收入囊中
- meanshift圖像聚類
- meanshift object detect
葵花寶典
今天有點累，理論就講少點吧T_T
meanshift中文是均值飄逸，就是給定一個點，然後會移動到概率密度最大的地方。
對於圖像，什麼是概率密度最大？
我們可以定義很多要素：
距離
RGB
HSV
下面我有個例子，就是用距離(x,y)和HSV(h,s,v)作圖像聚類的。
於是我們有5個要素，當前點與其他點的距離，HSV越接近，則概率密度越高。
假定我們有一點（m,n),如何選擇下一個點呢，如何在一個矩形中找到概率密度最高的那個點？
top,down,left,right是我們選的矩形，(m,n)是中心
for(s=top;s<=down;s++){
for(t=left;t<=right;t++)
{
ws=(s-m)*(s-m)+(t-n)*(t-n);//spatial information
ws/=(hs*hs);
ws=exp(-ws);
wr=(data[s*step+t*channels]-data[m*step+n*channels])*(data[s*step+t*channels]－data[m*step+n*channels]);
wr+=(data[s*step+t*channels+1]-data[m*step+n*channels+1])*(data[s*step+t*channels+1]-data[m*step+n*channels+1]);
wr+=(data[s*step+t*channels+2]-data[m*step+n*channels+2])*(data[s*step+t*channels+2]-data[m*step+n*channels+2]);
wr/=(hr*hr);
if(wr>1)
wr=0.;
else
wr=exp(-wr);
sumw+=wr*ws;
for(k=0;k<5;k++)//try
y[1][k]+=oridata[s*width+t][k]*wr*ws;
}
}
for(k=0;k<5;k++) //try
y[1][k]/=sumw;
//下一個要到的點
m=(int)(y[1][0]+0.5);
n=(int)(y[1][1]+0.5);圖像聚類
用meanshift可以實現圖像聚類，因爲本次重點是mean shift與back project的結合，所以關於聚類我就直接貼上代碼了
用meanshift可以實現圖像聚類，因爲本次重點是mean shift與back project的結合，所以關於聚類我就直接貼上代碼了
算法超級easy,就是循環圖像，對點作mean shift，每個點最大迭代100次，一直到點不動（收斂）爲止，然後用vector記錄下移動過程，將收斂點的顏色賦值給這些點，並記錄哪些點已經賦值過了，下次就不用循環直接continue.
hs,hr是尋找的矩形大小，比EPSLON小就收斂。
note:是用x,y和HSV
//#include <stdafx.h>
#include <cxcore.h>
#include <cv.h>
#include <highgui.h>
#include <stdio.h>
#include <vector>
#include <utility>
using namespace cv;
using namespace std;
#define hs 25
#define hr 25
#define maxstep 100
#define EPSLON 0.01
#define width 1920
#define height 1080
float oridata[height*width][5];//try
int visited[height][width];
int main(int argc,char **argv)
{
IplImage *oriImg,*luvImg,*fltImg,*afterImg;
char *filename = argv[1];
oriImg=cvLoadImage(filename,1);
if(!oriImg){
printf("cannot load the file.\n");
return -1;
}
luvImg=cvCreateImage(cvSize(width,height),oriImg->depth,oriImg->nChannels);
fltImg=cvCreateImage(cvSize(width,height),oriImg->depth,oriImg->nChannels);
afterImg=cvCreateImage(cvSize(width,height),oriImg->depth,oriImg->nChannels);
cvCvtColor(oriImg,luvImg,CV_RGB2Luv);
uchar *data,*newdata;
int channels, step,depth;
depth=luvImg->depth;
step=luvImg->widthStep;
channels=luvImg->nChannels;
data=(uchar *) luvImg->imageData;
newdata=(uchar *) fltImg->imageData;
int i,j,s,t,k,m,n,index;
int top,down,left,right;
float ws,wr;
float mhlength;
float sumw;
float y[2][5];//try, otherwise the second is 5
for(i=0;i<height;i++)
{
for(j=0;j<width;j++)
{
oridata[i*width+j][0]=i;
oridata[i*width+j][1]=j;
visited[i][j]=0;
oridata[i*width+j][2]=data[i*step+j*channels];
oridata[i*width+j][3]=data[i*step+j*channels+1];
oridata[i*width+j][4]=data[i*step+j*channels+2];
}
}
for(i=0;i<height;i++)
{
for(j=0;j<width;j++)
{
if(visited[i][j])
continue;
