OpenCV4之特征提取与对象检测

1、图像特征概述

图像特征的定义与表示

图像特征表示是该图像唯一的表述，是图像的DNA

图像特征提取概述

• 传统图像特征提取 - 主要基于纹理、角点、颜色分布、梯度、边缘等
• 深度卷积神经网络特征提取 - 基于监督学习、自动提取特征
• 特征数据/特征属性
• • 尺度空间不变性
  • 像素迁移不变性
  • 光照一致性原则
  • 旋转不变性原则

图像特征应用

图像分类、对象识别、特征检测、图像对齐/匹配、对象检测、图像搜索/比对

• 图像处理：从图像到图像
• 特征提取：从图像到向量（数据）

2、角点检测

• 什么是角点

  

• 各个方向的梯度变化

  

• Harris角点检测算法

  

  //函数说明：
  void cv::cornerHarris(
  	InputArray src,  //输入
      OutputArray dst,  //输出
      int blockSize,  //块大小
      int ksize,  //Sobel
      double k,  //常量系数
      int borderType = BORDER_DEFAULT  //
  )
  

• Shi-tomas角点检测算法

  

  //函数说明：
  void cv::goodFeaturesToTrack(
  	InputArray image,  //输入图像
      OutputArray corners,  //输出的角点坐标
      int maxCorners,  //最大角点数目
      double qualityLevel,  //质量控制，即λ1与λ2的最小阈值
      double minDistance,  //重叠控制，忽略多少像素值范围内重叠的角点
      InputArray mask = noArray(),
      int blockSize = 3,
      bool useHarrisDetector = false,
      double k = 0.04
  )
  

• 代码实现

  #include <opencv2/opencv.hpp>
  #include <iostream>
  
  using namespace cv;
  using namespace std;
  
  int main(int argc, char** argv) {
  
  	Mat src = imread("D:/images/building.png");
  	Mat gray;
  	cvtColor(src, gray, COLOR_BGR2GRAY);
  	namedWindow("src", WINDOW_FREERATIO);
  	imshow("src", src);
  
  	RNG rng(12345);
  	vector<Point> points;
  	goodFeaturesToTrack(gray, points, 400, 0.05, 10);
  	for (size_t t = 0; t < points.size(); t++) {
  		circle(src, points[t], 3, Scalar(rng.uniform(0, 255), rng.uniform(0, 255), rng.uniform(0, 255)), 1.5, LINE_AA);
  	}
  
  	namedWindow("out", WINDOW_FREERATIO);
  	imshow("out", src);
  
  	waitKey(0);
  	destroyAllWindows();
  
  	return 0;
  }
  

• 效果：
  
  

3、关键点检测

• 图像特征点/关键点
• 关键点检测函数
• 代码演示

ORB关键点检测（快速）

• ORB算法由两个部分组成：快速关键点定位+BRIEF描述子生成

• Fast关键点检测：选择当前像素点P，阈值T，周围16个像素点，超过连续N=12个像素点大于或者小于P，Fast1：优先检测1、5、9、13，循环所有像素点

  

关键点检测函数

//ORB对象创建
Orb = cv::ORB::create(500)

virtual void cv::Feature2D::detect(
	InputArray image,  //输入图像
    std::vector<KeyPoint>& keypoints,  //关键点
    InputArray mask = noArray()  //支持mask
)

KeyPoint数据结构-四个最重要属性：

• pt
• angle
• response
• size

代码实现：

#include <opencv2/opencv.hpp>
#include <iostream>

using namespace cv;
using namespace std;

int main(int argc, char** argv) {
	Mat src = imread("D:/images/building.png");
	imshow("input", src);
	//Mat gray;
	//cvtColor(src, gray, COLOR_BGR2GRAY);

	auto orb = ORB::create(500);
	vector<KeyPoint> kypts;
	orb->detect(src, kypts);

