欢迎来到天天文库
浏览记录
ID:36263736
大小:5.43 MB
页数:52页
时间:2019-05-07
《基于特征匹配的算法》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、基于特征的匹配一.特征匹配过程1特征提取2特征描述3特征匹配二.SIFT算法LocalfeaturesDetection:2x2matrixofimagederivatives(averagedinneighborhoodofapoint).(1)平移Translation(2)欧几里德几何(平移+旋转)(3)相似性变换(平移+旋转+尺度)(4)仿射变换(5)投影变换Theneedforinvariance1.几何变换2.光照变化一.特征匹配过程1特征提取1.1HarrisandHessianDetector1.2尺度不变特征检测1.3仿射不变特征
2、检测1.4特征提取总结2特征描述3特征匹配二.SIFT算法(1)Harrisdetector(Harris,1988)Secondmomentmatrix/autocorrelationmatrix公式由来说明影像信号的局部自相关函数给定点(x,y)及位移(△x,△y),窗口为W,用差平方和(SSD)近似自相关函数,计算窗口W和位移窗口内灰度的差别。位移后影像函数通过一阶泰勒展开式近似重新计算f(x,y):“secondmomentmatrixM”Autocorrelation(secondmoment)matrix—Mcanbeusedtoder
3、iveameasureof“cornerness”—Independentofvariousdisplacements(Δx,Δy)—Corner:significantgradientsin>1directionsrankM=2—Edge:significantgradientin1directionrankM=1—HomogeneousregionrankM=0Harrisdetector流程1.Imagederivatives2.Squareofderivatives3.Gaussianfilterg()4.Cornernessfunc
4、tion5.Non-maximasuppressioncHarris>tHarris(2)Hessiandetector(Beaudet,1978)Taylor二阶展开式得到Hessian矩阵IIxxIxyIyy小总结Harrisdetector◇Rotationinvariant?YesTheeigenvaluesofMrevealtheamountofintensitychangeinthetwoprincipalorthogonalgradientdirectionsinthewindow.◇Scaleinvariant?NoHessiand
5、etector◇Rotationinvariant?Yes◇Scaleinvariant?No一.特征匹配过程1特征提取1.1HarrisandHessianDetector1.2尺度不变特征检测1.3仿射不变特征检测1.4特征提取总结2特征描述3特征匹配二.SIFT算法1.2尺度不变特征检测(1)尺度选择(2)TheLaplacian-of-Gaussian(LoG)Detector(3)TheDifference-of-Gaussian(DoG)Detector(4)TheHarris-LaplacianDetector(5)TheHessia
6、n-LaplaceDetector1.2尺度不变特征检测(1)尺度选择1.2尺度不变特征检测(1)尺度选择1.2尺度不变特征检测(2)TheLaplacian-of-Gaussian(LoG)Detector◇1LaplacianfilterLaplacian算子具有旋转不变性,但对噪声很敏感,因此常需进行平滑操作◇2LoGfilter高斯滤波平滑,然后拉普拉斯滤波。SmoothLaplacianI(x,y)O(x,y)Laplacian-of-Gaussian(LoG)尺度空间的局部极大值点1.2尺度不变特征检测(3)TheDifference-
7、of-Gaussian(DoG)Detector可用高斯差分函数(DoG)近似LoGsOriginalimageSamplingwithsteps4=2sssComputationinGaussianscalepyramidLoGandDoGZerocrossings“Mexicanhat”,“Sombrero”Edgedetector!Lowe’sDoGkeypoints[Lowe]Edgezero-crossingBlobatcorrespondingscale:localextremum!Lowcontrastcornersuppressi
8、on:thresholdAssesscurvaturedistinguishcornersfromedgesKeypoi
此文档下载收益归作者所有