Episode Seven：从SIFT算法到dense SIFT算法

た入场券 2022-06-04 05:24 217阅读 0赞

**用SIFT算法出现的问题：**

在处理牙齿有关的数据时，由于牙齿的特征较少，用一般的SIFT算法求得的特征点是稀疏的（Sparse），导致处理其他特征较明显的数据时可以得到完美的结果，但是处理牙齿数据总是会得到错误的H矩阵，导致得到的图C是错误的。

**转战dense SIFT算法：**

dense SIFT算法可以得到相对稠密的特征点，则可以通过检测到的足够多的特征点进行特征匹配，并通过RANSAC方法进行筛选，得到相对正确的变换矩阵H，并得到正确的图C。

code：

///SIFT特征点匹配
    #include "opencv2/opencv.hpp"
    #include "opencv2/nonfree/nonfree.hpp"//SIFT相关
    #include "opencv2/legacy/legacy.hpp"//匹配器相关
    #include <opencv2/nonfree/features2d.hpp>
    #include <iostream>
    
    using namespace std;
    using namespace cv;
    
    int main()
    {
    	//1.SIFT特征点提取——detect()方法  
    	Mat srcImg1 = imread("pic/test2/a.jpg", CV_LOAD_IMAGE_COLOR);
    	Mat srcImg2 = imread("pic/test2/b.jpg", CV_LOAD_IMAGE_COLOR);
    	Mat srcImg1_copy = srcImg1;
    	Mat srcImg2_copy = srcImg2;
    	Mat dstImg1, dstImg2;
    	DenseFeatureDetector siftDetector;//SiftFeatureDetector是SIFT类的别名    
    	vector<KeyPoint> keyPoints1;
    	vector<KeyPoint> keyPoints2;
    	/*siftDetector.detect(srcImg1, keyPoints1);
    	siftDetector.detect(srcImg2, keyPoints2);
    	drawKeypoints(srcImg1, keyPoints1, dstImg1);
    	drawKeypoints(srcImg2, keyPoints2, dstImg2);
    	imshow("dstImg1", dstImg1);
    	imshow("dstImg2", dstImg2);
    	*/
    	//2.特征点描述符(特征向量)提取——compute()方法  
    	SiftDescriptorExtractor descriptor;//SiftDescriptorExtractor是SIFT类的别名    
    	Mat description1;
    	Mat description2;
    	
    
    
    	int step = 10; // 10 pixels spacing between kp's
    
    	
    	for (int i = step; i<srcImg1.rows - step; i += step)
    	{
    		for (int j = step; j<srcImg1.cols - step; j += step)
    		{
    			// x,y,radius
    			keyPoints1.push_back(KeyPoint(float(j), float(i), float(step)));
    		}
    	}
    
    	for (int i = step; i<srcImg2.rows - step; i += step)
    	{
    		for (int j = step; j<srcImg2.cols - step; j += step)
    		{
    			// x,y,radius
    			keyPoints2.push_back(KeyPoint(float(j), float(i), float(step)));
    		}
    	}
    
    	//Ptr<xfeatures2d::SIFT> sift = xfeatures2d::SIFT::create();
        //sift->compute(img, kps, features);
    
    
    
    
    	descriptor.compute(srcImg1, keyPoints1, description1);//注意原图和特征点要对应，不要写错  
    	descriptor.compute(srcImg2, keyPoints2, description2);
    	imshow("description1", description1);
    	imshow("description2", description2);
    	//3.使用暴力匹配器进行暴力匹配——BruteForceMatcher类的match()方法  
    	BruteForceMatcher<L2<float>> matcher;//实例化暴力匹配器  
    	vector<DMatch> matches; //定义匹配结果变量  
    	matcher.match(description1, description2, matches);//实现描述符之间的匹配  
    	
