kaze算法的图像配准研究（2）-匹配

在特征点检测过后，完成图像间特征点的匹配是非常重要的。对于图像配准工作而言。特征点匹配的准确度是最值得关注的点，宁愿少匹配，也不能误匹配。

我在此图像配准中使用的是KNN匹配：
下面引用自百度百科。

K最近邻(kNN，k-NearestNeighbor)分类算法是数据挖掘分类技术中最简单的方法之一，所谓K最近邻，就是k个最近的邻居的意思，说的是每个样本都可以用它最接近的k个邻居来代表。

如上图所示，如果k=5，那么对于未知点 xu x_u，离它最近的5个点有4个属于 w1 w_1，1个属于 w3 w_3，那么我们可以认为这个未知点 xu x_u是属于 w1 w_1的。
在特征点匹配的场景中，Lowe设定的匹配规则是1NN，即寻找特征向量离得最近的点进行匹配。但是这样会产生误匹配，因此它又定制了一个规则及1NN/2NN < ratio时，该匹配才生效。Lowe推荐的ratio为0.8。

得到匹配的一系列点后，接下来就是要产生两图像的对应仿射矩阵了。这里采用的是RANSAC法。

最终图像匹配源码为：


#include<iostream>//#include <opencv2/xfeatures2d.hpp>
#include <opencv2/opencv.hpp>using namespace std;
using namespace cv;void main()
{double start, duration_ms;cv::Mat mask1, mask2, imgshowcanny;cv::Mat imgShow1, imgShow2, imgShow3;// 声明并从data文件夹里读取两个rgb与深度图cv::Mat rgb1 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\1.jpg", 0);cv::Mat rgb2 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\2.jpg", 0);cv::Mat rgb3 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\3.jpg", 0);cv::Mat rgb4 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\4.jpg", 0);cv::Mat rgb5 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\5.jpg", 0);cv::Mat rgb6 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\6.jpg", 0);cv::Mat rgb7 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\7.jpg", 0);cv::Mat rgb8 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\8.jpg", 0);cv::Mat rgb9 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\9.jpg", 0);cv::Mat rgb10 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\10.jpg", 0);vector< cv::KeyPoint > kp1, kp2, kp3, kp4, kp5, kp6, kp7, kp8, kp9, kp10; //关键点mask1 = rgb1.clone();//边缘检测//cout << "Canny edge detection" << endl;//mask1 = Mat::zeros(rgb1.size(), CV_8UC1);//mask2 = Mat::zeros(rgb2.size(), CV_8UC1);cv::Canny(mask1, mask1, 50, 200, 3);//cv::Canny(rgb2, mask2, 50, 150, 3);cv::Mat element = getStructuringElement(MORPH_RECT, Size(3, 3));dilate(mask1, mask1, element);cv::imwrite("F:/study work/Visual Odometry dev/VO_practice_2016.4/算法测试/KAZE算法/data/mask1.jpg", mask1);imgshowcanny = mask1.clone();cv::resize(imgshowcanny, imgshowcanny, cv::Size(imgshowcanny.cols / 2, imgshowcanny.rows / 2));cv::imshow("mask", imgshowcanny);//特征检测算法：AKAZEcv::Ptr<cv::AKAZE> akaze = AKAZE::create(AKAZE::DESCRIPTOR_MLDB, 0, 3, 0.001f, 4, 4, KAZE::DIFF_PM_G2);Mat descriptors_1, descriptors_2, descriptors_3, descriptors_4, descriptors_5, descriptors_6, descriptors_7, descriptors_8, descriptors_9, descriptors_10;start = double(getTickCount());akaze->detectAndCompute(rgb1, mask1, kp1, descriptors_1, false);duration_ms = (double(getTickCount()) - start) * 1000 / getTickFrequency();//计时std::cout << "It took " << duration_ms << " ms to detect features in pic1 using AKAZE." << std::endl;start = double(getTickCount());akaze->detectAndCompute(rgb2, cv::Mat(), kp2, descriptors_2, false);duration_ms = (double(getTickCount()) - start) * 1000 / getTickFrequency();//计时std::cout << "It took " << duration_ms << " ms to detect features in pic2 using