摘要
传统SIFT算法采用128维描述算子表征一个关键点,计算量大、复杂度高,图像配准时间长,效率较低。此外,图像在采集的过程中因其尺度、旋转角度、明暗等不同因素的影响容易造成医学图像的误配准。因此,该文采用改进的SIFT算法进行配准。首先,利用快速近似最近邻搜索算法查找两幅图像上的关键点,并以其为中心取8×8的采样窗。把8×8的采样窗划分成4个模块,每一个模块是4×4的小窗口。在每一个模块内计算8个梯度方向信息,每个梯度信息就是一个特征点描述符。这样每个关键点就可以有32个SIFT特征描述符。然后,通过相似性度量判定两个关键点的相关性,并对其进行降序排序。理论上,当两幅图上的关键点一致时,两个关键点的相似性度量值等于1。本研究设置的阈值是0.95,当两点的比值大于0.95,就把两个点当作匹配点。为保证算法的精度,在传统SIFT算法的基础上采用双向匹配。只有两次匹配得到关键点的坐标之和相等,就把这对关键点当作匹配点。研究结果表明,采用改进的SIFT算法能提高配准速度,同时保持配准精度,得到较为理想的配准效果。
The traditional SIFT algorithm uses 128 dimension description operator to represent a key point,which has a large amount of calculation,high complexity,long image registration time,and low efficiency.In addition,in the process of medical image acquisition,due to the influence of different factors such as scale,rotation angle,light and shade,it is easy to cause medical image registration errors.Therefore,we propose an improved SIFT algorithm to matching.Firstly,the fast approximate nearest neighbor search algorithm is used to find the key points on the two images,and the key points is selected as the center of 8×8 sampling window.The 8×8 sampling window is divided into four modules,and each module is a 4×4 small window.Eight gradient directions are calculated in each module and each gradient information is a feature point descriptor.This allows each key point to have 32 SIFT feature descriptors.Then,the correlation between the two key points is determined by similarity measurement,and the two key points are sorted in descending order.Theoretically,when the key points of two images are consistent,the similarity measure of two key points is equal to 1.When the ratio of two points is greater than 0.95,the two points are regarded as matching points.In order to ensure the accuracy of the algorithm,dual-way matching is adopted on the basis of traditional SIFT algorithm.Only if the sum of the coordinates of the key points obtained by the two matches is equal,the key points are regarded as matching points.The results show that the improved SIFT algorithm can improve the registration speed and maintain the registration accuracy,and get a better registration effect.
作者
陈宗桂
董晓军
曾令容
张英俊
CHEN Zong-gui;DONG Xiao-jun;ZENG Ling-rong;ZHANG Ying-jun(Hunan University of Medicine,Huaihua 418000,China)
出处
《计算机技术与发展》
2022年第8期71-75,共5页
Computer Technology and Development
基金
湖南省自然科学基金面上项目(2019JJ40202)
湖南省教育厅资助科研项目(18C1135)。
关键词
高斯差分尺度空间
图像匹配
尺度不变
关键点描述
双向匹配算法
RANSAC算法
difference of Gaussian scale-space
image matching
scale invariant
key description
bidirectional matching algorithm
RANSAC algorithm
