opencv(C++)四种不同访问方式及速度对比

摘要：

比较一下opencv四种不同的访问方式的效率，多次测试的结论就是用指针的方式是最快的，方法2、3、4都是指针的方式，在release模式下，方法2、方法3、方法4很接近，没多少差别，在不同尺度下稍微各有一点点优劣，我个人常用方法3，因为它简洁高效，方法2看起来似乎也不错，方法4需要图像数据是连续的才能用。

假设：图像image为3通道8bit的图像，现在要访问它的第row行，第col列的r,g,b值

1. 方法一：Mat.at< vec >(row,col)

//Mat.at<vec>(i,j)方式访问
cv::Vec3b bgr = image.at<cv::Vec3b>(row, col);
blue  = bgr[0];
green = bgr[1];
red   = bgr[2];

2. 方法二：Mat.ptr< vec >(row)

//Mat.at<vec>(i,j)方式访问
cv::Vec3b *bgr = image.ptr<cv::Vec3b>(row);
blue  = bgr[col][0];
green = bgr[col][1];
red   = bgr[col][2];

3. 方法三：Mat.ptr< uchar >(row)

uchar *ptr = image.ptr<uchar>(row);//获得图像第row行的首地址
blue  = *(ptr + col*3 );
green = *(ptr + col*3 + 1);
red   = *(ptr + col*3 + 2);

4. 方法四：Mat.data + 偏移量

//image.data是图像矩阵的首地址
uchar *ptr = image.data + row * image.cols * 3 +  col*3;
blue  = *(ptr);
green = *(ptr + 1);
red   = *(ptr + 2);

注：其他数据类型的访问可参考https://www.jianshu.com/p/cfe373dc8c95

测试

图片尺寸427 * 640 * 3，重复颜色反转操作99次，显示平均时间在图片上

• debug模式下
  

• release模式下
  

完整源码

#include "opencv.hpp"
#include "imageProcess.h"

#define TIMES 1

//Mat.at<vec>(i,j)方式访问
void method_1(cv::Mat image)
{
	//cv::Mat image = image_.clone();
	double startTime = cv::getTickCount();
	int w = image.cols;
	int h = image.rows;
	for (int times = 0; times < TIMES; times++)
	{
		for (int row = 0; row < h; row++)
		{
			for (int col = 0; col < w; col++)
			{
				image.at<cv::Vec3b>(row, col)[0] = 255 - image.at<cv::Vec3b>(row, col)[0];
				image.at<cv::Vec3b>(row, col)[1] = 255 - image.at<cv::Vec3b>(row, col)[1];
				image.at<cv::Vec3b>(row, col)[2] = 255 - image.at<cv::Vec3b>(row, col)[2];
			}
		}
	}
	double endTime = cv::getTickCount();
	double t = ((endTime - startTime) / cv::getTickFrequency()) * 1000 / TIMES;
	std::ostringstream ss;
	ss << "Execute time : " << std::fixed << std::setprecision(2) << t << " ms ";
	putText(image, ss.str(), cv::Point(20, 20), cv::FONT_HERSHEY_SIMPLEX, 0.75, cv::Scalar(0, 0, 255), 2, 8);
	imshow("method1", image);
}

//Mat.ptr<vec3b>(row);
void method_2(cv::Mat image)
{
	//cv::Mat image = image_.clone();
	//cv::imshow("2", image);
	double startTime = cv::getTickCount();
	int w = image.cols;
	int h = image.rows;
	cv::Vec3b* curr=NULL;
	for (int times = 0; times < TIMES; times++)
	{
		for (int row = 0; row < h; row++) 
		{
			curr = image.ptr<cv::Vec3b>(row);
			for (int col = 0; col < w; col++) 
			{
				curr[col][0] = 255 - curr[col][0];
				curr[col][1] = 255 - curr[col][1];
				curr[col][2] = 255 - curr[col][2];
			}
		}
	}
	
