结合多尺度特征与局部采样描述的多模态图像配准方法
2025-08-03贾志有王国刚
Research on multimodal image registration method combining multi-scale features and local sampling description
Jia Zhiyou,Wang Guogang (School of Information Enginering,Shenyang Chemical University,Shenyang11O142,China)
Abstract:Aimingatthematching dificultiescausedbytheexistenceofserious geometricdiferencesandnonlinearntensity diference(NID)indiferentmodal images,this paperproposedamultimodal imagealignmentmethodcombining multi-scale featureswithalocalsamplingdescription.Firstly,themethodintroducedanonlineardifusionequationtoconstructanonlinearscalespace,andthen itcombineda phaseconsistencyandorientedFASTandrotatedBRIEF(ORB)algorithm to obtain multi-scalestable feature points.Then,the method proposedarotation-invariant doubleGaussiansamplingdescriptor,which could robustly span the rotation difference of [0∘,360∘) in the presence of NID. Finally,the method introduced an image recoverystrategy.The methodobtainedtheoptimal geometrictransformationmodel through primarymatching,corected the geometricdiferences existingbetween images,andthenperformed secondarymatching toimprovethematchingaccuracy. Experiments on multimodal data sets inremote sensing,medicine,andcomputer visionshowthattheroot-mean-squareerorof the proposed method can reach within 1.5 pixels and the correct matching rate can exceed 98% when there are geometric diferencessuchasscaleandrotation.Theresultsdemonstrate that this methodcanovercometheinfluenceof nonlinearradiation difference between images and achieve high precision registration.
Keywords:multimodalimages registration;nonlinear scale space;phase coherence;double Gausian sampling descriptor; nonlinear radial disparity
0 引言
图像匹配是计算机视觉领域的一个基础和关键问题,其目的是在两幅或多幅图像中提取可靠的特征对应,使之成为图像融合、图像检测、目标跟踪等多个领域的先决条件。而不同传感器类型、不同成像时间或视点的多模态图像存在尺度、旋转、仿射和非线性辐射(nonlinearintensitydifference,NID)等差异,从而导致匹配性能大幅下降,难以满足不断变化的实际应用要求[1]。针对这一难题,专家学者们提出了大量的方法。这些方法大致可划分为基于模板的匹配方法、基于深度学习的匹配方法和基于特征的匹配方法三类。
基于模板的匹配方法的关键是选择合适的相似性度量,之后利用相似性度量和优化方法来准确估计几何变换参数,从而驱动配准过程的优化。频域中的相位相关(PC)2是最广泛使用的相似性度量。Ye等人[34]提出了定向相位一致性直方图(HOPC)算法和基于定向梯度通道特征(CFOG)的算法。Fan等人[5提出了基于角度加权定向梯度的像素级特征。此类算法高度依赖先验信息来粗略地消除几何扭曲,包括尺度变化和图像旋转。尽管这些方法对NID表现出良好的鲁棒性,但它们很容易受到几何扭曲的影响。因此,在图像存在大幅度旋转以及尺度等几何差异时,往往会配准失败。
深度学习的快速发展为解决多模态图像问题提供了一种新的方法。为了解决多模态图像之间显著的非线性辐射差异,Baruch等人[提出了一种新的卷积神经网络(CNN)架构。该架构利用连体CNN和双非权重共享CNN,将生成的特征检测器与特征描述符紧密耦合,并在VIS-NIR跨模态场景中获得了良好的结果。……