AKAZE." << std::endl;akaze->detectAndCompute(rgb3, cv::Mat(), kp3, descriptors_3, false);akaze->detectAndCompute(rgb4, cv::Mat(), kp4, descriptors_4, false);akaze->detectAndCompute(rgb5, cv::Mat(), kp5, descriptors_5, false);akaze->detectAndCompute(rgb6, cv::Mat(), kp6, descriptors_6, false);akaze->detectAndCompute(rgb7, cv::Mat(), kp7, descriptors_7, false);akaze->detectAndCompute(rgb8, cv::Mat(), kp8, descriptors_8, false);akaze->detectAndCompute(rgb9, cv::Mat(), kp9, descriptors_9, false);akaze->detectAndCompute(rgb10, cv::Mat(), kp10, descriptors_10, false);// 可视化， 显示特征点cout << "Key points of two AKAZE images: " << kp1.size() << ", " << kp2.size() << endl;cv::drawKeypoints(rgb1, kp1, imgShow1, cv::Scalar::all(-1), cv::DrawMatchesFlags::DRAW_RICH_KEYPOINTS);cv::drawKeypoints(rgb2, kp2, imgShow2, cv::Scalar::all(-1), cv::DrawMatchesFlags::DRAW_RICH_KEYPOINTS);cv::drawKeypoints(rgb3, kp3, imgShow3, cv::Scalar::all(-1), cv::DrawMatchesFlags::DRAW_RICH_KEYPOINTS);cv::imwrite("F:/study work/Visual Odometry dev/VO_practice_2016.4/算法测试/KAZE算法/data/akaze1.jpg", imgShow1);cv::imwrite("F:/study work/Visual Odometry dev/VO_practice_2016.4/算法测试/KAZE算法/data/akaze2.jpg", imgShow2);cv::resize(imgShow2, imgShow2, cv::Size(imgShow2.cols / 2, imgShow2.rows / 2));cv::resize(imgShow1, imgShow1, cv::Size(imgShow1.cols / 2, imgShow1.rows / 2));cv::imshow("AKAZE_keypoints2", imgShow2);cv::imshow("AKAZE_keypoints1", imgShow1);cv::waitKey(0); //暂停等待一个按键//用Matcher匹配特征 :BFMatcher bf_matcher;FlannBasedMatcher flann_matcher;cv::Ptr<cv::DescriptorMatcher>  knn_matcher = cv::DescriptorMatcher::create("BruteForce-Hamming");std::vector< DMatch > flann_matches;std::vector< DMatch > bf_matches1;std::vector< DMatch > knn_match;std::vector< std::vector<cv::DMatch> > knn_matches1, knn_matches2, knn_matches3, knn_matches4, knn_matches5, knn_matches6, knn_matches7, knn_matches8, knn_matches9;start = double(getTickCount());bf_matcher.match(descriptors_1, descriptors_2, bf_matches1);//flann_matcher.match(descriptors_1, descriptors_2, flann_matches);duration_ms = (double(getTickCount()) - start) * 1000 / getTickFrequency();//计时std::cout << "It took " << duration_ms << " ms to do brutal-force-match on AKAZE features." << std::endl;start = double(getTickCount());knn_matcher->knnMatch(descriptors_1, descriptors_2, knn_matches1, 2);duration_ms = (double(getTickCount()) - start) * 1000 / getTickFrequency();//计时std::cout << "It took " << duration_ms << " ms to do knn-match on AKAZE features." << std::endl;knn_matcher->knnMatch(descriptors_1, descriptors_3, knn_matches2, 2);knn_matcher->knnMatch(descriptors_1, descriptors_4, knn_matches3, 2);knn_matcher->knnMatch(descriptors_1, descriptors_5, knn_matches4, 2);knn_matcher->knnMatch(descriptors_1, descriptors_6, knn_matches5, 2);knn_matcher->knnMatch(descriptors_1, descriptors_7, knn_matches6, 2);knn_matcher->knnMatch(descriptors_1, descriptors_8, knn_matches7, 2);knn_matcher->knnMatch(descriptors_1, descriptors_9, knn_matches8, 2);knn_matcher->knnMatch(descriptors_1, descriptors_10, knn_matches9, 2);//用RANSAC求取单应矩阵的方式去掉不可靠的BF匹配vector< DMatch > inliers1, inliers2;cout << "Computing homography (RANSAC)" << endl;vector<Point2f> points1(bf_matches1.size());vector<Point2f> points2(bf_matches1.size());for (size_t i = 0; i < bf_matches1.size(); i++){points1[i] = kp1[bf_matches1[i].queryIdx].pt;points2[i] = kp2[bf_matches1[i].trainIdx].pt;}/*for (int i = 0; i < points1.size(); i++) {cout << "points1_x :" << points1[i].x << endl;cout << "points1_y :" << points1[i].y << endl;}*///计算单应矩阵并找出inliers  vector<uchar> flag1(points1.size(), 0);Mat H1 = findHomography(points1, points2, CV_RANSAC, 3.0, flag1);Mat H2 = findHomography(points2, points1, CV_RANSAC, 3.0, flag1);cout << "points1_shape: " << points1.size() << endl;cout << "points2_shape: " << points2.size() << endl;cout << "H1_shape: " << H2.size() << endl;/*for (int i = 0; i < H1.rows; i++) {for (int j = 0; j < H1.cols; j++) {cout << "H1_value" << H1[i][i] << endl;}}*///std::cout << H2 << std::endl;//cout << H << endl << endl;  for (int i = 0; i < bf_matches1.size(); i++){if (flag1[i]){inliers1.push_back(bf_matches1[i]);}}cout << "AKAZE BF matches inliers size = " << inliers1.size() << endl;//去掉不可靠的HAMMING-KNN匹配vector< cv::DMatch > goodMatches1;for (size_t i = 0; i < knn_matches1.size(); i++){if (knn_matches1[i][0].distance < 0.3  *knn_matches1[i][1].distance)goodMatches1.push_back(knn_matches1[i][0]);}vector< cv::DMatch > goodMatches2;for (size_t i = 0; i < knn_matches2.size(); i++){if (knn_matches2[i][0].distance < 0.3  *knn_matches2[i][1].distance)goodMatches2.push_back(knn_matches2[i][0]);}vector< cv::DMatch > goodMatches3;for (size_t i = 0; i < knn_matches3.size(); i++){if (knn_matches3[i][0].distance < 0.3  *knn_matches3[i][1].distance)goodMatches3.push_back(knn_matches3[i][0]);}vector< cv::DMatch > goodMatches4;for (size_t i = 0; i < knn_matches4.size(); i++){if (knn_matches4[i][0].distance < 0.3  *knn_matches4[i][1].distance)goodMatches4.push_back(knn_matches4[i][0]);}vector< cv::DMatch > goodMatches5;for (size_t i = 0; i < knn_matches5.size(); i++){if (knn_matches5[i][0].distance < 0.3  *knn_matches5[i][1].distance)goodMatches5.push_back(knn_matches5[i][0]);}vector< cv::DMatch > goodMatches6;for (size_t i = 0; i < knn_matches6.size(); i++){if (knn_matches6[i][0].distance < 0.3  *knn_matches6[i][1].distance)goodMatches6.push_back(knn_matches6[i][0]);}vector< cv::DMatch > goodMatches7;for (size_t i = 0; i < knn_matches7.size(); i++){if (knn_matches7[i][0].distance < 0.3  *knn_matches7[i][1].distance)goodMatches7.push_back(knn_matches7[i][0]);}vector< cv::DMatch > goodMatches8;for (size_t i = 0; i < knn_matches8.size(); i++){if (knn_matches8[i][0].distance < 0.3  *knn_matches8[i][1].distance)goodMatches8.push_back(knn_matches8[i][0]);}vector< cv::DMatch > goodMatches9;for (size_t i = 0; i < knn_matches9.size(); i++){if (knn_matches9[i][0].distance < 0.3  *knn_matches9[i][1].distance)goodMatches9.push_back(knn_matches9[i][0]);}cout << "AKAZE good knn matches size = " << goodMatches1.size() << endl;cout << "I am before count point" << endl;vector<Point2f> points3(goodMatches1.size());vector<Point2f> points4(goodMatches1.size());for (size_t i = 0; i < goodMatches1.size(); i++){points3[i] = kp1[goodMatches1[i].queryIdx].pt;points4[i] = kp2[goodMatches1[i].trainIdx].pt;}cout << "I am before count point 1" << endl;vector<Point2f> points5(goodMatches2.size());vector<Point2f> points6(goodMatches2.size());for (size_t