OpenMP并行编程应用—加速OpenCV图像拼接算法.doc

OpenMP并行编程应用加速OpenCV图像拼接算法OpenMP是一种应用于多处理器程序设计的并行编程处理方案，它提供了对于并行编程的高层抽象，只需要在程序中添加简单的指令，就可以编写高效的并行程序，而不用关心具体的并行实现细节，降低了并行编程的难度和复杂度。也正因为OpenMP的简单易用性，它并不适合于需要复杂的线程间同步和互斥的场合。OpenCV中使用Sift或者Surf特征进行图像拼接的算法，需要分别对两幅或多幅图像进行特征提取和特征描述，之后再进行图像特征点的配对，图像变换等操作。不同图像的特征提取和描述的工作是整个过程中最耗费时间的，也是独立 运行的，可以使用OpenMP进行加速。以下是不使用OpenMP加速的Sift图像拼接原程序：cpp view plain copy print?在CODE上查看代码片派生到我的代码片#include highgui/highgui.hpp #include opencv2/nonfree/nonfree.hpp #include opencv2/legacy/legacy.hpp #include omp.h using namespace cv; /计算原始图像点位在经过矩阵变换后在目标图像上对应位置 Point2f getTransformPoint(const Point2f originalPoint, const Mat &transformMaxtri); int main(int argc, char *argv) float startTime = omp_get_wtime(); Mat image01 = imread(Test01.jpg); Mat image02 = imread(Test02.jpg); imshow(拼接图像1, image01); imshow(拼接图像2, image02); /灰度图转换 Mat image1, image2; cvtColor(image01, image1, CV_RGB2GRAY); cvtColor(image02, image2, CV_RGB2GRAY); /提取特征点 SiftFeatureDetector siftDetector(800); / 海塞矩阵阈值 vector keyPoint1, keyPoint2; siftDetector.detect(image1, keyPoint1); siftDetector.detect(image2, keyPoint2); /特征点描述，为下边的特征点匹配做准备 SiftDescriptorExtractor siftDescriptor; Mat imageDesc1, imageDesc2; siftDpute(image1, keyPoint1, imageDesc1); siftDpute(image2, keyPoint2, imageDesc2); float endTime = omp_get_wtime(); std:cout 不使用OpenMP加速消耗时间： endTime - startTime std:endl; /获得匹配特征点，并提取最优配对 FlannBasedMatcher matcher; vector matchePoints; matcher.match(imageDesc1, imageDesc2, matchePoints, Mat(); sort(matchePoints.begin(), matchePoints.end(); /特征点排序 /获取排在前N个的最优匹配特征点 vector imagePoints1, imagePoints2; for (int i = 0; i 10; i+) imagePoints1.push_back(keyPoint1matchePointsi.queryIdx.pt); imagePoints2.push_back(keyPoint2matchePointsi.trainIdx.pt); /获取图像1到图像2的投影映射矩阵，尺寸为3*3 Mat homo = findHomography(imagePoints1, imagePoints2, CV_RANSAC); Mat adjustMat = (Mat_(3, 3) 1.0, 0, image01.cols, 0, 1.0, 0, 0, 0, 1.0); Mat adjustHomo = adjustMat*homo; /获取最强配对点在原始图像和矩阵变换后图像上的对应位置，用于图像拼接点的定位 Point2f originalLinkPoint, targetLinkPoint, basedImagePoint; originalLinkPoint = keyPoint1matchePoints0.queryIdx.pt; targetLinkPoint = getTransformPoint(originalLinkPoint, adjustHomo); basedImagePoint = keyPoint2matchePoints0.trainIdx.pt; /图像配准 Mat imageTransform1; warpPerspective(image01, imageTransform1, adjustMat*homo, Size(image02.cols + image01.cols + 110, image02.rows); /在最强匹配点左侧的重叠区域进行累加，是衔接稳定过渡，消除突变 Mat image1Overlap, image2Overlap; /图1和图2的重叠部分 image1Overlap = imageTransform1(Rect(Point(targetLinkPoint.x - basedImagePoint.x, 0), Point(targetLinkPoint.x, image02.rows); image2Overlap = image02(Rect(0, 0, image1Overlap.cols, image1Overlap.rows); Mat image1ROICopy = image1Overlap.clone(); /复制一份图1的重叠部分 for (int i = 0; i image1Overlap.rows; i+) for (int j = 0; j image1Overlap.cols; j+) double weight; weight = (double)j / image1Overlap.cols; /随距离改变而改变的叠加系数 image1Overlap.at(i, j)0 = (1 - weight)*image1ROICopy.at(i, j)0 + weight*image2Overlap.at(i, j)0; image1Overlap.at(i, j)1 = (1 - weight)*image1ROICopy.at(i, j)1 + weight*image2Overlap.at(i, j)1; image1Overlap.at(i, j)2 = (1 - weight)*image1ROICopy.at(i, j)2 + weight*image2Overlap.at(i, j)2; Mat ROIMat = image02(Rect(Point(image1Overlap.cols, 0), Point(image02.cols, image02.rows); /图2中不重合的部分 ROIMat.copyTo(Mat(imageTransform1, Rect(targetLinkPoint.x, 0, ROIMat.cols, image02.rows); /不重合的部分直接衔接上去 namedWindow(拼接结果, 0); imshow(拼接结果, imageTransform1); imwrite(D:拼接结果.jpg, imageTransform1); waitKey(); return 0; /计算原始图像点位在经过矩阵变换后在目标图像上对应位置 Point2f getTransformPoint(const Point2f originalPoint, const Mat &transformMaxtri) Mat originelP, targetP; originelP = (Mat_(3, 1) originalPoint.x, originalPoint.y, 1.0); targetP = transformMaxtri*originelP; float x = targetP.at(0, 0) / targetP.at(2, 0); float y = targetP.at(1, 0) / targetP.at(2, 0); return Point2f(x, y); 之后在程序中加入OpenMP的头文件“omp.h”就可以了：cpp view plain copy print?