OpenCV 基于距离变换的高精度轮廓匹配

轮廓匹配在定位测量应用中对其匹配的精度有更高的要求,通常的像素级的匹配结果难以满足其要求。本文给出了一种具有亚像素精度的快速轮廓匹配定位方法,其进行数学计算的基础为二值图像的距离变换。

二值图像距离变换的概念由Rosenfeld和Pfaltz于1966年其论文中提出,目前广泛应用于计算机图形学,计算机视觉及GIS空间分析等领域,其基本含义是计算一个图像中非零像素点到最近的零像素点的距离,也就是对每一各非零像素点计算其到零像素点的最短距离,并将该距离值赋值给该非零像素位置,从而将一幅二值图像变换为一幅距离图像。OpenCV通过cv::distanceTransform()函数,给出了该功能的快速实现方法。并通过参数的方式给出了距离的几种定义方式,如欧式距离、马氏距离、倒角距离等。

高精度轮廓匹配的算法实现步骤:

假设待匹配的两个轮廓分别为A和B

(1)       将轮廓A栅格化并生成为一幅二值图像Ia,其中轮廓点为255,背景点为0;

(2)       将二值图Ia进行距离变换得到距离图Id;

(3)       通过一个变换描述轮廓B和轮廓A之间的关系,如平移变换、刚体变换、仿射变换等;

(4)       通过一定的搜索策略或者最优化方法搜索轮廓B在距离图Id中的最佳变换参数,并输出。

在上述过程中(4)是最关键的一步,具体的搜索策略可考虑使用Powell搜索、分支定界搜索或者Gauss-Newton方法。下面给出了使用Gauss-Newton方法实现求解平移变换的代码,仅供参考。其它复杂的变换可参考此过程完成。

1 数据栅格化,并距离变换

int GetRasterDistanceImage(cv::Point2f* refPoints, int pntsNum, double resolution, 
						   int& offX, int& offY, cv::Mat& rasterDistImg)
{
	double maxX, minX, maxY, minY;

	int i;
	for (i = 0; i < pntsNum; ++ i)
	{
		maxX = refPoints[i].x;
		minX = refPoints[i].x;
		maxY = refPoints[i].y;
		minY = refPoints[i].y;

		break;
	}

	for ( ; i < pntsNum; ++ i)
	{
		if (refPoints[i].x > maxX)
			maxX = refPoints[i].x;
		if (refPoints[i].x < minX)
			minX = refPoints[i].x;

		if (refPoints[i].y > maxY)
			maxY = refPoints[i].y;
		if (refPoints[i].y < minY)
			minY = refPoints[i].y;
	}

	int nX = int((maxX - minX) / resolution) + 200;
	int nY = int((maxY - minY) / resolution) + 200;

	cv::Mat rasterImg;
	rasterImg.create(nY, nX, CV_8UC1);
	memset(rasterImg.data, 0xFF, nY * nX);

	double realCenterX = (maxX + minX) / 2.0f;
	double realCenterY = (maxY + minY) / 2.0f;

	int offsetX = nX / 2 - int(realCenterX / resolution + 0.5f);
	int offsetY = nY / 2 - int(realCenterY / resolution + 0.5f);

	uchar* dat = (uchar*)(rasterImg.data);
	for (i = 0; i < pntsNum; ++ i)
	{
		if (validateMask[i] == 0)
			continue;

		int x = (int)floor(refPoints[i].x / resolution + 0.5f) + offsetX;
		int y = nY - ((int)floor(refPoints[i].y / resolution + 0.5f) + offsetY);

		*(dat + nX * y + x) = 0;
	}

	cv::distanceTransform(rasterImg, rasterDistImg, CV_DIST_C, 3);

	offX = offsetX;
	offY = offsetY;

	return 1;
}

2  高精度轮廓匹配,并输出平移量 offX,offY
void MatchReferencePoints(cv::Mat& rasterDistImg, cv::Point2f* rasterPoints, int pointsNum , double& offX, double& offY)
{
	offX = 0.0f;
	offY = 0.0f;

	cv::Mat imgB = rasterDistImg;

	double* pL = new double[pointsNum];
	double* pA = new double[pointsNum * 2];

	double a_b[2] = {0};
	double deltaX[2];
	double AL[2];
	double MM[4];

	double va0, va1, va2, va3, va4;
	double sx, sy;
	double pError = 0.0f;

	for (int i = 0; i < 100; ++ i)
	{
		double error = 0.0f;
		for (int j = 0; j < pointsNum; ++ j)
		{
			sx = myPoints[j].x + a_b[0];
			sy = myPoints[j].y + a_b[1];

			va0 = GetDataValue(imgB, sx, sy);

			va1 = GetDataValue(imgB, sx + 0.5, sy);
			va2 = GetDataValue(imgB, sx - 0.5, sy);
			va3 = GetDataValue(imgB, sx, sy + 0.5);
			va4 = GetDataValue(imgB, sx, sy - 0.5);

			pA[j] = va1 - va2;
			pA[j + pointsNum] = va3 - va4;

			pL[j] = 0.0f - va0;

			error += pL[j] * pL[j];
		}

		if (i == 0)
			pError = error;
		else if (error > pError)
			break;

		pError = error;

		int nOff_M = 0;
		for(int j = 0; j < 2; ++ j){
			for(int jj = 0; jj < 2; ++ jj){
				MM[nOff_M + jj] = 0.0;
				for(int kk = 0; kk < pointsNum; ++ kk)
					MM[nOff_M + jj] += pA[j * pointsNum + kk] * pA[jj * pointsNum + kk]; 
			}							// M = A * A(T)
			nOff_M += 2;
		}

		cv::Mat cvMatM(2, 2, CV_64FC1, MM);
		cv::invert(cvMatM, cvMatM);

		for (int j = 0; j < 2; ++ j)
		{
			AL[j] = 0.0f;
			for (int kk = 0; kk < pointsNum; ++ kk)
				AL[j] += pA[j * pointsNum + kk] * pL[kk];
		}

		for (int j = 0; j < 2; ++ j)
		{
			deltaX[j] = 0.0f;
			for (int jj = 0; jj < 2; ++ jj)
				deltaX[j] += MM[j * 2+ jj] * AL[jj];
		}

		int s = 0;
		for (; s < 2; ++ s)
		{
			if (fabs(deltaX[s]) > 0.0001)
				break;
		}

		if (s == 2)
			break;

		a_b[0] += deltaX[0];
		a_b[1] += deltaX[1];
	}

	offX = a_b[0];
	offY = a_b[1];

	delete[] pL;
	delete[] pA;

	delete[] myPoints;

	double temp = sqrt(pError / pointsNum);
}



你可能感兴趣的:(OpenCV,拾遗)