;
}
但是还有Scharr,可以看成对sobel的进一步优化?试试看效果
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int Elem_Size = 1;
int value1;
Mat srcImage,dstImage,dstImage2,dstImage3;
int main()
{
Mat satori;
satori = imread("satori.jpg");
if(satori.empty()){cout << "未能成功读取图片satori" << endl;return -1;}
bilateralFilter(satori,srcImage,25,25*2,25/2);
imshow("image",srcImage);
cvtColor(srcImage,srcImage,CV_RGB2GRAY);
Scharr(srcImage,dstImage,srcImage.depth(),1,0,1,0,BORDER_DEFAULT);
Scharr(srcImage,dstImage2,srcImage.depth(),0,1,1,0,BORDER_DEFAULT);
imshow("satori",dstImage);
imshow("satori2",dstImage2);
addWeighted(dstImage,1,dstImage2,1,1,dstImage3);
imshow("satori3",dstImage3);
while (char(waitKey(1)) != 'q');
return 0;
}
结果没看出了什么优化,反而引入了更多的噪声......可能是我参数没继续调吧(另一层面上来说更加敏锐?
好了,我们到了快乐的resize阶段,还有pryUp,pryDown这两个金字塔放大缩小函数
https://blog.csdn.net/poem_qianmo/article/details/26157633
我觉得没啥特别好说的,就是研究怎么样尽可能合理的采样或者插值,然后高斯函数的优势又一次被体现出来了。不得不这是个非常伟大的函数(我刚学到大数定律时就被这个函数的神奇性给吓到了)
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int main()
{
Mat tmpImage,dstImage;
tmpImage = imread("satori.jpg");
if(tmpImage.empty()){cout << "未能成功读取图片satori" << endl;return -1;}
dstImage = tmpImage;
while (1)
{
char key = waitKey(1);
switch (key)
{
case 'q':
return 0;
break;
case 'w':
resize(tmpImage,dstImage,Size(tmpImage.cols*2,tmpImage.rows*2));
break;
case 's':
resize(tmpImage,dstImage,Size(tmpImage.cols/2,tmpImage.rows/2));
break;
default:
break;
}
tmpImage = dstImage;
imshow("satori",dstImage);
}
}
不出意外的,缩小之后再放大之后会上天
霍夫线/圆检测算法
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int main()
{
Mat srcImage,dstImage,midImage;
srcImage = imread("test.jpg");
if(srcImage.empty()){cout << "未能成功读取图片" << endl;return -1;}
Canny(srcImage,midImage,400,100,3);
cvtColor(midImage,dstImage,CV_GRAY2BGR);
vector<Vec2f> lines;
HoughLines(midImage,lines,1,CV_PI/180,150,0,0);
for (size_t i = 0; i < lines.size(); i++)
{
float rho = lines[i][0] , theta = lines[i][1];
Point pt1,pt2;
double a = cos(theta),b = sin(theta);
double x0 = a*rho,y0 = b*rho;
pt1.x = cvRound(x0 + 1000 * (-b));
pt1.y = cvRound(y0 + 1000 * (a));
pt2.x = cvRound(x0 - 1000 * (-b));
pt2.y = cvRound(y0 - 1000 * (a));
line(dstImage,pt1,pt2,Scalar(55,100,95),1,CV_AA);
}
imshow("dst",dstImage);
while(char(waitKey(1)) != 'q');
return 0;
}
这个检测看得我头大
另外还有HoughLinesP这个检测方法,有点意思
试验了一下,HoughLinesP应该是相比HoughLine更优秀的一种检测方法,因为他是可以检测出线的起始的,而且有更多实用的可调参数(可以显示的最小/最大线段长度等)
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int main()
{
Mat srcImage,dstImage,midImage;
srcImage = imread("test.jpg");
if(srcImage.empty()){cout << "未能成功读取图片" << endl;return -1;}
Canny(srcImage,midImage,400,100,3);
cvtColor(midImage,dstImage,CV_GRAY2BGR);
vector<Vec4i> lines;
HoughLinesP(midImage,lines,1,CV_PI/180,150,0,0);
for (size_t i = 0; i < lines.size(); i++)
{
Vec4i l = lines[i];
line(dstImage,Point(l[0],l[1]),Point(l[2],l[3]),Scalar(0,100,0),5,CV_AA);
}
imshow("dst",dstImage);
while(char(waitKey(1)) != 'q');
return 0;
}
还有HoughCircles
RM上倒是不太用得上圆检测,除非要检测弹丸(谁没事干检测弹丸),RC这边倒是应该用得上?毕竟有几次赛题要扔球来着
HoughCircles里面,第5个参数规定了可检测的最大半径的圆形,有筛选作用,6,7参数则起到规定阈值的作用,也挺有用的
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int main()
{
Mat srcImage,dstImage,midImage;
srcImage = imread("circle.jpeg");
if(srcImage.empty()){cout << "未能成功读取图片" << endl;return -1;}
cvtColor(srcImage,midImage,CV_BGR2GRAY);
