一种改进的SIFT图像配准算法  

An Improved SIFT Image Registration Algorithm

作　　者：杨欢[1] 石俊生[1] 字崇德 李希才[1] 

机构地区：[1]云南师范大学颜色与图像视觉实验室,云南昆明650500

出　　处：《云南师范大学学报（自然科学版）》2017年第3期43-49,共7页Journal of Yunnan Normal University:Natural Sciences Edition

基　　金：国家自然科学基金资助项目(61178054);云南省教育厅重大专项基金资助项目(ZD2014004)

摘　　要：针对SIFT算法实时性差的缺点,提出一种改进的SIFT图像配准算法,该方法:(1)运用积分图像解决了构建尺度空间时间花费大的问题;(2)利用圆的特性对每一个特征向量进行序列化,以保证物体旋转不变性,同时将每一个描述子由原来128维降低到32维,解决了原SIFT算法维度较多所造成的算法效率低下的问题;(3)采用欧氏距离作为特征描述符之间的相似性度量并去除不具有配准意义的特征点对,提高匹配算法的鲁棒性.实验结果证明,该方法能在保证配准精度的同时降低配准的时间复杂度,在实时性要求高的应用中具有一定实用前景.SIFT algorithm is difficult to be applied to many occasions with high real-time requirement. For the shortcomings of SIFT algorithm, this paper proposes an improved SIFT algorithm, the methods are.（1） Using integral image to solve the construction problem of large scale space time; （2） Basing on the characteristics of a circle to serialize each feature vector to ensure the object rotation invariance at the same time,each descriptor from the original 128 dimension reduced to 32 dimensions, to solve the low efficiency of the original SIFT algorithm dimension caused by more problems; （3） U- sing Euclidean distance as the similarity between the feature descriptor and remove the feature points on does not have the significance of registration, improve the robustness of the matching algorithm the.The testing show that the proposed method could reduce the time complexity of registration and improve the registration accuracy.

关 键 词：尺度不变特征 SIFT 图像配准 算法改进 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]