QT + opencv 实现形状(轮廓)模板匹配

opencv__0">QT + opencv 实现形状(轮廓)模板匹配

实现思路

1.创建模板数据：主要是提取模板的轮廓信息，这一步通常通过边缘检测实现。将模板的轮廓信息存储起来。

代码：

//创建形状模板
bool cvLearnShapeMatchPattern_(Mat matDst, MyShapeUiParam param, MyShapeTemplData* pTemplData)
{//图像不存在if(matDst.empty())return false;//清除模板pTemplData->clear();//边缘检测Mat canny, gray;vector<Point> contours;cv::Canny(matDst, canny, pTemplData->minThresh, pTemplData->maxThresh);canny.copyTo(gray);/*if(!(pTemplData->sample==0)){Mat element=getStructuringElement(MORPH_ELLIPSE, Size(pTemplData->dilate, pTemplData->dilate));cv::dilate(canny, gray, element);}*/cvFindContour_(gray, contours, pTemplData->minLen, 10e9, pTemplData->sample);if((int)contours.size()==0)return false;//压缩金字塔int iTopLayer=param.m_iPyramid;buildPyramid(matDst, pTemplData->vecPyramid, iTopLayer);for(size_t i=0;i<pTemplData->vecPyramid.size();i++){//轮廓信息Mat src=pTemplData->vecPyramid[i];vector<Point> pyrContour=cvPyrContour_(contours, matDst.size(), src.size());pTemplData->vecPoints.push_back(pyrContour);//显示Mat dst=Mat::zeros(src.size(), CV_8UC1);for(size_t j=0;j<pyrContour.size();j++)dst.at<uchar>(pyrContour[j].y, pyrContour[j].x)=255;imshow(QString("%1").arg(i).toStdString(), dst);cv::waitKey(5);//梯度信息Mat gx, gy, mag, dir;Sobel(src, gx, CV_32F, 1, 0);Sobel(src, gy, CV_32F, 0, 1);cartToPolar(gx, gy, mag, dir);//提取梯度vector<SPtin> vecSPtin;for(size_t j=0;j<pyrContour.size();j++){SPtin info;Point p=pyrContour[j];info.derivativeX=gx.at<float>(p.y, p.x);info.derivativeY=gy.at<float>(p.y, p.x);info.magnitude=mag.at<float>(p.y, p.x);info.magnitudeN = info.magnitude==0 ? 0 : (1.0f / info.magnitude);vecSPtin.push_back(info);}pTemplData->vecSPtins.push_back(vecSPtin);}//训练标记pTemplData->bIsPatternLearned=true;return true;
}

2.模板匹配：同样需要利用压缩金字塔的方法来提升搜索速度，减少参数量。然后取顶层金字塔进行旋转，和模板轮廓进行匹配，找到大致角度后再进行细致的搜索。

部分代码

//模板匹配
bool cvShapeMatch_(Mat matSrc, vector<MyShapeTemplData> templDatas, MyShapeUiParam param, vector<MySingleTargetMatch>& vecSingleTargetData)
{//图像if(matSrc.empty())return false;//模板if((int)templDatas.size()==0)return false;//決定金字塔層數 總共為1 + iLayer層int iTopLayer=param.m_iPyramid;vector<Mat> vecMatSrcPyr;buildPyramid(matSrc, OutputArrayOfArrays(vecMatSrcPyr), iTopLayer);//取顶层图片的中心点int iTopSrcW = vecMatSrcPyr[iTopLayer].cols, iTopSrcH = vecMatSrcPyr[iTopLayer].rows;Point2f ptCenter((iTopSrcW - 1) / 2.0f, (iTopSrcH - 1) / 2.0f);//Caculate lowest score at every layervector<double> vecLayerScore(iTopLayer + 1, param.m_dScore);for (int iLayer = 1; iLayer <= iTopLayer; iLayer++)vecLayerScore[iLayer] = vecLayerScore[iLayer - 1] * param.m_dGreed;//clearvecSingleTargetData.clear();//并行加速omp_set_num_threads(4);#pragma omp parallel forfor(int k=0;k<(int)templDatas.size();k++){MyShapeTemplData* pTemplData = &templDatas[k];if((int)pTemplData->vecPyramid.size()==0 || !pTemplData->bIsPatternLearned)continue;Mat matDst=pTemplData->vecPyramid[0];if (matDst.empty())continue;if ((matDst.cols<matSrc.cols && matDst.rows>matSrc.rows) || (matDst.cols>matSrc.cols && matDst.rows<matSrc.rows))continue;if (matDst.size().area()>=matSrc.size ().area())continue;//匹配轮廓vector<Point> Contour=pTemplData->vecPoints[0];cvAffineTrans_(Contour, Point3f((float)(matDst.cols - 1)/2.0f, (float)(matDst.rows - 1)/2.0f, 0), Point3f(0, 0, 0));//第一階段以最頂層找出大致角度與ROIdouble dAngleStep=atan(2.0 / max(pTemplData->vecPyramid[iTopLayer].cols, pTemplData->vecPyramid[iTopLayer].rows)) * R2D;