GaussianBlur(midImage,midImage,Size(3,3),1,1);
vector<Vec3f> circles;
HoughCircles(midImage,circles,CV_HOUGH_GRADIENT,1.5,20,300,100,0,0);
for (size_t i = 0; i < circles.size(); i++)
{
Point center(cvRound(circles[i][0]),cvRound(circles[i][1]));
int radius = cvRound(circles[i][2]);
circle(srcImage,center,radius,Scalar(0,100,0),3);
}
imshow("dst",srcImage);
while(char(waitKey(1)) != 'q');
return 0;
}
调了一下参数,能够比较好的检测到图中想找的圆
去研究源码实现的话会发现其实HoughCircle,HoughCircleP其实都是基于HoughCircle2(旧的霍夫圆检测)实现的,这里就不去太扣底层的东西,来日方长
到快乐的漫水填充算法了
floodFill从功能上去理解就是和ps的魔术棒一样,总的来说是非常重要的一个功能
https://blog.csdn.net/poem_qianmo/article/details/28261997
算法原理其实挺好理解的,就是先选中一个点作为种子,以这个种子作为起点去计算周边像素差值,在阈值范围内的像素作为下一批种子
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int main()
{
Mat srcImage,dstImage,midImage;
srcImage = imread("satori.jpg");
if(srcImage.empty()){cout << "未能成功读取图片" << endl;return -1;}
Rect ccomp;
floodFill(srcImage,Point(0,0),Scalar(0,0,0),&ccomp,Scalar(10,10,10),Scalar(10,10,10));
imshow("dst",srcImage);
while(char(waitKey(1)) != 'q');
return 0;
}
最后两个Scalar参数用来框定选取阈值,调了一下之后可以比较好的把人物抠出来
floodFill还可以设置掩膜模式,避免漫水填充到掩膜内的非0像素
最后floodFill的最后参数还有一个32位操作数,高8位,中8位,低8位都有含义,相当的复杂,这里就不去过于细致的研究了
emmmm到角点检测了
角点检测应该是比较重要的,不管是RC还是RM,像RM里面识别到灯柱之后,要给灯柱四个角上的关键点都标注出来,然后才能做pnp结算
图像特征类型可以被分为如下三种:
- <1>边缘
- <2>角点 (感兴趣关键点)
- <3>斑点(Blobs)(感兴趣区域)
在当前的图像处理领域,角点检测算法可归纳为三类:
- <1>基于灰度图像的角点检测
- <2>基于二值图像的角点检测
- <3>基于轮廓曲线的角点检测
角点检测算法又是梯度运算的一大应用场景(这个想想就能知道)
配合角点检测的还有一个知名度非常高的方法那就是二值化——一共有5种方法(茴香豆的茴字有几种写法啊?)
先试一下二值化
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int thresholdvalue;
Mat src_image,dst_image1,dst_image2,dst_image3,dst_image4,dst_image5;
static void tracebar(int,void*)
{
threshold(src_image,dst_image1,thresholdvalue,255,THRESH_BINARY);
imshow("test",dst_image1);
}
int main()
{
src_image = imread("satori.jpg");
if(src_image.empty()){cout << "未能成功读取图片" << endl;return -1;}
cvtColor(src_image,src_image,CV_BGR2GRAY);
namedWindow("test");
createTrackbar("threshold","test",&thresholdvalue,255,tracebar);
tracebar(thresholdvalue,0);
while(char(waitKey(1)) != 'q');
return 0;
}
我就不尝试每一种方法了,总之还是挺立竿见影的
值得一提的是connerHarris之后的图像只有通过一个很小阈值的threshold之后才能显现出来
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int thresholdvalue;
Mat src_image,dst_image1,dst_image2,dst_image3,dst_image4,dst_image5;
Mat mid_image;
int main()
{
src_image = imread("satori.jpg");
if(src_image.empty()){cout << "未能成功读取图片" << endl;return -1;}
cvtColor(src_image,src_image,CV_BGR2GRAY);
namedWindow("test");
cornerHarris(src_image,mid_image,5,3,0.01);
threshold(mid_image,dst_image1,0.0001,255,THRESH_BINARY);
imshow("test",dst_image1);
while(char(waitKey(1)) != 'q');
return 0;
}
可以知道经过harris检测之后的图像值被放的很小,不符合我们常用的0-255的灰度规定,所以如果要用的话,一般来说还得经过操作,比如二值化,比如normolize
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int thresholdvalue;
Mat src_image,dst_image1,dst_image2,dst_image3,dst_image4,dst_image5;
Mat mid_image;
static void tracebar(int,void*)
{
threshold(dst_image1,dst_image2,thresholdvalue,255,THRESH_BINARY);
imshow("test",dst_image2);
}
int main()
{
src_image = imread("satori.jpg");
if(src_image.empty()){cout << "未能成功读取图片" << endl;return -1;}
cvtColor(src_image,src_image,CV_BGR2GRAY);
namedWindow("test");
cornerHarris(src_image,mid_image,5,3,0.01);
normalize(mid_image,dst_image1,0,255,NORM_MINMAX,CV_32FC1,Mat());