	Mat result01, result02, result03;
	drawKeypoints(src, kypts, result01, Scalar::all(-1), DrawMatchesFlags::DEFAULT);
	drawKeypoints(src, kypts, result02, Scalar::all(-1), DrawMatchesFlags::DRAW_RICH_KEYPOINTS);

	imshow("ORB关键点检测default", result01);
	imshow("ORB关键点检测rich", result02);
	waitKey(0);
	destroyAllWindows();
	return 0;
}

效果：

4、特征描述子

• 基于关键点周围区域
• 浮点数表示与二值编码
• 描述子长度

ORB特征描述子生成步骤：

• 提取特征关键点
• 描述子方向指派
• 特征描述子编码（二值编码32位）

代码实现：

#include <opencv2/opencv.hpp>
#include <iostream>

using namespace cv;
using namespace std;

int main(int argc, char** argv) {
	Mat src = imread("D:/images/building.png");
	imshow("input", src);
	//Mat gray;
	//cvtColor(src, gray, COLOR_BGR2GRAY);

	auto orb = ORB::create(500);    //获取500个关键点，每个关键点计算一个orb特征描述子
	vector<KeyPoint> kypts;
	orb->detect(src, kypts);

	Mat result01, result02, result03;
	//drawKeypoints(src, kypts, result01, Scalar::all(-1), DrawMatchesFlags::DEFAULT);
    //不同半径代表不同层级高斯金字塔中的关键点，即图像不同尺度中的关键点
	drawKeypoints(src, kypts, result02, Scalar::all(-1), DrawMatchesFlags::DRAW_RICH_KEYPOINTS);

	Mat desc_orb;
	orb->compute(src, kypts, desc_orb);

	std::cout << desc_orb.rows << " x " << desc_orb.cols << std::endl;

	//imshow("ORB关键点检测default", result01);
	imshow("ORB关键点检测rich", result02);
	waitKey(0);
	destroyAllWindows();
	return 0;
}

效果：

SIFT（尺度不变特征转换，Scale-invariant feature transform）特征描述子

• 尺度空间不变性
• 像素迁移不变性
• 角度旋转不变性

SIFT特征提取步骤

• 尺度空间极值检测
• 关键点定位
• 方向指派
• 特征描述子

尺度空间极值检测

• 构建尺度空间 -- 图像金字塔 + 高斯尺度空间
• 三层空间中的极值查找

关键点定位

• 极值点定位 - 求导拟合
• 删除低对比度与低响应候选点

方向指派

• 关键点方向指派
• Scale尺度最近的图像，1.5倍大小的高斯窗口

特征描述子

• 128维向量/特征描述子
• 描述子编码方式

代码实现：

#include <opencv2/opencv.hpp>
#include <iostream>

using namespace cv;
using namespace std;

int main(int argc, char** argv) {
	Mat src = imread("D:/images/building.png");
	imshow("input", src);
	//Mat gray;
	//cvtColor(src, gray, COLOR_BGR2GRAY);

	auto sift = SIFT::create(500);
	vector<KeyPoint> kypts;
	sift->detect(src, kypts);

	Mat result01, result02, result03;
	//drawKeypoints(src, kypts, result01, Scalar::all(-1), DrawMatchesFlags::DEFAULT);
	drawKeypoints(src, kypts, result02, Scalar::all(-1), DrawMatchesFlags::DRAW_RICH_KEYPOINTS);

	std::cout << kypts.size() << std::endl;

	for (int i = 0; i < kypts.size(); i++) {
		std::cout << "pt: " << kypts[i].pt << " angle: " << kypts[i].angle << " size: " << kypts[i].size << std::endl;
	}

	Mat desc_orb;
	sift->compute(src, kypts, desc_orb);

	std::cout << desc_orb.rows << " x " << desc_orb.cols << std::endl;

	//imshow("ORB关键点检测default", result01);
	imshow("ORB关键点检测rich", result02);
	waitKey(0);
	destroyAllWindows();
	return 0;
}

