#include "stdafx.h"
#include <core.hpp>
#include <highgui.hpp>
#include <iostream>
#include <fstream>
using namespace cv;
using namespace std;
#pragma comment(lib,"opencv_core220d.lib")
#pragma comment(lib,"opencv_highgui220d.lib")
int _tmain(int argc, _TCHAR* argv[])
{
ofstream outFile;
outFile.open("pixel_32f.txt");
char name[100] = "d://picture//lena.jpg";
namedWindow("show", CV_WINDOW_AUTOSIZE);
{
Mat pnm = imread(name, -1);
if (pnm.empty())
{
cout<<"read error"<<endl;
return -1;
}
Mat temp;
pnm.convertTo(temp, CV_32FC3);
vector<Mat> vec;
split(temp, vec);
cout<<"pnm.channels = "<<pnm.channels()<<endl;
cout<<"pnm.depth = "<<pnm.depth()<<endl;
cout<<"pnm.dims = "<<pnm.dims<<endl;
cout<<"pnm.elemsize = "<<pnm.elemSize()<<endl;
cout<<"pnm.elemsize1 = "<<pnm.elemSize1()<<endl;
for (int k=0; k<pnm.dims;++k)
{
cout<<"pnm.step = "<<pnm.step[k]<<endl;
}
cout<<"pnm.step1 = "<<pnm.step1()<<endl;
cout<<"pnm.type = "<<pnm.type()<<endl;
cout<<"pnm.total = "<<pnm.total()<<endl;
cout<<"pnm.rows*cols = "<<pnm.cols*pnm.rows<<endl;
cout<<"pnm.rows = "<<pnm.rows<<endl;
cout<<"pnm.cols = "<<pnm.cols<<endl;
cout<<endl;
cout<<"temp.channels = "<<temp.channels()<<endl;
cout<<"temp.depth = "<<temp.depth()<<endl;
cout<<"temp.dims = "<<temp.dims<<endl;
cout<<"temp.elemsize = "<<temp.elemSize()<<endl;
cout<<"temp.elemsize1 = "<<temp.elemSize1()<<endl;
for (int m=0; m<temp.dims; ++m)
{
cout<<"temp.step = "<<temp.step[m]<<endl;
}
cout<<"temp.step1 = "<<temp.step1()<<endl;
cout<<"temp.type = "<<temp.type()<<endl;
cout<<"temp.total = "<<temp.total()<<endl;
cout<<"temp.rows = "<<temp.rows<<endl;
cout<<"temp.cols = "<<temp.cols<<endl;
cout<<endl;
int i,j;
======section1======//
int64 beg = cvGetTickCount();
vector<Mat> spl;
split(temp, spl);
for (i=0; i<temp.rows; ++i)
{
float *pt = spl[0].ptr<float>(i);
for (j=0; j<temp.cols; ++j)
{
float mm = pt[j];
// if(i<6)
// outFile<<mm<<",";
mm = mm/(float)20.6;
}
// outFile<<endl;
}
merge(spl, temp);
int64 second = cvGetTickCount();
cout<<(second-beg)/cvGetTickFrequency()<<endl;
outFile<<(second-beg)/cvGetTickFrequency()<<endl;
//=========section2=========
int64 secbeg = cvGetTickCount();
for (i=0; i<temp.rows; ++i)
{
for (j=0; j<temp.cols; ++j)
{
float *mm = &(temp.ptr<float>(i)[3*j]);
// if(i<6)
// outFile<<*mm<<",";
*mm = *mm/(float)20.6;
}
// outFile<<endl;
}
int64 third = cvGetTickCount();
cout<<(third-secbeg)/cvGetTickFrequency()<<endl;
outFile<<(third-secbeg)/cvGetTickFrequency()<<endl;
//============section3==========//
int64 thridbeg = cvGetTickCount();
int col=temp.cols, row = temp.rows;
if (temp.isContinuous())
{
col*=row;
row =1;
}
for (i=0; i<row; ++i)
{
const float *pt = temp.ptr<float>(i);
for (j=0; j<col;++j)
{
float mm=pt[3*j];
// outFile<<mm<<",";
mm = mm/(float)20.6;
}
// outFile<<endl;
}
int64 four = cvGetTickCount();
cout<<(four-thridbeg)/cvGetTickFrequency()<<endl;
outFile<<(four-thridbeg)/cvGetTickFrequency()<<endl;
///===========section4===============/
int64 fourbeg = cvGetTickCount();
int step0=temp.step[0],step1=temp.step[1];
for (i=0; i<temp.rows; ++i)
{
for (j=0; j<temp.cols; ++j)
{
float *pix = (float *)(temp.data+i*step0+j*step1);
// if(i<6)
// outFile<<*pix<<",";
*pix = *pix/(float)20.6;
}
// outFile<<endl;
}
int64 fifth = cvGetTickCount();
cout<<(fifth-fourbeg)/cvGetTickFrequency()<<endl;
outFile<<(fifth-fourbeg)/cvGetTickFrequency()<<endl;
//============section5==========//
int64 fifthbeg = cvGetTickCount();
int step00=temp.step[0],step01=temp.step[1];
int col2=temp.cols, row2 = temp.rows;
if (temp.isContinuous())
{
col2*=row2;
row2 =1;
}
for (i=0; i<row2; ++i)
{
for (j=0; j<col2;++j)
{
float *mm= (float *)(temp.data+i*step00+j*step01);
//outFile<<mm<<",";
*mm = *mm/(float)20.6;
}
// outFile<<endl;
}
int64 sixth = cvGetTickCount();
cout<<(sixth-fifthbeg)/cvGetTickFrequency()<<endl;
outFile<<(sixth-fifthbeg)/cvGetTickFrequency()<<endl;
outFile.close();
imshow("show", pnm);
waitKey(0);
}
return 0;
}
下面是某次的運作結果以及對于Mat的幾個參數的分析。
//
pnm.channels = 3 ------------------通道數。
pnm.depth = 0 ------------------矩陣元素的基礎元素類型，CV_8U(uchar), CV_32F(float)...等.
