节点文献
基于特征的非刚性图像配准算法研究
Research on Non-rigid Image Registration Algorithms Based on Features
【作者】 王阳；
【作者基本信息】 天津大学 ， 电子与通信工程， 2019， 硕士
【摘要】 图像配准的主要任务是寻找两幅或多幅图像之间的空间变换关系,以实现图像在空间位置上的对齐。其中,非刚性图像配准是计算机视觉领域的研究热点,在医学图像处理、遥感图像分析、图像拼接与融合等领域都有着重要的理论意义和研究价值,代表了未来的研究方向。本文首先分析了非刚性图像配准的研究背景以及国内外研究现状,着重介绍了基于特征和光流估计的非刚性图像配准方法,并指出了这两类方法的优势与不足,同时给出了评价配准效果的主观和客观评价标准。基于特征的非刚性图像配准容易产生误匹配。为解决这一问题,本文提出一种基于超像素运动统计的误匹配去除方法。首先,将超像素划分的思想引入到特征匹配当中,采用改进的超像素分割算法对待配准图像进行分块;其次,提出了超像素运动的一致性约束,假设同一超像素区域内部的特征点具有相同或一致的运动趋势;最后,建立了超像素网格统计模型,将局部运动的平滑约束转化为匹配数量的统计似然函数,有效地实现了误匹配的自动检测。实验结果表明,与现有方法相比,本文方法对处理存在非刚性以及大位移形变的图像具有较高的鲁棒性。此外,论文提出一种基于矢量场插值模型的非刚性图像配准算法。首先,在特征矢量场的基础上添加了运动边界约束,获得运动边界处的特征矢量,去除误匹配后得到精确的稀疏矢量场;然后,提出了一种矢量场插值建模的方法,能够在稀疏矢量场的基础上进行矢量插值,得到稠密的位移矢量场;最后,根据稠密位移矢量场对浮动图像进行矫正,获得配准后的图像。该方法能基于少量特征矢量估计得到非刚性图像的全像素运动矢量场,可有效弥补光流估计算法的大位移信息缺失问题。多组实验结果证明了本文算法的可行性和鲁棒性。最后,对论文工作进行了总结,对未来的研究方向作了进一步规划和展望。
【Abstract】 Image registration aims to search the spatial transformation relationship between two or more images,and then align these images in spatial position.As a research hotspot in the field of computer vision,non-rigid image registration has important theoretical significance and research value in the fields of medical image processing,remote sensing image analysis,image mosaic and fusion,and so on.It represents the study direction in the future.Firstly,this thesis analyzes the research background of non-rigid image registration and the research status at home and abroad.Then the non-rigid image registration methods based on features and optical flow estimation are emphasized,and their advantages and disadvantages are also pointed out.In addition,the subjective and objective evaluation criteria of image registration are given.Feature-based non-rigid image registration is prone to mismatching.To solve this problem,this thesis proposes a mismatch removal method based on super-pixel motion statistics.Firstly,the idea of super-pixel segmentation is introduced into feature matching,and an improved super-pixel segmentation algorithm is used to segment the images for registration.Secondly,this thesis proposes a consistency constraint of super-pixel motion by assuming that the feature points in the same super-pixel region are of the same or consistent motion trend.Finally,the super-pixel grids statistical model is established,and the local motion smoothing constraint is transformed into the statistical likelihood function of the matching number,thus the automatic identification and mismatch removal can be realized.The simulation results compared to the state-of-the-art algorithms shown that the proposed method is robust to the non-rigid images with large displacement deformation.In addition,this thesis proposes a non-rigid image registration algorithm based on vector field interpolation model.Firstly,motion boundary constraints are added to the feature vector field,and obtain the edge vector field on the moving boundary.After removing the mismatches,the precise sparse vector field is obtained.Then,a vector field interpolation modeling method is proposed,which can realize the interpolation of sparse vector field.Finally,the floating image is aligned according to the obtained dense displacement vector field,and then the final registration can be realized.This method can estimate the full-pixel motion vector field between non-rigid images using a small number of feature vectors,and then solve the problem of large displacement information missing in optical flow estimation algorithms.Multiple experimental results demonstrate the effectiveness and robustness of the proposed algorithm.Finally,the work in the thesis is summarized,and the study direction in the future is also expected.