效果：

5、特征匹配

• 特征匹配算法
• 特征匹配函数
• 特征匹配方法对比

特征匹配算法

• 暴力匹配，全局搜索，计算最小距离，返回相似描述子合集

  

• FLANN匹配，2009年发布的开源高维数据匹配算法库，全称Fast Library for Approximate Nearest Neighbors

• 支持KMeans、KDTree、KNN、多探针LSH等搜索与匹配算法

描述子	匹配方法
SIFT, SURF, and KAZE	L1 Norm
AKAZE, ORB, and BRISK	Hamming distance（二值编码）

// 暴力匹配
auto bfMatcher = BFMatcher::create(NORM_HAMMING, false);
std::vector<DMatch> matches;
bfMatcher->match(box_descriptors, scene_descriptors, matches);
Mat img_orb_matches;
drawMatches(box, box_kpts, box_in_scene, scene_kpts, matches, img_orb_matches);
imshow("ORB暴力匹配演示", img_orb_matches);

// FLANN匹配
auto flannMatcher = FlannBasedMatcher(new flann::LshIndexParams(6, 12, 2));
flannMatcher.match(box_descriptors, scene_descriptors, matches);
Mat img_flann_matches;
drawMatches(box, box_kpts, box_in_scene, scene_kpts, matches, img_flann_matches);
namedWindow("FLANN匹配演示", WINDOW_FREERATIO);
imshow("FLANN匹配演示", img_flann_matches);

特征匹配DMatch数据结构

DMatch数据结构：

• queryIdx
• trainIdx
• distance

distance表示距离，值越小表示匹配程度越高。

OpenCV特征匹配方法对比

代码实现：

#include <opencv2/opencv.hpp>
#include <iostream>

using namespace cv;
using namespace std;

int main(int argc, char** argv) {

	Mat book = imread("D:/images/book.jpg");
	Mat book_on_desk = imread("D:/images/book_on_desk.jpg");
	imshow("book", book);
	//Mat gray;
	//cvtColor(src, gray, COLOR_BGR2GRAY);

	vector<KeyPoint> kypts_book;
	vector<KeyPoint> kypts_book_on_desk;
	Mat desc_book, desc_book_on_desk;

	//auto orb = ORB::create(500);
	//orb->detectAndCompute(book, Mat(), kypts_book, desc_book);
	//orb->detectAndCompute(book_on_desk, Mat(), kypts_book_on_desk, desc_book_on_desk);

	auto sift = SIFT::create(500);
	sift->detectAndCompute(book, Mat(), kypts_book, desc_book);
	sift->detectAndCompute(book_on_desk, Mat(), kypts_book_on_desk, desc_book_on_desk);

	Mat result;
	vector<DMatch> matches;

	//// 暴力匹配
	//auto bf_matcher = BFMatcher::create(NORM_HAMMING, false);
	//bf_matcher->match(desc_book, desc_book_on_desk, matches);
	//drawMatches(book, kypts_book, book_on_desk, kypts_book_on_desk, matches, result);

	// FLANN匹配
	//auto flannMatcher = FlannBasedMatcher(new flann::LshIndexParams(6, 12, 2));
	auto flannMatcher = FlannBasedMatcher();
	flannMatcher.match(desc_book, desc_book_on_desk, matches);
	Mat img_flann_matches;
	drawMatches(book, kypts_book, book_on_desk, kypts_book_on_desk, matches, img_flann_matches);
	namedWindow("SIFT-FLANN匹配演示", WINDOW_FREERATIO);
	imshow("SIFT-FLANN匹配演示", img_flann_matches);

	//namedWindow("ORB暴力匹配演示", WINDOW_FREERATIO);
	//imshow("ORB暴力匹配演示", result);
	waitKey(0);
	destroyAllWindows();
	return 0;
}

效果：

1、ORB描述子匹配效果

2、SIFT描述子匹配效果

6、单应性变换/透视变换

Mat cv::findHomography(
	InputArray    srcPoints,    // 输入
    InputArray    dstPoints,    // 输出
    int    method = 0,
    double    ransacReprojThreshold = 3,
    OuputArray    mask = noArray(),
    const int    maxIters = 2000,
    const double    confidence = 0.995
)