在CODE上查看代码片派生到我的代码片#include highgui/highgui.hpp #include opencv2/nonfree/nonfree.hpp #include opencv2/legacy/legacy.hpp #include omp.h using namespace cv; /计算原始图像点位在经过矩阵变换后在目标图像上对应位置 Point2f getTransformPoint(const Point2f originalPoint, const Mat &transformMaxtri); int main(int argc, char *argv) float startTime = omp_get_wtime(); Mat image01, image02; Mat image1, image2; vector keyPoint1, keyPoint2; Mat imageDesc1, imageDesc2; SiftFeatureDetector siftDetector(800); / 海塞矩阵阈值 SiftDescriptorEactor siftDescriptor; /使用OpenMP的sections制导指令开启多线程 #pragma omp parallel sections #pragma omp section image01 = imread(Test01.jpg); imshow(拼接图像1, image01); /灰度图转换 cvtColor(image01, image1, CV_RGB2GRAY); /提取特征点 siftDetector.detect(image1, keyPoint1); /特征点描述，为下边的特征点匹配做准备 siftDpute(image1, keyPoint1, imageDesc1); #pragma omp section image02 = imread(Test02.jpg); imshow(拼接图像2, image02); cvtColor(image02, image2, CV_RGB2GRAY); siftDetector.detect(image2, keyPoint2); siftDpute(image2, keyPoint2, imageDesc2); float endTime = omp_get_wtime(); std:cout 使用OpenMP加速消耗时间： endTime - startTime std:endl; /获得匹配特征点，并提取最优配对 FlannBasedMatcher matcher; vector matchePoints; matcher.match(imageDesc1, imageDesc2, matchePoints, Mat(); sort(matchePoints.begin(), matchePoints.end(); /特征点排序 /获取排在前N个的最优匹配特征点 vector imagePoints1, imagePoints2; for (int i = 0; i 10; i+) imagePoints1.push_back(keyPoint1matchePointsi.queryIdx.pt); imagePoints2.push_back(keyPoint2matchePointsi.trainIdx.pt); /获取图像1到图像2的投影映射矩阵，尺寸为3*3 Mat homo = findHomography(imagePoints1, imagePoints2, CV_RANSAC); Mat adjustMat = (Mat_(3, 3) 1.0, 0, image01.cols, 0, 1.0, 0, 0, 0, 1.0); Mat adjustHomo = adjustMat*homo; /获取最强配对点在原始图像和矩阵变换后图像上的对应位置，用于图像拼接点的定位 Point2f originalLinkPoint, targetLinkPoint, basedImagePoint; originalLinkPoint = keyPoint1matchePoints0.queryIdx.pt; targetLinkPoint = getTransformPoint(originalLinkPoint, adjustHomo); basedImagePoint = keyPoint2matchePoints0.trainIdx.pt; /图像配准 Mat imageTransform1; warpPerspective(image01, imageTransform1, adjustMat*homo, Size(image02.cols + image01.cols + 110, image02.rows); /在最强匹配点左侧的重叠区域进行累加，是衔接稳定过渡，消除突变 Mat image1Overlap, image2Overlap; /图1和图2的重叠部分 image1Overlap = imageTransform1(Rect(Point(targetLinkPoint.x - basedImagePoint.x, 0), Point(targetLinkPoint.x, image02.rows); image2Overlap = image02(Rect(0, 0, image1Overlap.cols, image1Overlap.rows); Mat image1ROICopy = image1Overlap.clone(); /复制一份图1的重叠部分 for (int i = 0; i image1Overlap.rows; i+) for (int j = 0; j image1Overlap.cols; j+) double weight; weight = (double)j / image1Overlap.cols; /随距离改变而改变的叠加系数 image1Overlap.at(i, j)0 = (1 - weight)*image1ROICopy.at(i, j)0 + weight*image2Overlap.at(i, j)0; image1Overlap.at(i, j)1 = (1 - weight)*image1ROICopy.at(i, j)1 + weight*image2Overlap.at(i, j)1; image1Overlap.at(i, j)2 = (1 - weight)*image1ROICopy.at(i, j)2 + weight*image2Overlap.at(i, j)2; Mat ROIMat = image02(Rect(Point(image1Overlap.cols, 0), Point(image02.cols,