m=i;n=j;//當前的中心點
for(k=0;k<5;k++)//try
y[0][k]=oridata[i*width+j][k];//向量的初始值
vector<pair<int,int> >vss;
for(index=0;index<maxstep;index++)//對當前的結點而言，最多迭代100次
{
pair<int,int>newone;
newone = make_pair(m,n);
vss.push_back(newone);
for(k=0;k<5;k++)//try
y[1][k]=0;
mhlength=0.;
sumw=0.;
top=m-hs;
down=m+hs;
left=n-hs;
right=n+hs;
if(top<0) top =0;
if(down>height-1) down=height-1;
if(left<0) left=0;
if(right>width-1) right=width-1;
for(s=top;s<=down;s++)
{
for(t=left;t<=right;t++)
{
ws=(s-m)*(s-m)+(t-n)*(t-n);//spatial information
ws/=(hs*hs);
ws=exp(-ws);
//ws=1-ws+(ws*ws)/2-(ws*ws*ws)/6+(ws*ws*ws*ws)/24-(ws*ws*ws*ws*ws)/120;
wr=(data[s*step+t*channels]-data[m*step+n*channels])*(data[s*step+t*channels]-data[m*step+n*channels]);
wr+=(data[s*step+t*channels+1]-data[m*step+n*channels+1])*(data[s*step+t*channels+1]-data[m*step+n*channels+1]);
wr+=(data[s*step+t*channels+2]-data[m*step+n*channels+2])*(data[s*step+t*channels+2]-data[m*step+n*channels+2]);
wr/=(hr*hr);
if(wr>1)
wr=0.;
else
wr=exp(-wr);
sumw+=wr*ws;
for(k=0;k<5;k++)//try
y[1][k]+=oridata[s*width+t][k]*wr*ws;
}
}
for(k=0;k<5;k++) //try
y[1][k]/=sumw;
//下一個要到的點
m=(int)(y[1][0]+0.5);
n=(int)(y[1][1]+0.5);
if(visited[m][n])
break;
if(m<hs||m>height-hs||n<hs||n>width-hs)
break;
else{
for(k=0;k<5;k++)//try
{
mhlength+=(y[1][k]-y[0][k])*(y[1][k]-y[0][k]);
y[0][k]=y[1][k];
}
mhlength=sqrt(mhlength);
if(mhlength<EPSLON)//找到極值點
break;
}
}//鄰域處理結束
for(int ii = 0;ii < vss.size();ii++)
{
int row = vss[ii].first;
int line = vss[ii].second;
newdata[row*step+line*channels]=int(y[1][2]+0.5);//try
newdata[row*step+line*channels+1]=int(y[1][3]+0.5);
newdata[row*step+line*channels+2]=int(y[1][4]+0.5);
visited[row][line]=1;
}
}
}
cvCvtColor(fltImg,afterImg,CV_Luv2RGB);
cvNamedWindow("ori",1);
cvNamedWindow("filtered",1);
cvShowImage("ori",oriImg);
cvShowImage("filtered",afterImg);
cvWaitKey(0);
cvReleaseImage(&oriImg);
cvDestroyWindow("image");
cvReleaseImage(&afterImg);
cvDestroyWindow("filtered");
return 0;
}初識API
- C++: int meanShift(InputArray probImage, Rect& window, TermCriteria criteria)
-
- probImage – 反向投影矩陣
- window – 一開始的窗口,就是自己用矩形圈起來的地方
- criteria – Stop criteria for the iterative search algorithm.
TermCriteria::TermCriteria
The constructors.
- C++: TermCriteria::TermCriteria()
- C++: TermCriteria::TermCriteria(int type, int maxCount, double epsilon)
- C++: TermCriteria::TermCriteria(const CvTermCriteria& criteria)
-
- type – The type of termination criteria: TermCriteria::COUNT, TermCriteria::EPS or TermCriteria::COUNT +TermCriteria::EPS.
- maxCount – The maximum number of iterations or elements to compute.
- epsilon – The desired accuracy or change in parameters at which the iterative algorithm stops.
- criteria – Termination criteria in the deprecated CvTermCriteria format.