拟合方法：

• 最小二乘法（0）
• 随机采样一致性（RANSC）
• 渐进采样一致性（RHO）

代码实现：

#include <opencv2/opencv.hpp>
#include <iostream>

using namespace std;
using namespace cv;

int main(int argc, char** argv) {

	Mat input = imread("D:/images/book_on_desk.jpg");
	Mat gray, binary;
	cvtColor(input, gray, COLOR_BGR2GRAY);
	threshold(gray, binary, 0, 255, THRESH_BINARY | THRESH_OTSU);

	vector<vector<Point>> contours;
	vector<Vec4i> hierachy;
	findContours(binary, contours, hierachy, RETR_EXTERNAL, CHAIN_APPROX_SIMPLE);
	int index = 0;
	for (size_t i = 0; i < contours.size(); i++) {
		if (contourArea(contours[i]) > contourArea(contours[index])) {
			index = i;
		}
	}

	Mat approxCurve;
	approxPolyDP(contours[index], approxCurve, contours[index].size() / 10, true);

	//imshow("approx", approxCurve);

	//std::cout << contours.size() << std::endl;
	vector<Point2f> srcPts;
	vector<Point2f> dstPts;

	for (int i = 0; i < approxCurve.rows; i++) {
		Vec2i pt = approxCurve.at<Vec2i>(i, 0);
		srcPts.push_back(Point(pt[0], pt[1]));
		circle(input, Point(pt[0], pt[1]), 12, Scalar(0, 0, 255), 2, 8, 0);
		putText(input, std::to_string(i), Point(pt[0], pt[1]), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(255, 0, 0), 1);
	}

	dstPts.push_back(Point2f(0, 0));
	dstPts.push_back(Point2f(0, 760));
	dstPts.push_back(Point2f(585, 760));
	dstPts.push_back(Point2f(585, 0));

	Mat h = findHomography(srcPts, dstPts, RANSAC);    //计算单应性矩阵
	Mat result;
	warpPerspective(input, result, h, Size(585, 760));    //对原图进行透视变换获得校正后的目标区域

	namedWindow("result", WINDOW_FREERATIO);
	imshow("result", result);

	drawContours(input, contours, index, Scalar(0, 255, 0), 2, 8);
	namedWindow("轮廓", WINDOW_FREERATIO);
	imshow("轮廓", input);
	waitKey(0);

	return 0;
}

效果：

7、基于匹配的对象检测

• 基于特征的匹配与对象检测
• ORB/AKAZE/SIFT
• 暴力/FLANN
• 透视变换
• 检测框

代码实现：

#include <opencv2/opencv.hpp>
#include <iostream>

using namespace cv;
using namespace std;

int main(int argc, char** argv) {
	// Mat image = imread("D:/images/butterfly.jpg");
	Mat book = imread("D:/images/book.jpg");
	Mat book_on_desk = imread("D:/images/book_on_desk.jpg");
	namedWindow("book", WINDOW_FREERATIO);
	imshow("book", book);
	auto orb = ORB::create(500);
	vector<KeyPoint> kypts_book;
	vector<KeyPoint> kypts_book_on_desk;
	Mat desc_book, desc_book_on_desk;
	orb->detectAndCompute(book, Mat(), kypts_book, desc_book);
	orb->detectAndCompute(book_on_desk, Mat(), kypts_book_on_desk, desc_book_on_desk);
	Mat result;
	auto bf_matcher = BFMatcher::create(NORM_HAMMING, false);
	vector<DMatch> matches;
	bf_matcher->match(desc_book, desc_book_on_desk, matches);

	float good_rate = 0.15f;
	int num_good_matches = matches.size() * good_rate;
	std::cout << num_good_matches << std::endl;
	std::sort(matches.begin(), matches.end());

	matches.erase(matches.begin() + num_good_matches, matches.end());

	drawMatches(book, kypts_book, book_on_desk, kypts_book_on_desk, matches, result);
	vector<Point2f> obj_pts;
	vector<Point2f> scene_pts;
	for (size_t t = 0; t < matches.size(); t++) {
		obj_pts.push_back(kypts_book[matches[t].queryIdx].pt);
		scene_pts.push_back(kypts_book_on_desk[matches[t].trainIdx].pt);
	}