	double endTime = cv::getTickCount();
	double t = ((endTime - startTime) / cv::getTickFrequency()) * 1000 / TIMES;
	std::ostringstream ss;
	ss << "Execute time : " << std::fixed << std::setprecision(2) << t << " ms ";
	putText(image, ss.str(), cv::Point(20, 20), cv::FONT_HERSHEY_SIMPLEX, 0.75, cv::Scalar(0, 0, 255), 2, 8);
	imshow("method2", image);
}


void method_3(cv::Mat image)
{
	//cv::Mat image = image_.clone();
	double startTime = cv::getTickCount();
	int w = image.cols;
	int h = image.rows;
	uchar * ptr=NULL;//指针定义应该放在这里，如果放在for循环中定义会消耗一部分变量分配时间；
	for (int times = 0; times < TIMES; times++)
	{
		for (int row = 0; row < h; row++) 
		{
			ptr = image.ptr<uchar>(row);
			for (int col = 0; col < w; col++)
			{
				*(ptr + 3 * col) = 255 - *(ptr + 3 * col);
				*(ptr + 3 * col + 1) = 255 - *(ptr + 3 * col + 1);
				*(ptr + 3 * col + 2) = 255 - *(ptr + 3 * col + 2);
			}
		}
	}
	
	double endTime = cv::getTickCount();
	double t = ((endTime - startTime) / cv::getTickFrequency()) * 1000 / TIMES;
	std::ostringstream ss;
	ss << "Execute time : " << std::fixed << std::setprecision(2) << t << " ms ";
	putText(image, ss.str(), cv::Point(20, 20), cv::FONT_HERSHEY_SIMPLEX, 0.75, cv::Scalar(0, 0, 255), 2, 8);
	imshow("method3", image);
}

//当图像数据是连续的，就可以从矩阵的第一个元素开始
//按相对位置访问整个矩阵
void method_4(cv::Mat image)
{
	//assert(image.isContinuous());
	//std::cout <<"image.isContinuous():"<< image.isContinuous()<< std::endl;

	double startTime = cv::getTickCount();
	int w = image.cols;
	int h = image.rows;
	uchar * imgPtr = NULL;
	for (int times = 0; times < TIMES; times++)
	{
		//image.data与image.ptr<uchar>(0)等价；
		for (int row = 0; row < h; row++) {
			imgPtr = image.data + row * image.step;
			for (int col = 0; col < w; col++) {
				//以下三种方式都是等价的，
				*(imgPtr + col*3) = 255 - *(imgPtr + col * 3);
				*(imgPtr + col * 3 + 1) = 255 - *(imgPtr + col * 3 + 1);
				*(imgPtr + col * 3 + 2) = 255 - *(imgPtr + col * 3 + 2);

				/*imgPtr[col * 3] = 255 - imgPtr[col * 3];
				imgPtr[col * 3 + 1] = 255 - imgPtr[col * 3 + 1];
				imgPtr[col * 3 + 2] = 255 - imgPtr[col * 3 + 2];*/

				/**(imgPtr++) = 255 - *(imgPtr);
				*(imgPtr++) = 255 - *(imgPtr);
				*(imgPtr++) = 255 - *(imgPtr);*/
			}
		}
	}
	double endTime = cv::getTickCount();
	double t = ((endTime - startTime) / cv::getTickFrequency()) * 1000 / TIMES;
	std::ostringstream ss;
	ss << "Execute time : " << std::fixed << std::setprecision(2) << t << " ms ";
	putText(image, ss.str(), cv::Point(20, 20), cv::FONT_HERSHEY_SIMPLEX, 0.75, cv::Scalar(0, 0, 255), 2, 8);
	imshow("method4", image);
}


int main()
{
	std::string imagePath = "K:\\deepImage\\building.jpg";
	cv::Mat m1 = cv::imread(imagePath);
	//cv::resize(m1, m1, cv::Size(1000, 1000));
	method_1(m1.clone());
	method_2(m1.clone());
	method_3(m1.clone());
	method_4(m1.clone());
	cv::waitKey(0);

	return 0;
}