基于Shape Context和尺度不变特征变换的多模图像自动配准方法  

Multimodal image registration algorithm based on Shape Context and scale-invariant feature transform

作　　者：王炳健[1] 卢刚[1] 黄洋[1] 李庆[2] 秦翰林[1] 

机构地区：[1]西安电子科技大学技术物理学院,西安710071 [2]中国科学院微电子研究所,北京100029

出　　处：《强激光与粒子束》2011年第12期3292-3296,共5页High Power Laser and Particle Beams

基　　金：中央高校基本科研业务费专项资金项目(72104699);航空科学基金项目(20100181006);陕西省自然科学基础研究计划项目(2011JQ8012);国家自然科学基金项目(60902080)

摘　　要：提出了基于修正的尺度不变特征变换(SIFT)特征提取和Shape Context特征描述算子相结合的多模图像自动配准算法,该算法利用修正的SIFT算法提取多模图像中的特征点,然后采用Shape Context算子描述特征点,利用特征点周围区域边缘点的梯度方向形成特征向量。采用欧氏距离作为匹配标准对多模图像中特征点进行初始匹配,然后通过RANSAC算法消除误匹配的特征点对,并采用最小二乘法计算仿射变换参数,最后通过仿射变换和双线性插值实现图像配准。对红外图像和可见光图像的配准实验结果表明了本算法的有效性和稳定性。An image registration method for multimodal images is proposed.This method is a combination of modified scale-invariant feature transform（SIFT） feature extraction algorithm and Shape Context feature descriptor.Feature points of multimodal images are extracted by modified SIFT feature extraction algorithm.Then each feature point is described by modified shape context descriptor,which forms a feature vector from the orientation histograms of sub-region around each feature point.After feature points matching with Euclidean distance as measure,the RANSAC algorithm is used to eliminate wrong corresponding pairs.At last,multimodal images registration is achieved by affine transformation and bilinear interpolation.Experimental results for registration of infrared images and visible images show that this method almost has the same precision as the algorithm using manually chosen corresponding points.

关 键 词：多模图像配准 尺度不变特征变换 ShapeContext特征描述算子 RANSAC算法 

分 类 号：TN911.73[电子电信—通信与信息系统]