	Mat h = findHomography(obj_pts, scene_pts, RANSAC);    // 计算单应性矩阵h
	vector<Point2f> srcPts;
	srcPts.push_back(Point2f(0, 0));
	srcPts.push_back(Point2f(book.cols, 0));
	srcPts.push_back(Point2f(book.cols, book.rows));
	srcPts.push_back(Point2f(0, book.rows));

	std::vector<Point2f> dstPts(4);
	perspectiveTransform(srcPts, dstPts, h);    // 计算转换后书的四个顶点

	for (int i = 0; i < 4; i++) {
		line(book_on_desk, dstPts[i], dstPts[(i + 1) % 4], Scalar(0, 0, 255), 2, 8, 0);
	}
	namedWindow("暴力匹配", WINDOW_FREERATIO);
	imshow("暴力匹配", result);
	namedWindow("对象检测", WINDOW_FREERATIO);
	imshow("对象检测", book_on_desk);
	//imwrite("D:/object_find.png", book_on_desk);
	waitKey(0);
	return 0;
}

效果：

8、文档对齐

• 模板表单/文档
• 特征匹配与对齐

代码实现：

#include <opencv2/opencv.hpp>
#include <iostream>

using namespace cv;
using namespace std;

int main(int argc, char** argv) {
	Mat ref_img = imread("D:/images/form.png");
	Mat img = imread("D:/images/form_in_doc.jpg");
	imshow("表单模板", ref_img);
	auto orb = ORB::create(500);
	vector<KeyPoint> kypts_ref;
	vector<KeyPoint> kypts_img;
	Mat desc_book, desc_book_on_desk;
	orb->detectAndCompute(ref_img, Mat(), kypts_ref, desc_book);
	orb->detectAndCompute(img, Mat(), kypts_img, desc_book_on_desk);
	Mat result;
	auto bf_matcher = BFMatcher::create(NORM_HAMMING, false);
	vector<DMatch> matches;
	bf_matcher->match(desc_book_on_desk, desc_book, matches);
	float good_rate = 0.15f;
	int num_good_matches = matches.size() * good_rate;
	std::cout << num_good_matches << std::endl;
	std::sort(matches.begin(), matches.end());
	matches.erase(matches.begin() + num_good_matches, matches.end());
	drawMatches(ref_img, kypts_ref, img, kypts_img, matches, result);
	imshow("匹配", result);
	imwrite("D:/images/result_doc.png", result);

	// Extract location of good matches
	std::vector<Point2f> points1, points2;
	for (size_t i = 0; i < matches.size(); i++)
	{
		points1.push_back(kypts_img[matches[i].queryIdx].pt);
		points2.push_back(kypts_ref[matches[i].trainIdx].pt);
	}
	Mat h = findHomography(points1, points2, RANSAC);    // 尽量用RANSAC，比最小二乘法效果好一些
	Mat aligned_doc;
	warpPerspective(img, aligned_doc, h, ref_img.size());    // 单应性矩阵h决定了其他无效区域不会被变换，只会变换target区域
	imwrite("D:/images/aligned_doc.png", aligned_doc);
	waitKey(0);
	destroyAllWindows();
	return 0;
}

效果：

9、图像拼接

• 特征检测与匹配
• 图像对齐与变换
• 图像边缘融合

代码实现：

#include <opencv2/opencv.hpp>
#include <iostream>
#define RATIO    0.8
using namespace std;
using namespace cv;

void linspace(Mat& image, float begin, float finish, int number, Mat &mask);
void generate_mask(Mat &img, Mat &mask);
int main(int argc, char** argv) {
	Mat left = imread("D:/images/q11.jpg");
	Mat right = imread("D:/images/q22.jpg");
	if (left.empty() || right.empty()) {
		printf("could not load image...\n");
		return -1;
	}

