【摘要】 针对大尺寸图像在图像配准过程中运算量大的问题,提出一种基于SIFT(scale-invariant feature transform)特征点检测的图像配准算法,检测前通过对待配准原图的下采样预处理,降低运算的复杂度,减少构建图像金字塔过程中高斯核卷积的运算量.采用BBF(best bin first)算法实现k-d树中k近邻点搜索,快速得到对应特征点的初始匹配对,再运用RANSAC(random sample consensus)算法在对误匹配对进行迭代剔除,得出能拟合所有内点变换模型参数的最优解,通过坐标变换和插值实现图像配准,并以峰值信噪比为指标衡量配准后的图像与参考图像之间的相似程度.实验结果表明,与传统的直接配准相比,在保证较好的配准效果条件下,本文方法能大幅缩短运行时间.更多还原

【Abstract】 In order to reduce the computational complexity in large-size image registration,the author proposes a new image registration method based on the SIFT feature point detection algorithm.Using this method,the original image is down-sampled firstly as preprocessing before detecting feature points by SIFT.Therefore,the complexity of operation,especially Gaussian-kernel convolution computation in the image pyramid building phase is reduced.In addition,the algorithm makes use of BBF algorithm to realize k neighbor points search in the k-d tree and gets initial matches of the corresponding feature points quickly.RANSAC algorithm in iteration is also used to eliminate false matches and obtain the optimal solution of the parameters of transform model that can fit all the inner-points.Finally through coordinate transformation and interpolation,image registration is realized.The similarity between the registrated image and the reference image is measured by the peak signal-to-noise ratio.The experimental results show that,the proposed method can shorten the running time and guarantee good registration performance simultaneously compared with the traditional direct registration.更多还原