pnm.dims = 2 ------------------矩陣的維數，圖像一般為2維的;
pnm.elemsize = 3 ------------------矩陣每個元素的大小（所占位元組數）=channel×sizeof(elemsize1);
pnm.elemsize1 = 1 ------------------矩陣每個元素的基礎元素大小=sizeof(uchar, float......等);
pnm.step = 1536 [0]---------------矩陣每行所占的位元組數，包括用于位元組對齊的位元組,如果不用補齊位元組則=cols*elemsize;
pnm.step = 3 [1]---------------矩陣中每個元素所占位元組數，對于2維矩陣=elemsize。
pnm.step1 = 1536 ------------------return normalized step; =step[0]/elemsize1，即每行的步長，it can be useful for fast access to arbitrary matrix element;
pnm.type = 16 ------------------傳回flags中表示每個元素類型，即CV_8UC3，CV_32FC1等表示的整數
pnm.total = 262144 ------------------=rows*cols
pnm.rows*cols = 262144
pnm.rows = 512
pnm.cols = 512
temp.channels = 3
temp.depth = 5
temp.dims = 2
temp.elemsize = 12
temp.elemsize1 = 4
temp.step = 6144
temp.step = 12
temp.step1 = 1536
temp.type = 21
temp.total = 262144
temp.rows = 512
temp.cols = 512
37098.3
34995
4189.36
6048.81
5776.15
//
最後的幾個運作結果每次是不一樣的，但是基本可以肯定的是前兩種是差不多的，第三種最少，第四和第五不相上下，但是第三種的缺點是丢失了原始的行列資訊，對于不計較元素位置的處理來說是首選，正如opencv2.2的手冊中所說，在存儲結構中沒有gap，計算不是太複雜時，可以提升10%-20%的效率。而對于多通道下的需要處理行列資訊的計算來說，個人偏好第四種，也就是opencv1.x的方式，直接使用原始指針操作，習慣上用着友善。第一二種方法是C++API的新的擷取方式，當然還有使用疊代器方式的存取等，詳細的見手冊或者cheet_sheet中提供的方法。
當然上面隻是對于元素類型是CV_32FC3的進行的測試，對于基礎的CV_8UC1/3沒有測試。
------後記-------
對于第四種費了很長時間才搞明白，開始的時候沒弄清楚，取出的資料不是不對就是直接出現通路異常。對于新的Mat結構所有的有效像素資料都是存儲在uchar* data中，是以取用時需要自己顯式的轉換類型，開始時沒有意識到在記憶體中存儲的是uchar類型，直接使用了*(temp.data+i*step0+j*step1)來擷取資料，結果總是取到錯誤的資料，後來意識到存儲的是float資料，需要轉換，就直接在前面加了(float)結果還是不行，最後仔細分析之後終于發現問題了====*(temp.data+i*step0+j*step1)是按照uchar将資料取出，轉換為uchar型的資料，如果在前面加上float轉換，隻是将uchar提升為了float而不是取出的是float，原來如此，将其改為(float *)(temp.data+i*step0+j*step1)，也就是按照float的形式取出，也就是一次取出4個位元組作為一個資料，這樣就可以了。
![利用Fiddler進行移動端弱網測試[圖]](data:image/gif;base64,R0lGODlhAQABAIAAAP///wAAACwAAAAAAQABAAACAkQBADs=)