	// 提取特征点与描述子
	vector<KeyPoint> keypoints_right, keypoints_left;
	Mat descriptors_right, descriptors_left;
	auto detector = AKAZE::create();
	detector->detectAndCompute(left, Mat(), keypoints_left, descriptors_left);
	detector->detectAndCompute(right, Mat(), keypoints_right, descriptors_right);

	// 暴力匹配
	vector<DMatch> matches;
	auto matcher = DescriptorMatcher::create(DescriptorMatcher::BRUTEFORCE);

	// 发现匹配
	std::vector< std::vector<DMatch> > knn_matches;
	matcher->knnMatch(descriptors_left, descriptors_right, knn_matches, 2);
	const float ratio_thresh = 0.7f;
	std::vector<DMatch> good_matches;
	for (size_t i = 0; i < knn_matches.size(); i++)
	{
		if (knn_matches[i][0].distance < ratio_thresh * knn_matches[i][1].distance)
		{
			good_matches.push_back(knn_matches[i][0]);
		}
	}
	printf("total good match points : %d\n", good_matches.size());
	std::cout << std::endl;

	Mat dst;
	drawMatches(left, keypoints_left, right, keypoints_right, good_matches, dst);
	imshow("output", dst);
	imwrite("D:/images/good_matches.png", dst);

	//-- Localize the object
	std::vector<Point2f> left_pts;
	std::vector<Point2f> right_pts;
	for (size_t i = 0; i < good_matches.size(); i++)
	{
		// 收集所有好的匹配点
		left_pts.push_back(keypoints_left[good_matches[i].queryIdx].pt);
		right_pts.push_back(keypoints_right[good_matches[i].trainIdx].pt);
	}

	// 配准与对齐，对齐到第一张
	Mat H = findHomography(right_pts, left_pts, RANSAC);

	// 获取全景图大小
	int h = max(left.rows, right.rows);
	int w = left.cols + right.cols;
	Mat panorama_01 = Mat::zeros(Size(w, h), CV_8UC3);
	Rect roi;
	roi.x = 0;
	roi.y = 0;
	roi.width = left.cols;
	roi.height = left.rows;

	// 获取左侧与右侧对齐图像
	left.copyTo(panorama_01(roi));
	imwrite("D:/images/panorama_01.png", panorama_01);
	Mat panorama_02;
	warpPerspective(right, panorama_02, H, Size(w, h));
	imwrite("D:/images/panorama_02.png", panorama_02);

	// 计算融合重叠区域mask
	Mat mask = Mat::zeros(Size(w, h), CV_8UC1);
	generate_mask(panorama_02, mask);

	// 创建遮罩层并根据mask完成权重初始化
	Mat mask1 = Mat::ones(Size(w, h), CV_32FC1);
	Mat mask2 = Mat::ones(Size(w, h), CV_32FC1);

	// left mask
	linspace(mask1, 1, 0, left.cols, mask);

	// right mask
	linspace(mask2, 0, 1, left.cols, mask);

	namedWindow("mask1", WINDOW_FREERATIO);
	imshow("mask1", mask1);
	namedWindow("mask2", WINDOW_FREERATIO);
	imshow("mask2", mask2);

	// 左侧融合
	Mat m1;
	vector<Mat> mv;
	mv.push_back(mask1);
	mv.push_back(mask1);
	mv.push_back(mask1);
	merge(mv, m1);
	panorama_01.convertTo(panorama_01, CV_32F);
	multiply(panorama_01, m1, panorama_01);

	// 右侧融合
	mv.clear();
	mv.push_back(mask2);
	mv.push_back(mask2);
	mv.push_back(mask2);
	Mat m2;
	merge(mv, m2);
	panorama_02.convertTo(panorama_02, CV_32F);
	multiply(panorama_02, m2, panorama_02);

