像素级别遍历是我们在图像任务中经常遇到的问题，在实时的图像处理中，能够高效的访问像素数据是很重要的。OpenCV中的数据容器是cv::Mat，cv::Mat提供了三种数据访问的方式分别是下标寻址，指针访问，迭代器访问。下面我们对比下这几种不同方式的访问速度。
#include <iostream>
#include <assert.h>
#include "opencv2/core/core.hpp"
using namespace std;
void method1(cv::Mat);
void method2(cv::Mat);
void method3(cv::Mat);
void method4(cv::Mat);
void method5(cv::Mat);
void method6(cv::Mat);
void method7(cv::Mat);
int main(int argc, char* argv[])
{
cv::Size imgSize(6400,4800);
cv::Mat image = cv::Mat(imgSize, CV_8UC3, cv::Scalar(1,1,1));
method1(image);
method2(image);
method3(image);
method4(image);
method5(image);
method6(image);
method7(image);
}
void method1(cv::Mat img){
// at access with Vec3b Vector
double t0 = (double) cv::getTickCount();
int height = img.rows;
int width = img.cols;
int sum = 0;
for(int row=0; row < height; row++){
for(int col=0; col < width; col++){
cv::Vec3b uc_pixel = img.at<cv::Vec3b>(row, col);
int a = uc_pixel[0];
int b = uc_pixel[1];
int c = uc_pixel[2];
sum += a + b + c;
}
}
assert(sum==3*height*width);
double time = ((double) cv::getTickCount() - t0) / cv::getTickFrequency();
std::cout << "Time for method1: " << time << std::endl;
}
void method2(cv::Mat img){
// direct at access
double t0 = (double) cv::getTickCount();
int height = img.rows;
int width = img.cols;
int sum = 0;
for(int row=0; row < height; row++){
for(int col=0; col < width; col++){
int a = img.at<cv::Vec3b>(row, col)[0];
int b = img.at<cv::Vec3b>(row, col)[1];
int c = img.at<cv::Vec3b>(row, col)[2];
sum += a + b + c;
}
}
assert(sum==3*height*width);
double time = ((double) cv::getTickCount() - t0) / cv::getTickFrequency();
std::cout << "Time for method2: " << time << std::endl;
}
void method3(cv::Mat img){
// pointer + Vec3b vector
double t0 = (double) cv::getTickCount();
int height = img.rows;
int width = img.cols;
int sum = 0;
for(int row=0; row < height; row++){
cv::Vec3b *ptr = img.ptr<cv::Vec3b>(row);
for(int col=0; col < width; col++){
cv::Vec3b pixel = ptr[col];
int a = pixel[0];
int b = pixel[1];
int c = pixel[2];
sum += a + b + c;
}
}
assert(sum==3*height*width);
double time = ((double) cv::getTickCount() - t0) / cv::getTickFrequency();
std::cout << "Time for method3: " << time << std::endl;
}
void method4(cv::Mat img){
// pointer
double t0 = (double) cv::getTickCount();
int height = img.rows;
int width = img.cols;
int sum = 0;
for(int row=0; row < height; row++){
cv::Vec3b *ptr = img.ptr<cv::Vec3b>(row);
for(int col=0; col < width; col++){
int a = ptr[col][0];
int b = ptr[col][1];
int c = ptr[col][2];
sum += a + b + c;
}
}
assert(sum==3*height*width);
double time = ((double) cv::getTickCount() - t0) / cv::getTickFrequency();
std::cout << "Time for method4: " << time << std::endl;
}
void method5(cv::Mat img){
// raw pointer
double t0 = (double) cv::getTickCount();
int height = img.rows;
int width = img.cols;
int sum=0;
for(int row=0; row < height; row++){
const uchar *ptr = img.ptr(row);
for(int col=0; col < width; col++){
const uchar *uc_pixel = ptr;
int a = uc_pixel[0];
int b = uc_pixel[1];
int c = uc_pixel[2];
// int a = ptr[0];
// int b = ptr[1];
// int c = ptr[2];
sum += a + b + c;
ptr += 3;
}
}
assert(sum==3*height*width);
double time = ((double) cv::getTickCount() - t0) / cv::getTickFrequency();
std::cout << "Time for method5: " << time << std::endl;
}
void method6(cv::Mat img){
// raw pointer + raw step
double t0 = (double) cv::getTickCount();
int height = img.rows;
int width = img.cols;
int sum = 0;
const uchar *uc_pixel = img.data;
for(int row=0; row < height; row++){
uc_pixel = img.data + row*img.step;
for(int col=0; col < width; col++){
int a = uc_pixel[0];
int b = uc_pixel[1];
int c = uc_pixel[2];
sum += a + b + c;
uc_pixel += 3;
}
}
assert(sum==3*height*width);
double time = ((double) cv::getTickCount() - t0) / cv::getTickFrequency();
std::cout << "Time for method6: " << time << std::endl;
}
void method7(cv::Mat image){
double t0 = (double) cv::getTickCount();
int height = image.rows;
int width = image.cols;
cv::MatConstIterator_<cv::Vec3b> it = image.begin<cv::Vec3b>(), it_end = image.end<cv::Vec3b>();
int sum = 0;
for(; it != it_end; ++it){
int a = (*it)[0];
int b = (*it)[1];
int c = (*it)[2];
sum += a + b + c;
}
assert(sum==3*height*width);
double time = ((double) cv::getTickCount() - t0) / cv::getTickFrequency();
std::cout << "Time for method7: " << time << std::endl;
}
Time for method1: 0.586296
Time for method2: 0.414636
Time for method3: 0.504703
Time for method4: 0.207575
Time for method5: 0.111554
Time for method6: 0.0940078
Time for method7: 0.522204
对比这几种方式我们可以发现，最为高效的还是直接使用指针计算地址偏移量, 然而这种方式必须保证Mat在内存的存储是连续的，可以通过cv::Mat::isContinous()函数检测，如果是连续的则可以处理为单行向量，使用最为高效的方式访问。如果不想这么麻烦，其实method5是一种较为可取的方式，通过从cv::Mat::ptr()得到每一行的首地址，这样就不需要保证连续存储，速度和纯粹使用指针也差不了多少。
实际上对于method5，不使用中间指针进行改写的话:
void method5(cv::Mat img){
// raw pointer
double t0 = (double) cv::getTickCount();
int height = img.rows;
int width = img.cols;
int sum=0;
for(int row=0; row < height; row++){
const uchar *ptr = img.ptr(row);
for(int col=0; col < width; col++){
// const uchar *uc_pixel = ptr;
// int a = uc_pixel[0];
// int b = uc_pixel[1];
// int c = uc_pixel[2];
// 不使用中间指针
int a = ptr[0];
int b = ptr[1];
int c = ptr[2];
sum += a + b + c;
ptr += 3;
}
}
assert(sum==3*height*width);
double time = ((double) cv::getTickCount() - t0) / cv::getTickFrequency();
std::cout << "Time for method5: " << time << std::endl;
}
重新测试下：
Time for method1: 0.58601
Time for method2: 0.416404
Time for method3: 0.507943
Time for method4: 0.208068
Time for method5: 0.0918915
Time for method6: 0.0917811
Time for method7: 0.523099
时间上已经十分接近method6，实际操作的时候直接使用method5，不使用中间指针即可。
网友评论