i = 0; i < goodMatches2.size(); i++){points5[i] = kp1[goodMatches2[i].queryIdx].pt;points6[i] = kp3[goodMatches2[i].trainIdx].pt;}cout << "I am before count point 2" << endl;vector<Point2f> points7(goodMatches3.size());vector<Point2f> points8(goodMatches3.size());cout << "goodMatchers3.size" << goodMatches3.size() << endl;for (size_t i = 0; i < goodMatches3.size(); i++){points7[i] = kp1[goodMatches3[i].queryIdx].pt;points8[i] = kp4[goodMatches3[i].trainIdx].pt;}cout << "I am before count point 3" << endl;vector<Point2f> points9(goodMatches4.size());vector<Point2f> points10(goodMatches4.size());for (size_t i = 0; i < goodMatches4.size(); i++){points9[i] = kp1[goodMatches4[i].queryIdx].pt;points10[i] = kp5[goodMatches4[i].trainIdx].pt;}cout << "I am before count point 4" << endl;vector<Point2f> points11(goodMatches5.size());vector<Point2f> points12(goodMatches5.size());for (size_t i = 0; i < goodMatches5.size(); i++){points11[i] = kp1[goodMatches5[i].queryIdx].pt;points12[i] = kp6[goodMatches5[i].trainIdx].pt;}vector<Point2f> points13(goodMatches6.size());vector<Point2f> points14(goodMatches6.size());for (size_t i = 0; i < goodMatches6.size(); i++){points13[i] = kp1[goodMatches6[i].queryIdx].pt;points14[i] = kp7[goodMatches6[i].trainIdx].pt;}vector<Point2f> points15(goodMatches7.size());vector<Point2f> points16(goodMatches7.size());for (size_t i = 0; i < goodMatches7.size(); i++){points15[i] = kp1[goodMatches7[i].queryIdx].pt;points16[i] = kp8[goodMatches7[i].trainIdx].pt;}cout << "I am before count point 4" << endl;vector<Point2f> points17(goodMatches8.size());vector<Point2f> points18(goodMatches8.size());for (size_t i = 0; i < goodMatches8.size(); i++){points17[i] = kp1[goodMatches8[i].queryIdx].pt;points18[i] = kp9[goodMatches8[i].trainIdx].pt;}vector<Point2f> points19(goodMatches9.size());vector<Point2f> points20(goodMatches9.size());for (size_t i = 0; i < goodMatches9.size(); i++){points19[i] = kp1[goodMatches9[i].queryIdx].pt;points20[i] = kp10[goodMatches9[i].trainIdx].pt;}cout << "I am after count point" << endl;Mat H3 = findHomography(points4, points3, CV_RANSAC, 3.0, flag1);Mat H4 = findHomography(points6, points5, CV_RANSAC, 3.0, flag1);Mat H5 = findHomography(points8, points7, CV_RANSAC, 3.0, flag1);Mat H6 = findHomography(points10, points9, CV_RANSAC, 3.0, flag1);Mat H7 = findHomography(points12, points11, CV_RANSAC, 3.0, flag1);Mat H8 = findHomography(points14, points13, CV_RANSAC, 3.0, flag1);Mat H9 = findHomography(points16, points15, CV_RANSAC, 3.0, flag1);Mat H10 = findHomography(points18, points17, CV_RANSAC, 3.0, flag1);Mat H11 = findHomography(points20, points19, CV_RANSAC, 3.0, flag1);cv::Mat imdst_2, imdst_3, imdst_4, imdst_5, imdst_6, imdst_7, imdst_8, imdst_9, imdst_10, imsrc_1, imsrc_2, imsrc_3, imsrc_4, imsrc_5, imsrc_6, imsrc_7, imsrc_8, imsrc_9, imsrc_10, warp_dst, imresult;imsrc_1 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\1.jpg");imsrc_2 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\2.jpg");imsrc_3 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\3.jpg");imsrc_4 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\4.jpg");imsrc_5 