	// 合并全景图
	Mat panorama;
	add(panorama_01, panorama_02, panorama);
	panorama.convertTo(panorama, CV_8U);
	imwrite("D:/images/panorama.png", panorama);
	waitKey(0);
	return 0;
}

void generate_mask(Mat &img, Mat &mask) {
	int w = img.cols;
	int h = img.rows;
	for (int row = 0; row < h; row++) {
		for (int col = 0; col < w; col++) {
			Vec3b p = img.at<Vec3b>(row, col);
			int b = p[0];
			int g = p[1];
			int r = p[2];
			if (b == g && g == r && r == 0) {
				mask.at<uchar>(row, col) = 255;
			}
		}
	}
	imwrite("D:/images/mask.png", mask);
}

// 对mask中的0区域，进行逐行计算每个像素的权重值
void linspace(Mat& image, float begin, float finish, int w1, Mat &mask) {
	int offsetx = 0;
	float interval = 0;
	float delta = 0;
	for (int i = 0; i < image.rows; i++) {
		offsetx = 0;
		interval = 0;
		delta = 0;
		for (int j = 0; j < image.cols; j++) {
			int pv = mask.at<uchar>(i, j);
			if (pv == 0 && offsetx == 0) {
				offsetx = j;
				delta = w1 - offsetx;
				interval = (finish - begin) / (delta - 1);    // 计算每个像素变化的大小
				image.at<float>(i, j) = begin + (j - offsetx)*interval;
			}
			else if (pv == 0 && offsetx > 0 && (j - offsetx) < delta) {
				image.at<float>(i, j) = begin + (j - offsetx)*interval;
			}
		}
	}
}

效果：

1、图像拼接重合区域mask

2、拼接前图像

3、图像拼接效果图

10、条码标签定位与有无判定

代码实现：

ORBDetector.h

#pragma once
#include <opencv2/opencv.hpp>

class ORBDetector {
public:
	ORBDetector(void);
	~ORBDetector(void);
	void initORB(cv::Mat &refImg);
	bool detect_and_analysis(cv::Mat &image, cv::Mat &aligned);
private:
	cv::Ptr<cv::ORB> orb = cv::ORB::create(500);
	std::vector<cv::KeyPoint> tpl_kps;
	cv::Mat tpl_descriptors;
	cv::Mat tpl;
};

ORBDetector.cpp

#include "ORBDetector.h"

ORBDetector::ORBDetector() {
	std::cout << "create orb detector..." << std::endl;
}

ORBDetector::~ORBDetector() {
	this->tpl_descriptors.release();
	this->tpl_kps.clear();
	this->orb.release();
	this->tpl.release();
	std::cout << "destory instance..." << std::endl;
}

void ORBDetector::initORB(cv::Mat &refImg) {
	if (!refImg.empty()) {
		cv::Mat tplGray;
		cv::cvtColor(refImg, tplGray, cv::COLOR_BGR2GRAY);
		orb->detectAndCompute(tplGray, cv::Mat(), this->tpl_kps, this->tpl_descriptors);
		tplGray.copyTo(this->tpl);
	}
}

bool ORBDetector::detect_and_analysis(cv::Mat &image, cv::Mat &aligned) {
	// keypoints and match threshold
	float GOOD_MATCH_PERCENT = 0.15f;
	bool found = true;
	// 处理数据集中每一张数据
	cv::Mat img2Gray;
	cv::cvtColor(image, img2Gray, cv::COLOR_BGR2GRAY);
	std::vector<cv::KeyPoint> img_kps;
	cv::Mat img_descriptors;
	orb->detectAndCompute(img2Gray, cv::Mat(), img_kps, img_descriptors);

	std::vector<cv::DMatch> matches;
	cv::Ptr<cv::DescriptorMatcher> matcher = cv::DescriptorMatcher::create("BruteForce-Hamming");
	// auto flann_matcher = cv::FlannBasedMatcher(new cv::flann::LshIndexParams(6, 12, 2));
	matcher->match(img_descriptors, this->tpl_descriptors, matches, cv::Mat());

