Abstract: Image registration of endoscope video is the most important problem for Computer-aided Diagnosis(CAD) in endoscopy.For gastroscopy,this paper presents an image registration method based on homography assumption.It gets initial matching point pairs by feature point detection algorithm,and uses Delaunay triangulation to cluster the initial matching point pairs into corresponding triangules.If the triangle patch satisfies homography theorem,it marks three vertexes as the matching points.Otherwise,it registers the inscribed circle center of corresponding triangle patch with epipolar constraint.Afterwards,the corresponding triangle patches are divided into smaller ones with the inscribed circle centers and vertexes,and those newly generated patches are fed into the next loop process.The method is completed until no new patches and matching point pairs can be generated.Experimental results show that this method has better performance than FAST,SIFT,SURF and GFTT in both distribution uniformity and registration precision.

Key words: gastroscope image, image registration, triangle patch, homography, distribution uniformity, registration precision

摘要： 内镜视频图像配准是内镜下计算机辅助诊断要解决的首要问题。为此,针对胃镜场景,提出一种基于单应性假设的图像配准方法。通过特征点检测算法获得初始配准点对并进行Delaunay三角剖分,若三角面片满足单应性定理,则将其3个顶点标记为配准点,否则使用外极线约束对该面片的内接圆圆心进行配准,并通过配准点和三角面片顶点将面片分裂为更小的面片后进行迭代配准,直至没有新的面片和配准点对产生。实验结果表明,与FAST,SIFT,SURF,GFTT算法获得的初始配准点对相比,该方法获得的配准点分布均匀性较好,配准精度较高。

关键词: 胃镜图像, 图像配准, 三角面片, 单应性, 分布均匀性, 配准精度

CLC Number: 

• R318.0