    	//4.对匹配结果进行筛选（依据DMatch结构体中的float类型变量distance进行筛选）  
    	float minDistance = 100;
    	float maxDistance = 0;
    	for (int i = 0; i < matches.size(); i++)
    	{
    		if (matches[i].distance < minDistance)
    			minDistance = matches[i].distance;
    		if (matches[i].distance > maxDistance)
    			maxDistance = matches[i].distance;
    	}
    	cout << "minDistance: " << minDistance << endl;
    	cout << "maxDistance: " << maxDistance << endl;
    
    	vector<DMatch> goodMatches;
    	for (int i = 0; i < matches.size(); i++)
    	{
    		if (matches[i].distance < 3 * minDistance)
    		{
    			goodMatches.push_back(matches[i]);
    		}
    	}
    	//5.绘制匹配结果——drawMatches()  
    	Mat dstImg3;
    	drawMatches(srcImg1, keyPoints1, srcImg2, keyPoints2, goodMatches, dstImg3);
    	namedWindow("dstImg3", WINDOW_NORMAL);//为防止图像太大，屏幕显示不下，采用可以手动拉伸的方法；或者可以采用提前把图片缩小的方法
    	imshow("dstImg3", dstImg3);
    
    
    	vector<Point2f> obj;
    	vector<Point2f> scene;
    
    	for (int i = 0; i < goodMatches.size(); i++)
    	{
    		obj.push_back(keyPoints1[goodMatches[i].queryIdx].pt);
    		scene.push_back(keyPoints2[goodMatches[i].trainIdx].pt);
    	}
    
    
    
    	/*
    	//计算原图像和目的图像之间的Homography  //用粗线标出要检测的物体
    	Mat H = findHomography(obj , scene, CV_RANSAC);
    
    	vector<Point2f>obj_corners(4);
    	obj_corners[0] = cvPoint(0, 0);
    	obj_corners[1] = cvPoint(srcImg1.cols, 0);
    	obj_corners[2] = cvPoint(srcImg1.cols, srcImg1.rows);
    	obj_corners[3] = cvPoint(0, srcImg1.rows);
    
    	vector<Point2f>scene_corners(4);
    	perspectiveTransform(obj_corners, scene_corners, H);
    
    	line(dstImg3, scene_corners[0] + Point2f(static_cast<float>(srcImg1.cols), 0), scene_corners[1] + Point2f(static_cast<float>(srcImg1.cols), 0), Scalar(255, 0, 123), 4);
    
    	line(dstImg3, scene_corners[1] + Point2f(static_cast<float>(srcImg1.cols), 0), scene_corners[2] + Point2f(static_cast<float>(srcImg1.cols), 0), Scalar(255, 0, 123), 4);
    
    	line(dstImg3, scene_corners[2] + Point2f(static_cast<float>(srcImg1.cols), 0), scene_corners[3] + Point2f(static_cast<float>(srcImg1.cols), 0), Scalar(255, 0, 123), 4);
    
    	line(dstImg3, scene_corners[3] + Point2f(static_cast<float>(srcImg1.cols), 0), scene_corners[0] + Point2f(static_cast<float>(srcImg1.cols), 0), Scalar(255, 0, 123), 4);
    	namedWindow("lalal",WINDOW_NORMAL);
    	imshow("lalal", dstImg3);
    	*/
    
    	Mat H = findHomography(scene,obj,CV_RANSAC);
    	Mat srcImg3;
    	//Size size(500, 500);
    	//原图像变形  
    	warpPerspective(srcImg2_copy, srcImg3, H, srcImg2_copy.size());
    	
    	namedWindow("1", WINDOW_NORMAL);
    	imshow("1", srcImg1_copy);
    
    	namedWindow("2", WINDOW_NORMAL);
    	imshow("2", srcImg2_copy);
    
    	namedWindow("3",WINDOW_NORMAL);
    	imshow("3", srcImg3);
    
    	
    	waitKey(0);
    	return 0;
    }

Smart is the new sexy!