	// Sort matches by score
	std::sort(matches.begin(), matches.end());

	// Remove not so good matches
	const int numGoodMatches = matches.size() * GOOD_MATCH_PERCENT;
	matches.erase(matches.begin() + numGoodMatches, matches.end());
	// std::cout << numGoodMatches <<"distance:"<<matches [0].distance<< std::endl;
	if (matches[0].distance > 30) {
		found = false;
	}
	// Extract location of good matches
	std::vector<cv::Point2f> points1, points2;
	for (size_t i = 0; i < matches.size(); i++)
	{
		points1.push_back(img_kps[matches[i].queryIdx].pt);
		points2.push_back(tpl_kps[matches[i].trainIdx].pt);
	}

	cv::Mat H = findHomography(points1, points2, cv::RANSAC);
	cv::Mat im2Reg;
	warpPerspective(image, im2Reg, H, tpl.size());

	// 逆时针旋转90度
	cv::Mat result;
	cv::rotate(im2Reg, result, cv::ROTATE_90_COUNTERCLOCKWISE);
	result.copyTo(aligned);
	return found;
}

object_analysis.cpp

#include "ORBDetector.h"
#include <iostream>

using namespace cv;
using namespace std;

int main(int argc, char** argv) {
	Mat refImg = imread("D:/facedb/tiaoma/tpl2.png");
	ORBDetector orb_detector;
	orb_detector.initORB(refImg);
	vector<std::string> files;
	glob("D:/facedb/orb_barcode", files);
	cv::Mat temp;
	for (auto file : files) {
		std::cout << file << std::endl;
		cv::Mat image = imread(file);
		int64 start = getTickCount();
		bool OK = orb_detector.detect_and_analysis(image, temp);
		double ct = (getTickCount() - start) / getTickFrequency();
		printf("decode time: %.5f ms\n", ct * 1000);
		std::cout << "标签: " << (OK == true) << std::endl;
		imshow("temp", temp);
		waitKey(0);
	}
}

效果：

1、检测图片

2、模板图片

11、DNN概述

DNN模块介绍：

• DNN - Deep Neutal Network
• OpenCV3.3 开始发布
• 支持VOC与COCO数据集的对象检测模型
• 包括SSD/Faster-RCNN/YOLOv4等
• 支持自定义对象检测
• 支持人脸检测

函数知识：

• 读取模型
• 转换数据与设置
• 推理输出

Net net = readNetFromTensorflow(model, config);    // 支持tensorflow
Net net = readNetFromCaffe(config, model);    // 支持caffe
Net net = readNetFromONNX(onnxfile);

// 读取数据
Mat image = imread("D:/images/example.png");
Mat blob_img = blobFromImage(image, scalefactor, size, mean, swapRB);
net.setInput(blob_img);

// 推理输出
Mat result = net.forward();

后处理/输出解析：

• 不同网络的输出不同
• 如何解析要根据模型输出
• 对象检测网络SSD/Faster-RCNN解析

SSD的输出解析：

• http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2017_11_17.tar.gz
• 1 x 1 x N x 7 - DetectOutput
• [image_id, label, conf, x_min, y_min, x_max, y_max]

Faster-RCNN输出解析：

• http://download.tensorflow.org/models/object_detection/faster_rcnn_inception_v2_coco_2018_01_28.tar.gz
• 1 x 1 x N x 7 - DetectOutput
• [image_id, label, conf, x_min, y_min, x_max, y_max]

YOLOv4输出解析：

• 解析多个输出层，80个类别 - N x W x H x D
• 4 + 80 预测，三个输出层
• centerx，centery，width，height
• 13 x 13 x 255 = 13 x 13 x 3 x 85
• NMS
• https://github.com/AlexeyAB/darknet/wiki/YOLOv4-model-zoo