createTrackbar("threshold","test",&thresholdvalue,255,tracebar);
tracebar(thresholdvalue,0);
while(char(waitKey(1)) != 'q');
return 0;
}
这下拖条总算有点用了
重映射和surf特征点检测,surf应该是比较通用的特征点检测算法了,总体而言
https://www.cnblogs.com/dengxiaojun/p/5302778.html
重映射主要是使用remap这个函数,remap的数学定义如下
其中 \(map_1\)和\(map_2\)都是作为参数输入remap函数的(值得一提的是map的类型不是随意的,create的时候要创建CV_32FC1)
这里有个简单的应用,比如将一个图片给镜像翻转
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "iostream"
using namespace cv;
using namespace std;
int thresholdvalue;
Mat src_image,dst_image1,dst_image2;
Mat mid_image;
int main()
{
src_image = imread("satori.jpg");
if(src_image.empty()){cout << "未能成功读取图片" << endl;return -1;}
Mat map_x,map_y;
map_x.create(src_image.size(),CV_32FC1);
map_y.create(src_image.size(),CV_32FC1);
for (size_t i = 0; i < src_image.cols; i++)
{
for (size_t j = 0; j < src_image.rows; j++)
{
map_x.at<float>(j,i) = static_cast<float>(i);
map_y.at<float>(j,i) = static_cast<float>(src_image.rows-j);
}
}
remap(src_image,dst_image1,map_x,map_y,CV_INTER_LINEAR,BORDER_CONSTANT,Scalar(0,0,0));
imshow("dst1",dst_image1);
while(char(waitKey(1)) != 'q');
return 0;
}
这个没有什么难度的,像素操作的时候稍微注意一点就行了
然后我们来关注一下重点——我们亲爱的SURF特征检测算法,在opencv里面SURF被封成了一个类,有一堆可以执行的乱七八糟的操作
我们跑个drawKeypoints试试
然后发现好玩的事情
https://blog.csdn.net/zhounanzhaode/article/details/50302385
所以为了使用SURF我还得再操作一下,真的烦人
直接使用apt-get的方法失败了,会在添加ppa时报错——
没有 Release 文件。 N: 无法安全地用该源进行更新,所以默认禁用该源。
采用自己摸索的方法操作成功,具体的操作方法为:
去github上面下载这个库
https://github.com/opencv/opencv_contrib/tree/3.4
然后checkout到和opencv版本一致的tag下面
将opencv_contrib/modules/....下面所需要的模组复制粘贴到opencv/modules/下面去
然后修改opencv/modules下的cmakelist中的
set(FIXED_ORDER_MODULES core imgproc imgcodecs videoio highgui video calib3d features2d objdetect dnn ml flann photo stitching xfeatures2d)
添加自己需要的新的模组即可
然后就是和安装时一样的cmake,make,make install三连,成功的话就可以正常include nonfree.hpp了
emmmmmm
看了一下,更新到opencv3之后,SURF的使用和opencv2完全不一样了,除了drawKeypoint这个api仍然保留之外,其他的好像都变化了
随手找了个教程,结果又找到RM相关的了,哈哈哈哈哈哈
https://www.cnblogs.com/long5683/p/9692987.html
结果发现之前的操作没弄干净......编译的时候出现了这个
error: (-213:The function/feature is not implemented) This algorithm is patented and is excluded in this configuration; Set OPENCV_ENABLE_NONFREE CMake option and rebuild the library in function 'create'
结果就是还得在cmake的时候操作一下.....干啊
https://blog.csdn.net/zhoukehu_CSDN/article/details/83145026
按照这个博客走,记得在cmake-gui里面把OPENCV_ENABLE_NONFREE这个勾上......唉,这个是真的恶心,又要make整整20分钟了
重新make完之后就一切ok了
#include "opencv2/core/core.hpp"
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/opencv.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/features2d/features2d.hpp"
#include "opencv2/xfeatures2d/nonfree.hpp"
#include "iostream"
using namespace cv;
using namespace cv::xfeatures2d;
using namespace std;
int thresholdvalue;
Mat src_image,dst_image1,dst_image2;
Mat mid_image;
int main()
{
src_image = imread("satori.jpg");
if(src_image.empty()){cout << "未能成功读取图片" << endl;return -1;}
int minHessian = 400;
Ptr<SURF> detector = SURF::create(minHessian);
vector<KeyPoint> keypoints;
detector->detect(src_image,keypoints,Mat());
drawKeypoints(src_image,keypoints,dst_image1,Scalar(0,0,0));
imshow("dst",dst_image1);
while(char(waitKey(1)) != 'q');
return 0;
}
暂时跟着demo跑的内容就先这么多吧
实际上还需要学习的内容包括相机有关的一些知识(内参矩阵标定),solvePnp的使用等等,包括video模块下的一些有用的功能我也还没去做了解
后面就开始看几个实际项目研究研究,一边巩固已经学习的东西,一边积累工作经验吧