舉個例子：假設我們對(110,260)座標處寬35長40的方框感興趣，最大迭代10次，差值（精確度）爲0.01
result是我們的反向投影矩陣，那就可以如下調用
Rect rect(110,260,35,40);
TermCriteria criteria(TermCriteria::MAX_ITER,10,0.01);
meanShift(result,rect,criteria);
TermCriteria criteria(TermCriteria::MAX_ITER,10,0.01);
meanShift(result,rect,criteria);
荷槍實彈
下面我附上完整代碼，並有詳細註釋
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/imgproc/imgproc.hpp"
#include "opencv2/video/tracking.hpp"
#include <iostream>
using namespace cv;
using namespace std;
class ColorHistogram {
private:
int histSize[3];
float hranges[2];
const float* ranges[3];
int channels[3];
public:
ColorHistogram() {
// Prepare arguments for a color histogram
histSize[0]= histSize[1]= histSize[2]= 256;
hranges[0]= 0.0; // BRG range
hranges[1]= 255.0;
ranges[0]= hranges; // all channels have the same range
ranges[1]= hranges;
ranges[2]= hranges;
channels[0]= 0; // the three channels
channels[1]= 1;
channels[2]= 2;
}
//這裏獲取HSV的第一個通道的直方圖
Mat getHueHistogram(const cv::Mat &image,int minSaturation=0) {
Mat hist;
//色彩空間轉換
Mat hsv;
cvtColor(image, hsv, CV_BGR2HSV);
// mask,要注意哦
Mat mask;
if (minSaturation>0) {
// 三通道分開
vector<Mat> v;
split(hsv,v);
// 去除低飽和的值
threshold(v[1],mask,minSaturation,255,THRESH_BINARY);
}
hranges[0]= 0.0; //hue的range是[0,180]
hranges[1]= 180.0;
channels[0]= 0; calcHist(&hsv,1,channels,mask,hist,1,histSize,ranges);//注意，mask不爲0的地方計算在內，也就是高飽和的Hue計算在內
return hist;
}
};
class ContentFinder {
private:
float hranges[2];
const float* ranges[3];
int channels[3];
float threshold;
Mat histogram;
public:
ContentFinder() : threshold(-1.0f) {
hranges[0]= 0.0; // range [0,255]
hranges[1]= 255.0;
channels[0]= 0; // the three channels
channels[1]= 1;
channels[2]= 2;
ranges[0]= hranges; // all channels have same range
ranges[1]= hranges;
ranges[2]= hranges;
}
// Sets the reference histogram
void setHistogram(const Mat& h) {
histogram= h;
normalize(histogram,histogram,1.0);
}
cv::Mat find(const cv::Mat& image, float minValue, float maxValue, int *channels, int dim) {
cv::Mat result;
hranges[0]= minValue;
hranges[1]= maxValue;
ranges[0]= hranges; // all channels have same range
ranges[1]= hranges;
ranges[2]= hranges;
calcBackProject(&image,1,channels,histogram,result,ranges,255.0);
return result;
}
};
int main( int, char** argv )
{
Mat image= cv::imread("baboon1.jpg");
// Baboon's face ROI
Mat imageROI= image(cv::Rect(110,260,35,40));
// 獲得Hue直方圖
int minSat=65;
ColorHistogram hc;
Mat colorhist = hc.getHueHistogram(imageROI,minSat);
ContentFinder finder;
finder.setHistogram(colorhist);
image= cv::imread("baboon3.jpg");
// 色彩空間轉換
Mat hsv;
cvtColor(image, hsv, CV_BGR2HSV);
vector<Mat> v;
split(hsv,v);
//除去低飽和度
threshold(v[1],v[1],minSat,255,cv::THRESH_BINARY);
// 獲得反向投影
int ch[1]={0};
Mat result= finder.find(hsv,0.0f,180.0f,ch,1);
// 除去低飽和度的點
bitwise_and(result,v[1],result);
Rect rect(110,260,35,40);
rectangle(image, rect, Scalar(0,0,255));
TermCriteria criteria(TermCriteria::MAX_ITER,10,0.01);
meanShift(result,rect,criteria);
rectangle(image, rect, cv::Scalar(0,255,0));
namedWindow("result");
imshow("result",image);
waitKey(0);
return 0;
} 看下效果
舉一反三
camshift算法是對meanshift算法的改進，它先用mean shift找到物體，再調整窗的大小，然後可以旋轉窗找到最合適的旋轉角度
計算機視覺討論羣：162501053
轉載請註明：http://blog.csdn.net/abcd1992719g
