基于图像特征和光流场的非刚性图像配准

Non-rigid registrations based on image characteristics and optical flows

考虑传统非刚性图像配准方法无法同时满足配准精度和配准时间要求,综合图像的特征和灰度信息,提出了几种改进的非刚性图像配准方法:基于圆形描述子特征的非刚性配准方法(Circle Descriptor Feature,CDF),基于动态驱动力Demons的非刚性配准方法(Dynamic Driving Force Demons,DDFD),和基于图像特征和光流场的非刚性配准方法。CDF方法通过提取图像的特征点,采用圆形描述子代替传统方法的正方形描述子来保证图像的旋转不变性,提高配准速度;DDFD方法通过引入驱动力系数动态改变驱动力,有效地解决了传统方法配准时间和配准精度低的问题;基于图像特征和光流场的非刚性配准方法则首先提取浮动图像和参考图像的特征点,然后利用提取的特征点进行粗配准(特征级配准),再采用基于光流场的方法进行精细配准(像素级配准),最终实现配准精度和配准时间的兼顾。对checkboard测试图像、自然图像、脑部MR图像、肝部CT图像进行了实验测试,结果表明,本文方法在配准时间、配准精度及对大形变图像的适应性方面均优于传统尺度不变特征转换(SIFT)、加速鲁棒特征(SURF)、Demons、Active Demons和全变差正则项-L^1范数项(TV-L^1)等方法。

As the non-rigid image registration methods can not meet the requirements of registration accuracy and registration time simultaneously,three kinds of improved non-rigid registration methods are proposed based on image characteristics and image gray.These non-rigid registration methods were based on the Circle Descripto increases Feature（CDF）,Dynamic Driving Force Demons（DDFD）and image characteristics and optical flow,respectively.In CDF method,feature points were extracted from the images,and the circle descriptor is used in the method instead of square descriptor in classical methods,by which the rotation invariance was maintained and the speed of the registration was increased.In DDFD method,the driving force was changed by introducing the driving force coefficient,so that the registration time and registration accuracy were improved effectively.In registration methods based on image characteristics and optical flow,the feature points were extractedfrom a float image and a reference image by using registration method based on image characteristics,and these extracted feature points were used to get a coarse registered image（feature level registration）;then the optical-flow method was used to register accurately（pixel level registration）for the coarse registered image and to achieves the purpose of taking account of the registration accuracy and registration time.The experiments on checkboard images,natural images,brain MR images and liver CT images were performed and the results show that the proposed methods are better than the classical methods such as Scale-invariant Feature Transform（SIFT）,Speeded-Up Robust Features（SURF）,Demons,Active Demons and Total Variation Regularization/L^1 norm（TV-L^1）in registration time,registration accuracy and adaptability for large-deformation images.