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\5.jpg");imsrc_6 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\6.jpg");imsrc_7 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\7.jpg");imsrc_8 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\8.jpg");imsrc_9 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\9.jpg");imsrc_10 = cv::imread("C:\\opencv-cpp\\opencvtest\\img_source\\1\\10.jpg");//warpAffine(im_src, warp_dst, H2, im_src.size());warpPerspective(imsrc_2, imdst_2, H3, imsrc_2.size());warpPerspective(imsrc_3, imdst_3, H4, imsrc_3.size());warpPerspective(imsrc_4, imdst_4, H5, imsrc_4.size());warpPerspective(imsrc_5, imdst_5, H6, imsrc_5.size());warpPerspective(imsrc_6, imdst_6, H7, imsrc_6.size());warpPerspective(imsrc_7, imdst_7, H8, imsrc_7.size());warpPerspective(imsrc_8, imdst_8, H9, imsrc_8.size());warpPerspective(imsrc_9, imdst_9, H10, imsrc_9.size());warpPerspective(imsrc_10, imdst_10, H11, imsrc_10.size());cv::Mat imsrc_1_, imdst_2_, imdst_3_, imdst_4_, imdst_5_, imdst_6_, imdst_7_, imdst_8_, imdst_9_, imdst_10_;imsrc_1.convertTo(imsrc_1_, CV_32F, 1.0 / 255.0);imdst_2.convertTo(imdst_2_, CV_32F, 1.0 / 255.0);imdst_3.convertTo(imdst_3_, CV_32F, 1.0 / 255.0);imdst_4.convertTo(imdst_4_, CV_32F, 1.0 / 255.0);imdst_5.convertTo(imdst_5_, CV_32F, 1.0 / 255.0);imdst_6.convertTo(imdst_6_, CV_32F, 1.0 / 255.0);imdst_7.convertTo(imdst_7_, CV_32F, 1.0 / 255.0);imdst_8.convertTo(imdst_8_, CV_32F, 1.0 / 255.0);imdst_9.convertTo(imdst_9_, CV_32F, 1.0 / 255.0);imdst_10.convertTo(imdst_10_, CV_32F, 1.0 / 255.0);imresult = (imsrc_1_ + imdst_2_ + imdst_3_ + imdst_4_ + imdst_5_ + imdst_6_ + imdst_7_ + imdst_8_ + imdst_9_ + imdst_10_) / 10;cv::imshow("result", imresult);cv::imwrite("D:\\opencv_cpp\\opencv_cpp\\source_images\\after_transactions\\2_after_transaction.jpg", imdst_2);cv::imwrite("D:\\opencv_cpp\\opencv_cpp\\source_images\\after_transactions\\3_after_transaction.jpg", imdst_3);// 可视化：显示匹配的特征cv::Mat AKAZE_BF_imgMatches;cv::Mat AKAZE_KNN_imgMatches; 筛选匹配，把距离太大的去掉 这里使用的准则是去掉大于四倍最小距离的匹配//vector< cv::DMatch > goodMatches;//double minDis = 9999;//for (size_t i = 0; i<matches.size(); i++)//{//  if (matches[i].distance < minDis)//      minDis = matches[i].distance;//}//for (size_t i = 0; i<matches.size(); i++)//{//  if (matches[i].distance < 8 * minDis)//      goodMatches.push_back(matches[i]);//}// 显示 good matchescv::drawMatches(rgb1, kp1, rgb2, kp2, inliers1, AKAZE_BF_imgMatches);cv::imwrite("F:/study work/Visual Odometry dev/VO_practice_2016.4/算法测试/KAZE算法/data/AKAZE-BF-Matches-inliers.jpg", AKAZE_BF_imgMatches);cv::drawMatches(rgb1, kp1, rgb2, kp2, goodMatches1, AKAZE_KNN_imgMatches);cv::imwrite("F:/study work/Visual Odometry dev/VO_practice_2016.4/算法测试/KAZE算法/data/AKAZE-Good-KNN-Matches.jpg", AKAZE_KNN_imgMatches);cv::resize(AKAZE_BF_imgMatches, AKAZE_BF_imgMatches,
Size(AKAZE_BF_imgMatches.cols / 2, AKAZE_BF_imgMatches.rows / 2));cv::imshow("AKAZE-BF-Matches-inliers", AKAZE_BF_imgMatches);cv::resize(AKAZE_KNN_imgMatches, AKAZE_KNN_imgMatches, Size(AKAZE_KNN_imgMatches.cols / 2, AKAZE_KNN_imgMatches.rows / 2));cv::imshow("AKAZE-Good-KNN-Matches", AKAZE_KNN_imgMatches);cv::waitKey(0);
}

