【作者】 张乐；

【导师】 文颖；

【作者基本信息】 华东师范大学 ， 计算机应用技术， 2017， 硕士

【摘要】 大形变图像配准在计算机图像处理尤其医学图像处理中有重要的研究价值和应用意义。由于在配准过程中形变量较大,传统的Demons算法仅仅利用图像的梯度信息驱使像素朝梯度下降的方向扩散,使图像发生形变,导致在图像平坦或者边缘处梯度信息缺失的情况下很难完成配准。因此,本文提出了基于驱动力的Log-Demons算法,首先通过提取图像的结构张量作为约束,使配准过程保持更多的结构信息,其次利用基于描述子匹配获得的驱动力提高图像在大形变情况下的配准精度,避免算法配准陷入局部最优,最终构建了基于Log-Demons算法的大形变图像高精度配准模型。本文的主要工作包括:1)提出基于结构张量的Log-Demons算法:图像的结构张量能提取更多的局部结构信息,并且对外部光照变化不敏感。通过张量守恒准则将其融合进Log-Demons算法中来约束像素点的扩散,使其能保持良好的局部结构并获得更精确的形变场。2)提出基于驱动力的Log-Demons算法:为了能实现大形变图像高精度配准,本文提出将边缘点匹配算法获得的矢量位移作为驱动力,拉动图像边缘周围像素点的运动,使其朝正确的方向扩散,解决了大形变情况下图像边缘梯度相同导致像素点随机扩散的问题,使其能够适应大形变图像配准。3)提出基于MROGH描述子匹配获得驱动力的算法:为了获得更准确的驱动力,并适应大形变带来的大角度变化,本文提出采用MROGH构建描述子并对其进行一对一的精确匹配以获得匹配点的位移向量作为驱动力。4)提出基于特征驱动的Log-Demons融合算法:将驱动力作为常量单独计算,在迭代配准的过程中以指数减少的方式与Log-Demons的更新形变场相融合,使其在形变开始时给予较大的影响力,形变减小时由Log-Demons自身的驱动力完成配准,既加快配准过程又减少了特征点误匹配所带来的影响。在更新形变场计算过程中,本文提出在李群中利用群性质融合驱动力和Log-Demons本身的更新形变场,使融合后的形变场也能保持微分同胚。通过算法(1)-(4),最终构建了基于驱动力的Log-Demons算法模型,实现了大形变图像配准。本文方法在模拟形变图像、真实场景大位移图像以及脑图像上进行了测试,并与其它方法(如Log-Demons、SpectralLog-Demons、LDDMM)进行了比较,实验表明本文方法都取得了较好的效果。本文方法不仅能较好地完成大形变图像配准,而且计算得到的形变场具有更好的局部结构和精度。更多还原

【Abstract】 Large deformation image registration has important research value and application significance in computer vision especially medical image processing.Due to large amount of deformation in the process of registration,the traditional Demons algorithm only uses gradient information of the image to drive the pixel to diffuse,which is difficult to complete the registration if the image is flat or the gradient information is missing at the edge.Therefore,this paper proposes a Log-Demons algorithm based on driving force,which can keep more structural information in the registration process by using the structural tensor of the image as the constraint.This paper uses the driving force based on the descriptor matching to avoid the algorithm to fall into the local optimal and improve the precision in large deformation registration.Finally,this paper builds a high precision registration model based on Log-Demons algorithm for large deformation image registration.The main work of this paper includes:1)Propose Log-Demons algorithm based on structural tensor:the structural tensor of the image can extract more local structure information and is insensitive to external illumination changes.It is incorporated into the Log-Demons algorithm by the tensor conservation criterion to constrain the diffusion of the pixels so that it can maintain a good local structure and obtain a more accurate deformation field.2)Propose a driving-based Log-Demons algorithm:in order to obtain high-precision registration of large deformation images,this paper proposes to use the vector displacement obtained by edge points matching algorithm as a driving force,to pull the motion of the points around the edge of the image,which solves the problem of random diffusion in the case of large deformation,so that it can adapt to large deformation registration.3)Propose the algorithm of obtaining the driving force based on MROGH descriptor matching:in order to obtain more accurate driving force and adapt to the large angle change caused by large deformation,this paper proposes to use MROGH to construct descriptor and perform one-to-one point matching to obtain the displacement vector as the driving force.4)Propose a feature-driven Log-Demons fusion algorithm:take the driving force as a constant and combine the update of Log-Demons with driving force in the way of exponential reduction,so that it can give a great influence at the beginning.When deformation is reduced,it can complete the registration by the Log-Demons own driving force,which can not only speed up the registration process but also reduce the impact of wrong points matching.In the process of calculating update field,this paper proposes to use the property of fusion of the Lie Group to combine the driving force with the update field calculated by Log-Demons,so that the fusion field can also maintain the diffeomorphism.Through the algorithm(1)-(4),the Log-Demons algorithm model based on driving force is constructed,and the large deformation image registration is realized.The method is compared with other methods,and the experimental results show that the method has achieved good results.The results show that the proposed method has a good effect on the simulated image,the real scene large displacement image and the brain image.The method can not only complete the large deformation image registration,but also has better local structure and precision.更多还原