代码中选用了10张图片进行配准。最后结果为：

kaze算法的图像配准研究（2）-匹配相关推荐

kaze算法的图像配准研究（1）-KAZE算法原理
一:KAZE算法的由来在2012年,ECCV会议中出现了一种比SIFT更稳定的特征检测算法.KAZE的取名是为了纪念尺度空间分析的开创者-日本学者Iijima.KAZE在日语中是'风'的谐音,寓意是 ...
SIFT-FCACO算法的图像配准
SIFT-FCACO算法的图像配准 2014年微型机与应用第15期作者:吴金津,文志强,龙永新,武岫缘 2015/6/1 19:12:00 关键词: SIFT-FCACO算法蚁群算法 RANSAC ...
c语言编程图像拼接,一种基于Lucas-Kanade算法的图像配准和拼接方法
一种基于Lucas-Kanade算法的图像配准和拼接方法 [技术领域] [0001 ]本发明涉及图像处理技术领域,具体涉及一种基于Lucas-Kanade算法的图像配准和拼接方法. [背景技术] [ ...
图像配准系列之基于FFD形变与粒子群算法的图像配准
在之前的文章中,我们分别使用了梯度下降发与LM算法来优化FFD形变的控制参数,达到图像配准的目的: 图像配准系列之基于FFD形变与梯度下降法的图像配准图像配准系列之基于FFD形变与LM算法的图像配准 ...
python实现基于SIFT算法的图像配准（仿射变换）
话不多说,直接上代码,可以用的话别忘了请喝可乐!(手动笑哭脸) [用法] 第45.46行的输入: img1 = cv2.imread('sift/3.jpg') img2 = cv2.imread(' ...
【图像配准】多图配准/不同特征提取算法/匹配器比较测试
前言本文首先完成之前专栏前置博文未完成的多图配准拼接任务,其次对不同特征提取器/匹配器效率进行进一步实验探究. 各类算法原理简述看到有博文[1]指出,在速度方面SIFT<SURF<BR ...
图像配准的前世今生：从人工设计特征到深度学习
点击上方"小白学视觉",选择加"星标"或"置顶" 重磅干货,第一时间送达机器之心编译参与:Nurhachu Null,Geek AI 作 ...
【计算机视觉】图像配准(Image Registration)
(Source:https://blog.sicara.com/image-registration-sift-deep-learning-3c794d794b7a) 图像配准方法概述图像配准广泛 ...
机器学习笔记 - 基于传统方法/深度学习的图像配准
一.图像配准图像配准是将一个场景的不同图像变换到同一坐标系的过程.这些图像可以在不同的时间(多时间配准).由不同的传感器(多模态配准)和/或从不同的视点拍摄.这些图像之间的空间关系可以是刚性的 ...

kaze算法的图像配准研究（2）-匹配

kaze算法的图像配准研究（2）-匹配相关推荐

最新文章

热门文章